Three Essays on Edgeworth Price Cycles in Western Australia

THREE ESSAYS ON EDGEWORTH PRICE CYCLES IN WESTERN AUSTRALIA

A dissertation presented by

Sean Isakower

to the Department of Economics

Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the field of Economics

Northeastern University Boston, MA April 2014

THREE ESSAYS ON EDGEWORTH PRICE CYCLES IN WESTERN AUSTRALIA

Sean Isakower

ABSTRACT OF DISSERTATION

Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Economics in the College of Social Sciences and Humanities at Northeastern University April 2014

Abstracts

Chapter One: Regular Price Cycles in Liquefied Petroleum Gas and a Comparison with Gasoline Price Cycles

This is the first paper to document regular price cycles in liquefied petroleum gas (LPG) and to compare LPG price cycles with gasoline price cycles in the same area. LPG price cycles in the

Perth area of Western Australia are well characterized by the theoretical Edgeworth price cycle model. LPG price cycles are much longer and more asymmetric than gasoline price cycles, which is consistent with Noel’s (2008) prediction that Edgeworth cycles are longer and more asymmetric when aggregate demand is more elastic. The aggregate demand for LPG is much more elastic than the demand for gasoline because most LPG vehicles are dual-fuel.

Chapter Two: Price Cycles and the Level of Margin in Retail Fuel Markets

This paper uses a difference-in-difference framework to estimate the effect of Edgeworth price cycles on the level of retail competition in the Perth metropolitan area LPG market using gasoline price and price-cost margins as a control for any large (but unobservable) changes to the

Perth retail fuel market. I take advantage of Fuelwatch’s collection of both retail prices and benchmark wholesale costs to calculate average market price and margin. I find that both average price and average margin increase after price cycles disappear from the LPG retail market. My findings are consistent with previous literature that suggests Edgeworth price cycles are an intermediate level of competition.

Chapter Three: Spatial Competition and Edgeworth Price Cycles

This paper analyzes the effect of a novel price law on the intensity of spatial competition in the

Perth metropolitan area. This law, which enables an online consumer information service called

Fuelwatch, requires stations to announce future price changes and maintain each price for at least

24 hours. To investigate the effect of this law on competition between stations, I construct numerous spatial weight matrices and select the appropriate spatial scope of competition. I then utilize demographic data from the Australian Census Bureau to fit a 2SLS model using the appropriate spatial scope to estimate consistent coefficients for time-invariant station characteristics. I find that while the law fails to destabilize the Edgeworth price cycles found in this area, the law does intensify the scope of spatial competition. Paradoxically, this increase in cross-price elasticity leads to less aggressive undercutting behavior, a finding that is unintuitive but consistent with the Edgeworth price cycle model.

Dedicated to my family and friends

Acknowledgements

There are many people who aided me in the completion of this dissertation. I am extremely grateful to my committee, family, and friends for all of their help and support.

First, I would like to thank my committee, James Dana, John Kwoka, and Zhongmin

Wang, for their invaluable insight and suggestions. I would also like to thank Zhongmin Wang for introducing me to my research topic and providing me with data I would not have been able to acquire otherwise.

I also thank the administrative staff in the Economics department, specifically Cheryl

Fonville, Kathleen Downey, and William Dirtion for their help and finding the solutions to issues nobody else could.

I also would like to thank my family, who supported me not only during my graduate studies but also during my entire academic career. I thank you for your guidance and unwavering belief in me.

I’m also grateful for the support of my fellow students, without whose help I certainly would have been lost. I would especially like to thank Laura Wholley for her companionship and for putting up with me in our shared office, and Shaun O’Brien for keeping me somewhat sane during exams.

Finally, I wish to express my gratitude to Erica Rosenfeld, who always believed in me, even when I felt discouraged. Thank you for everything, from being willing to do the dishes because I needed to study for an eleventh hour that day to meticulously proofreading this dissertation.

TABLE OF CONTENTS

Abstract 2

Dedication 5

Acknowledgements 6

Table of Contents 7

Chapter One: Regular Price Cycles in Liquefied Petroleum Gas and a Comparison with Gasoline Price Cycles

Section 1: Introduction 9

Section 2: Background 12

Section 3: Theory and Literature 15

Section 4: Empirical Analysis 18

Section 5: Conclusion 30

References 31

Chapter Two: Price Cycles and the Level of Margin in Retail Fuel Markets

Section 1: Introduction 33

Section 2: Literature Review 33

Section 3: Data 35

Section 4: Implications of Edgeworth Price Cycles for Consumers 41

Section 5: The Collapse of Edgeworth Price Cycles 43

Section 6: Empirical Analysis of Cycling Behavior on Average Price-cost Margins 45

Section 7: Discussion 56

Section 8: Conclusion 58

References 60

Chapter Three: Spatial Competition and Edgeworth Price Cycles

Section 1: Introduction 62

Section 2: Literature Review 64

Section 3: Data 67

Section 4: Estimation Model and a Model of Spatial Competition 71

Section 5: Weight Matric Calculation and Calculation of Weighted Prices 74

Section 6: Empirical Results 77

Section 7: Analysis of the 24-Hour Rule on Price Cycle Characteristics 84

Section 8: Policy Recommendation 86

Section 9: Concluding Remarks 88

References 90

Chapter 1: Regular Price Cycles in Liquefied Petroleum Gas and a Comparison with Gasoline Price Cycles

1. Introduction

Retail gasoline prices in many markets of the U.S., Canada, Australia, and some European countries move in sawtooth-shaped cycles: prices increase rapidly and then decrease slowly.

These regular, frequent, and asymmetric gasoline price cycles have been the subject of public debate and government investigations. A growing economic literature has examined the mechanism underlying these gasoline price cycles and found that they are well characterized by the Edgworth price cycle equilibrium in the dynamic oligopoly model of Maskin and Tirole

(1988). This literature has informed economists and antitrust authorities of the cause and welfare implications of gasoline price cycles. More generally, these studies have made important contributions to the long-standing economic literature on oligopoly pricing by carefully documenting firms’ pricing strategy with high-frequency data.

This is the first paper in this literature to document regular price cycles in liquefied petroleum gas (LPG) and, importantly, to compare LPG price cycles with gasoline price cycles observed in the same geographic area. We find that the LPG price cycles in the Perth metropolitan area of Western Australia, similar to the gasoline price cycles in the same area (e.g.,

Wang 2009), are well characterized by the Edgeworth price cycle equilibrium. We also find that

LPG price cycles are much longer and more asymmetric than gasoline price cycles. This finding is consistent with Noel’s (2008) prediction that Edgeworth cycles are longer and more asymmetric when demand is more elastic.

In a comprehensive review of the gasoline price cycle literature, Noel (2011a) writes that

“[a]n obvious direction for future work [in this area] is to search for and uncover additional

9 examples of Edgeworth Price Cycles outside of retail gasoline.” The markets for LPG and gasoline are very similar on the supply side, but differ significantly on the demand side. In the

Perth area, retail fuel stations that sell LPG almost always sell gasoline. As a consequence, LPG and gasoline are supplied by the same firms through the same network of retail fuel stations. The aggregate demand for LPG, however, is much more elastic than the demand for gasoline.1 One reason is that gasoline is a substitute for LPG because most LPG-capable vehicles are gasoline-

LPG dual-fuel vehicles.2 The fact that LPG and gasoline differ in demand but not supply offers a rare opportunity to observe the impact of aggregate demand elasticity on price cycles. It would be more difficult to separate the impact of demand from the impact of supply factors if we were to compare price cycles in a non-fuel product (i.e., the market for online keyword advertising studied by Zhang 2005) with gasoline price cycles.

For a quick overview of our results, consider three figures. Figure 1 shows Maskin and

Tirole’s (1988) numerical example of an Edgeworth price cycle.3 Figure 2 shows the brand average LPG retail prices (in Australian cents per liter before sales tax) and the wholesale price of LPG paid by one retailer in the Perth market over a period of 78 days. In both figures, firms hike prices substantially and sequentially and decrease prices gradually. These two figures, along with more detailed evidence presented in this paper, indicate that the LPG price cycles in the

Perth market are well characterized by Edgeworth price cycles. Figure 3 shows the market average gasoline and LPG retail prices during a period of 5 months. LPG prices are scaled by a factor of 1.5 because a gasoline-LPG dual-fuel car that uses 1 liter of gasoline per 10 kilometers

1 We are not aware of any studies that estimate the price elasticity of demand for LPG in Australia, but Mehrara and Ahmadi (2011) find that the own price elasticity of demand for LPG as a transport fuel in Iran is -3.58, much larger than the estimated price elasticity of demand for gasoline in the literature. 2 In contrast, about 97.5% of the vehicles during our sample period use gasoline only. For these vehicles, LPG is not a substitute for gasoline. 3 In this example, market demand is D(p) = 6-p, production cost is 0 for both firms, and price must be an integer. 10

may use 1.5 liters or more of LPG per 10 kilometers. LPG price cycles are much longer and

more asymmetric than gasoline price cycles, and the timing of the two cycles is quite different.

Figure 1: Maskin and Tirole’s (1988) Edgeworth Price Cycle Example

6 Firm 1 price Firm 2 price

Price

0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 Time

Note: Market demand is D(p) = 6-p, production cost is 0 for both firms, and price must be an integer.

Figure 2: Daily Brand Average LPG Prices Before Sales Tax in the Perth Market 2/20/02 – 5/08/02

Price (cents per liter) (cents per Price

34 02/20/02 03/02/02 03/12/02 03/22/02 04/01/02 04/11/02 04/21/02 05/01/02

BP Shell Caltex Peak Wholesale price

Figure 3: Daily Market Average Gasoline and Scaled LPG Prices in the Perth Market 11/01/02 - 4/01/03

100

Mean ULP price LPG price X 1.5

70 11/01/02 12/01/02 01/01/03 02/01/03 03/01/03 04/01/03 date

Note: LPG price is scaled by a factor of 1.5.

2. Background

LPG, a mixture of propane and butane, is a popular alternative transport fuel heavily promoted by a number of countries (e.g., Australia, Turkey, Poland, Italy, and South Korea) primarily due to environmental considerations. However, very little has been written about LPG in the academic literature, so we first offer an overview of the significance and characteristics of

LPG as a transport fuel. We then describe the Perth LPG market and the data used in this paper.

2.1. LPG as a Transport Fuel in Australia

For many years, the Australian government has promoted LPG as an alternative transport fuel on the grounds of environmental protection and energy security. LPG is thought to be less harmful to the environment,4 and Australia is thought to have an abundant indigenous supply.5

To promote its use as a transport fuel, the Australian government exempted LPG from the fuel

4 In a report to the Australian Greenhouse Office, Beer et al. (2004) find that the life-cycle emissions of LPG vehicles are below those of the equivalent class of gasoline vehicles for all emissions except carbon monoxide. 5 See Australian Senate Standing Committee on Rural and Regional Affairs and Transport (2007, p.104) for a discussion of Australia’s LPG reserves.

12 excise tax that was imposed on gasoline and diesel and provides subsidies for consumers who purchase LPG vehicles or retrofit their existing gasoline or diesel vehicles to use LPG.6

After gasoline and diesel, LPG is the third most popular vehicle fuel in Australia. Australia has a well-developed LPG distribution system with slightly more than half of the retail gasoline stations also selling LPG (Australian Department of the Environment, Water, Heritage and the

Arts 2010, p.7). Australia’s Motor Vehicle Census reported a total of 329,592 LPG or LPG- gasoline dual-fuel motor vehicles (2.5% of all motor vehicles) in 2003 and a total of 513,562

LPG vehicles (3.1% of all motor vehicles) in 2011. These LPG vehicles consume about 7% of total fuel by energy content because LPG is heavily utilized by taxis and other high-mileage fleet vehicles (Australian Transport Council and Environment Protection and Heritage Council 2008, p.14).

Most LPG-powered vehicles in Australia are conversions of gasoline engine models

(Australian Transport Council and Environment Protection and Heritage Council 2008, p.30).

For example, only 12,900 new LPG-capable vehicles were sold in 2007, while 92,000 LPG conversions were undertaken in the same period (Australian Department of the Environment,

Water, Heritage and the Arts 2010, p.8). It is important to note that most LPG vehicles in

Australia are gasoline-LPG dual-fuel.7 This means that when an LPG system was added to a car, the gasoline fuel system was retained. A major advantage of dual-fuel vehicles is that they offer drivers the option of switching between gasoline and LPG, thus extending the effective driving range of the vehicle. The downside of dual-fuel vehicles is that they are not optimized for LPG

6 See Australian Senate Standing Committee on Rural and Regional Affairs and Transport (2007, p.103) for a discussion of Australian government’s LPG vehicle subsidy program. 7 See statements by Royal Automobile Club of Australia (RACV) at http://www.racv.com.au/wps/wcm/connect/racv/Internet/Primary/my+car/advice+_+information/fuel/liquefied+petr oleum+gas+(LPG)+member+information+sheet?CACHE=NONE

13 use. Industry publications often recommend that drivers of dual-fuel vehicles use gasoline “from time to time to keep everything [in the gasoline fuel system] working properly.”8

LPG has a lower energy density than gasoline, so vehicles must use a higher volume of LPG than gasoline to cover the same distance. For example, in 2006, a Ford BF Falcon XT sedan with an automatic transmission and a gasoline engine was rated by the Australian government at

10.9 liters per 100 km, and the same vehicle fitted with an LPG-dedicated system was rated at

15.9 liters per 100 km (Morley 2006), implying a fuel consumption ratio of LPG to gasoline of

1.46. LPG-gasoline dual-fuel vehicles are not optimized for LPG use as LPG-dedicated vehicles are, so the fuel consumption ratio for dual-fuel vehicles may be considerably higher than 1.46.

2.2. Market Setting and Dataset

Our study is based primarily on a dataset that covers the daily LPG and gasoline prices at every fuel station in the Perth metropolitan area for the period January 3, 2001 to October 31,

2003. The price data was downloaded from a website (www.fuelwatch.wa.gov.au) that was established by the Western Australian government through a law commonly referred to as the

24-hour-rule. This law requires every fuel station in Western Australia to: (1) report to the government its fuel prices for the next day by 2:00 pm today and (2) post the reported prices on its price board at the beginning of the next day and maintain the prices for 24 hours. As a result, fuel firms decide their prices simultaneously (i.e., without knowing rivals’ prices) and once every

24 hours.

8 “The Pros and Cons of Converting to LPG,” http://www.tradingpost.com.au/Research/Cars/The-Pros-and-Cons-of- Converting-to-LPG

Of the 377 fuel stations in our dataset, a total of 234 sell LPG9: 229 stations sell both gasoline and LPG, and 5 stations sell LPG only. However, the sales volume of LPG is only about

7% of the volume of gasoline.10 Most of the fuel stations that sell LPG carry the brands of BP

(63), Ampol/Caltex (63), Shell (30), Mobil (17), Gull (17), and Peak (17). The first four brands belong to major oil companies, and the latter two are large independent brands in Western

Australia. The remaining fuel stations that sell LPG belong to Kleenheat (7), Woolworths Plus

(7), and other small brands (13). Woolworths Plus is the fuel brand of Woolworths, a supermarket chain that also sells fuels. Most of the 372 fuel stations that sell gasoline also carry the brands of BP (67), Ampol/Caltex (88), Shell (46), Mobil (23), Gull (38), and Peak (18).

LPG sold at fuel stations in Western Australia is produced by three firms: BP Australia,

Wesfarmers, and BHP Petroleum. These three firms’ total capacity to produce LPG far exceeds local demand. It is useful to note that the wholesale arrangements in the LPG and gasoline markets were such that each brand can control the retail prices of most of the stations that carry the brand.11 We have access to the confidential daily wholesale LPG prices paid by a multi-site

BP franchisee for the period of January 1, 2001 to October 31, 2003. These wholesale prices are reasonable proxies for the marginal costs of supplying retail LPG.

3. Theory and Literature

The theoretical foundation of the gasoline price cycle literature is the Edgeworth price cycle equilibrium in Maskin and Tirole’s (1988) dynamic oligopoly model. In this model, two

9 When analyzing cycling behavior, we do not consider those retail stations whose prices never cycled. There are about 10 such stations, typically in isolated areas. 10 According to Western Australian Select Committee on Pricing of Petroleum Products (2000, p.1 and p. 17), 1.85 billion liters of gasoline was sold in Western Australia in the 1999/2000 financial year, and approximately 132 million liters of LPG were sold as a transport fuel in Western Australia in 1999. 11 See Western Australian Select Committee on Pricing of Petroleum Products (2000, p. 17) for details of the wholesale arrangements in the LPG market and Wang (2009) for details of the gasoline market. 15 identical firms compete over the price of a homogenous product for an infinite number of periods in a market with fixed demand and production cost where firms set price alternatingly. An

Edgeworth price cycle has three phases. Let the initiation of a price restoration be the start of a new price cycle. In the falling phase, the two firms undercut each other gradually until price reaches marginal cost. At this point, the firms are in the third phase called the war of attrition phase: both firms wish prices to be increased, but neither wants to raise its price first. Once a firm relents by hiking its price first, the other firm follows with a slightly smaller price hike, and the consecutive price hikes constitute the restoration phase.

When studying the extent to which the observed price cycles can be captured by the

Edgworth price cycle theory, the existing gasoline price cycle literature has focused primarily on two sets of features of Edgworth price cycles.12 The first set of features characterizes cycle structure. A basic and obvious feature of the Edgeworth price cycle is that (1a) the length of the restoration phase is shorter than that of the non-restoration phase.13 Another basic feature of

Edgeworth cycle is that (1b) changes in cost or demand cannot explain its existence since both remain fixed in the basic model. Changes in cost, however, are expected to have subtle impacts on cycle characteristics. In the price cycle equilibrium, a restoration does not start until price approaches marginal cost. If marginal cost itself has decreased since the initial restoration, it would naturally take a longer time for price to fall close to marginal cost again. This logic implies that (1c) the length of a cycle tends to be longer if marginal cost is lower at the end of the cycle than at the start of the cycle.

12 The gasoline price cycle literature also examines two other questions that we do not address in this paper: in what types of gasoline markets are price cycles more likely to arise (e.g., Eckert 2003, Noel 2007a, Doyle et al. 2010) and what are the welfare implications of gasoline price cycles (e.g., Lewis and Noel 2011, Noel 2011b, Zimmerman et al. 2011). 13 The term non-restoration phase refers to the sum of the falling phase and the war of attrition phase. Since the start of a war of attrition phase is hard to identify in practice, much of the gasoline price literature addresses the falling phase and the war of attrition phase together. 16

The second set of features characterizes firm behavior, emphasizing that the Edgeworth price cycle is a theory of oligopoly price competition. For this oligopoly theory to be applicable, the price setters in a retail gasoline or LPG market should be a few firms instead of the large number of individual stations. Moreover, the war of attrition problem embedded in the price cycle equilibrium generates the need for oligopoly firms to coordinate price increases (Noel

2008, Wang 2008). The basic intuition is that firms have strong incentive to be the last to increase price; those that increase their prices early lose market share. The gasoline price cycle literature finds that price hikes in cycling markets exhibit high degrees of intrabrand synchronization and uniformity (e.g., Wang 2009 and Lewis 2012), suggesting a small number of brands are price setters. This literature also finds that firms use the facilitating practice of price leadership and followership to coordinate price increases; large brands tend to be leaders, small brands tend to be followers, and followers tend to undercut leaders (e.g., Eckert and West

2004, Noel 2007b, Atkinson 2009, Lewis 2012, Byrne and Ware 2011). The coordination problem at the cycle bottom makes larger firms, which control a large number of retail stations, more natural and effective price leaders.14 We examine (2a) whether LPG price hikes exhibit high degrees of intrabrand synchronization and uniformity and (2b) whether firms use price leadership to coordinate price hikes and, if so, how.

Noel (2008) uses computational methods to extend the Maskin and Tirole (1988) model along several dimensions. In particular, Noel finds that aggregate demand elasticity impacts the shape of the price cycle. When aggregate demand is more elastic, firms are less aggressive in undercutting. By undercutting its competitors, a firm can increase its sales volume through two

14 The two firms in the Maskin and Tirole model alternate as the price leader since they relent with equal probability. Eckert (2003) considers two firms of different size and find that for some parameters, the larger firm is more likely to be the price leader. Noel (2007b) and Wang (2008) make the informal argument that the coordination problem at the cycle bottom makes larger firms more natural and effective price leaders. 17 channels: stealing existing consumers from its competitors or attracting new consumers to the market. Firms are less aggressive in undercutting in a market with a more elastic demand curve because a small price cut in such a market can lead to the same sales increase as a larger price cut in a market with a less elastic demand curve. Less aggressive undercutting implies that the non- restoration phase is longer, and given that the length of the restoration phase is fixed in a model with two players, cycles are more asymmetric. Since the aggregate demand for LPG is much more elastic than that for gasoline, Noel’s model would predict that LPG price cycles are longer and more asymmetric than gasoline price cycles.

We close this section by noting a discrepancy between theory and reality. Firms in the

Maskin and Tirole model, by assumption, always set their prices sequentially, while firms in the

Perth LPG market have to set their prices simultaneously due to the 24-hour rule. While fully addressing this discrepancy is beyond the scope of this study, Figure 2 indicates that firms in the

Perth LPG market do increase their prices sequentially ex post.

4. Empirical Analysis

Figure 4 shows the market average LPG retail prices for our entire sample period. LPG price cycles started to emerge in the Perth market in June 2001 but were not regular until January

2002. Since our focus is on equilibrium behavior, we focus on the period of January 2002 through October 2003. We organize our empirical analyses of LPG price cycles into four subsections: basic LPG cycle characteristics, impact of cost changes, intrabrand pricing behavior, and price leadership and followership. Subsection 5 compares LPG price cycles with gasoline price cycles.

4.1. Basic LPG Cycle Characteristics

As in the Edgeworth cycle theory, we define the lead price hike of a cycle as the first price hike by any brand after a series of small price decreases.15 The day on which a lead price hike takes place is the start of a new price cycle. The firms that initiate lead price hikes are called price leaders. The length of a price cycle is then the number of days from the start through the end of the cycle. The length of the restoration phase is the number of days from the start day through the peak day. A cycle peak day is a day on which the market average price reaches the maximum level within the cycle period. Figures 2 and 3 indicate that we can identify the cycle start days and the cycle peak days in a clear-cut way.

For the sample period of January 2002 through October 2003, we identify 37 cycles. We do not observe the end of the 37th cycle, so we do not know the length of this cycle. The length of the first 36 cycles ranges from 10 days to 30 days, with the mean and the median both 18 days.

The average length of the restoration phase and the non-restoration phase is 4.3 and 13.9 days, respectively. A simple paired t test, with a p-value of 0.0000, strongly confirms cycle feature

(1a): the length of the restoration phase is smaller than that of the non-restoration phase.

15 For a brand’s price increase to be considered a lead price hike, the brand’s average price increase must be at least 1 cent per liter. Almost all lead price hikes are as clear cut as those shown in Figure 2. 19

Figure 4: Perth Market Average LPG Retail Prices Before Sales Tax

Prices: Australian cents perliter Australian Prices:

35 Perth market average before-sales-tax price Wholesale price a multi-site BP retailer pays

01/08/01 01/22/01 02/05/01 02/19/01 03/05/01 03/19/01 04/02/01 04/16/01 04/30/01 05/14/01 05/28/01 06/11/01 06/25/01 07/09/01 07/23/01 08/06/01 08/20/01 09/03/01 09/17/01 10/01/01 10/15/01 10/29/01 11/12/01 11/26/01 12/10/01 12/24/01 Date

Prices: Australian cents perliter Australian Prices:

Perth market average before-sales-tax price Wholesale price a multi-site BP retailer pays

01/08/02 01/22/02 02/05/02 02/19/02 03/05/02 03/19/02 04/02/02 04/16/02 04/30/02 05/14/02 05/28/02 06/11/02 06/25/02 07/09/02 07/23/02 08/06/02 08/20/02 09/03/02 09/17/02 10/01/02 10/15/02 10/29/02 11/12/02 11/26/02 12/10/02 12/24/02 Date

55 Perth market average before-sales-tax price Wholesale price a multi-site BP retailer pays

Prices: Australian cents perliter Australian Prices:

01/01/03 01/15/03 01/29/03 02/12/03 02/26/03 03/12/03 03/26/03 04/09/03 04/23/03 05/07/03 05/21/03 06/04/03 06/18/03 07/02/03 07/16/03 07/30/03 08/13/03 08/27/03 09/10/03 09/24/03 10/08/03 10/22/03 11/05/03 Date

Cycle amplitude can be measured by the height of the restoration phase (i.e., the restoration amplitude) or the height of the falling phase (i.e., the falling amplitude). The average restoration amplitude of the 37 cycles is 5.9 cents per liter, and the average falling amplitude of the first 36 cycles is 6.0 cents per liter. Cycle amplitude is, on average, 14.6% of LPG retail prices, the average of which is 41.1 cents per liter during the sample period with regular cycles. Even though the two average values are close to each other, the values of these two measures are often very different for a given cycle. In fact, the difference between these two measures ranges from

-9.1 to 4.2 cents per liter. The two amplitude measures are always equal to each other in the

Maskin and Tirole model where marginal cost is fixed, but these two measures may differ for some LPG cycles as the marginal cost of supplying retail LPG changes over time. The next subsection studies the difference between the two amplitude measures further.

Of the 37 cycles, 15 start on Sunday, 3 on Saturday, 4 on Monday, 4 on Tuesday, and 5 on

Friday. The fact that nearly half of the cycles start during the weekend leads one to suspect that demand might play a role. A price leader has the incentive to start a price hike on a day with relatively low demand so that its loss of market share is relatively small. However, we cannot offer any evidence for or against this hypothesis as we do not observe daily LPG demand.

4.2. Impact of Cost Changes

Figures 2 and 4 support cycle feature (1b) that changes in cost cannot explain the existence or the dynamics of the LPG price cycles.16 Retail LPG prices change much more frequently than wholesale LPG prices do. The data also supports cycle feature (1c) that cycle length tends to be longer if marginal cost is lower at the end of a cycle than at the beginning of the cycle. Figure 5 shows the impact of cost changes on cycle length. Cost change is defined as

16 Demand changes cannot explain the existence of LPG price cycles either. Firms face the same market demand every day, yet some firms hike price on certain days while others do not. 21 the wholesale price paid by the multi-site BP franchisee at the end of a cycle minus the wholesale price paid at the beginning of the cycle. If we regress cycle length on cost change and a constant, the coefficient on cost change is – 0.92, with a standard error of 0.33.

Recall that the logic behind Figure 5 is that a price cycle does not end until price falls close to marginal cost, and that if marginal cost has fallen since the start of the price cycle, it takes longer for retail price to fall close to marginal cost again. The same logic also implies that if marginal cost has become lower than it was at the start of a cycle, price should experience a larger fall before it approaches marginal cost again, implying that the amplitude of the falling phase is larger than the amplitude of the restoration phase.

Figure 6 confirms this testable implication. If wholesale price is lower at the end of a cycle than at the beginning, the height of its falling phase tends to be larger than the height of its restoration phase. In a linear regression in which the dependent variable is the height of the falling phase minus the height of the restoration phase and the explanatory variables are cost change and a constant, the cost change coefficient is – 0.95, with a standard error of 0.13.

Figure 5: Impact of Cost Changes on Cycle Length

Cycle length (days) Cycle length

10 -10 -5 0 5 Change in wholesale price

Note: The x axis is the wholesale price paid by the multi-site BP franchisee at the end of a cycle minus the wholesale price paid at the beginning of the cycle.

Figure 6: Impact of Cost Changes on Cycle Amplitude

Falling amplitude Falling - amplitude restoration

-5 -10 -5 0 5 Change in wholesale price

Note: The x axis is the wholesale price paid by the BP multi-site franchisee at the end of a cycle minus the wholesale price paid at the beginning of the cycle. Falling amplitude is the height of the falling phase, and the restoration amplitude is the height of the restoration phase.

4.3. Intrabrand Pricing Behaviors

A conspicuous feature in Figure 2 is that firms hike their prices sequentially. For this

sequential pattern of price hikes to emerge, there must be some degree of intrabrand

synchronization in price hikes. In this subsection, we present further evidence that LPG price

hikes, but not price decreases, exhibit a high degree of intrabrand synchronization and

uniformity. Before doing so, it is useful to recall why this pattern is consistent with the

Edgeworth price cycle equilibrium. Intrabrand synchronization and uniformity dramatically

reduces the number of price setters from over 200 fuel stations to a few oligopoly firms and

allows firms to send unambiguous signals to competitors, thus facilitating pricing coordination.

Why are intrabrand synchronization and uniformity present in price hikes but not in price

decreases? At the bottom of each cycle, a war of attrition problem exists that requires

coordination, while no such problem exists during the falling phase.

Figure 7(a) shows, by cycle day, the box-whisker plots of BP-branded stations’ LPG prices

before and during the 15th cycle in our sample. This cycle was led by BP alone. The x-axis

indicates cycle days: 1 is the first day of the 15th cycle, and -1 is the day immediately before the

start of the 15th cycle or the last day of the 14th cycle. This figure shows that most of the BP-

branded stations’ LPG prices were hiked to a single price (49.9 cents per liter) in the first three

days of the 15th cycle, but these stations’ prices exhibit considerable levels of price dispersion

before and after the restoration phase of this cycle.

Figure 7(b) shows, by cycle day, the box-whisker plots of Shell-branded stations’ LPG

prices before and during the 15th cycle. Shell’s prices exhibit some dispersion on the first day of

the cycle as Shell was not a leader for this cycle. On the second day of the cycle, Shell followed

BP’s lead price hike by hiking most of its prices to exactly 49.9 cents per liter. This is why

Shell’s prices collapsed to a single point in Figure 7(b) on the second day of the cycle. It is

useful to point out that BP, Caltex, Shell, and Mobil, if not a leader of a cycle, almost always

hiked the majority of their retail LPG prices to exactly the same level as the one to which the

price leader hiked (most of) its prices.

Figure 7(a): Box Plots of BP Stations’ Prices before and during a Price Cycle Led by BP

Cents perCents liter

42 -3 -2 -1 1 2 3 4 5 6 7 8 9 excludes outside values

Note: The x-axis is cycle days: 1 is the first day of a cycle, and -1 is the day immediately before the start of the cycle.

Figure 7(b): Box Plots of Shell Stations’ Prices before and during a Price Cycle Led by BP

Cents perCents liter

-3 -2 -1 1 2 3 4 5 6 7 8 9 excludes outside values

Note: The x-axis is cycle days: 1 is the first day of a cycle, and -1 is the day immediately before the start of the cycle.

Figure 7(c): Box Plots of BP Stations’ Prices across 11 Cycles Led by BP Alone

-5

-10

Cents perCents liter

-15

-20

-3 -2 -1 1 2 3 4 5 6 7 8 9 excludes outside values

Note: The x-axis is cycle days: 1 is the first day of a cycle, and -1 is the day immediately before the start of the cycle.

The patterns observed in Figures 7(a) and 7(b) hold across other cycles. To illustrate, we

follow Lewis’ (2011) method of plotting normalized prices. Define the restoration price of a

cycle as the mode price charged by the leader brand. For example, the restoration price of the

15th cycle is 49.9 cents per liter, the mode price charged by BP stations. We then obtain the

normalized price of a station on any day of a cycle by subtracting the restoration price of the

cycle from this station’s price on that day. Normalized prices on any day measure the difference

from the restoration price of the corresponding cycle, making them comparable across cycles.

Figure 7(c) shows, by cycle day, the box plots of BP stations’ LPG prices across the 11 cycles for which BP was the sole leader. Again, the distribution of BP stations’ prices collapsed on the first two days of the 11 cycles that it led alone. In fact, of the 583 normalized BP prices on the first days of the 11 cycles, 455 prices are 0, 123 prices are smaller than 0, and only 6 are bigger than 0. Almost all of the 123 normalized prices that are smaller than 0 are those of the stations that had not yet hiked prices on the first day of a cycle. BP’s prices are dispersed on the third day in this figure because BP started to cut some stations’ prices on the third day of some cycles.

4.4. Price Leadership and Followership

Figure 2 indicates firms hike their prices sequentially, generating a conspicuous pattern of price leadership and followership. In this subsection, we summarize the price leadership and followership patterns across the 37 regular LPG price cycles in our sample.

Of the 37 cycles, 18 are led by Shell alone, 11 by BP alone, 3 by Caltex alone, and 1 by

Mobil alone. Four cycles are co-led, with multiple brands initiating price hikes on the same day.

Two of these cycles are co-led by Shell and BP, 1 by Caltex and BP, and 1 by Caltex and Shell.

In sum, BP was a price leader for 14 cycles, Caltex for 5 cycles, and Shell for 20 cycles. Price leaders in this LPG market are essentially the three largest LPG firms. Co-leadership is a by- product of the 24-hour-rule, under which firms must decide whether to hike price without knowing rivals’ prices. This uncertainty about rivals’ actions may lead firms to hike price on the same day.

One way to summarize firms’ leader-follower behavior is to consider a representative cycle.

We use normalized prices to construct a representative or average cycle. We subtract the market average price on the day immediately before the start of the current cycle from each station’s

LPG price on each day of a cycle. Figure 8 shows, by brand and cycle day, the normalized

prices averaged across the 37 cycles. Market average price reaches its highest level on the fourth

day of the representative cycle. The six brands in this figure sort themselves into two groups:

Shell, BP, and Caltex, the price leaders, generally have above market average prices; and Gull,

Peak, and Woolworths Plus, the price followers, have below market average prices. During the

restoration phase, the three follower brands’ prices are much lower than those of the leader

brands, and during the falling phase, the price differences among the brands are fairly small.

Figure 8: Normalized LPG Prices Averaged across the 37 Cycles

Normalized prices averaged across Normalizedcycles averaged across prices

-2

-3 -2 -1 1 2 3 4 5 6 7 8 9 Day of cycle

Shell BP Caltex Market average Gull Peak Woolworths Plus

Note: The x-axis is cycle days: 1 is the first day of a cycle, and -1 is the day immediately before the start of the cycle.

To show the precise price differences among the brands on various cycle days, we report

in Table 1 the results of a regression in which the dependent variable is a station’s daily price and

the independent variables are brand dummies. (The results are very similar if we use normalized

prices as the dependent variable in this regression.) The constant term represents the price level

of Peak-branded stations, and the slope coefficients are differences from Peak stations’ prices.

The prices of the major oil firms are always higher than those of Peak or Woolworths Plus, and

the difference is larger during the restoration phase than during the falling phase. On cycle day

3, the prices of Shell, BP, Caltex, and Mobil are all over 5 cents per liter higher than those of

Peak or Woolworths. The major oil firms’ prices are also higher than Gull’s during the restoration phase, but not during some of the falling days. Woolworths Plus tends to have the lowest retail LPG prices. On cycle day 1, Shell and BP have the highest prices as they are the most frequent price leaders.

Table 1 Estimated Price Differences among Brands on Various Cycle Days Cycle Days -1 1 3 4 5 8 9 0.89*** 4.45*** 6.38*** 3.06*** 1.84*** 1.20*** 0.85*** Shell (0.12) (0.64) (0.56) (0.45) (0.22) (0.15) (0.17) 0.79*** 3.99*** 5.90*** 2.97*** 1.49*** 1.23*** 0.94*** BP (0.16) (0.71) (0.38) (0.44) (0.17) (0.16) (0.14) 0.78*** 1.79*** 5.31*** 2.50*** 1.46*** 0.87*** 0.68*** Caltex (0.13) (0.47) (0.59) (0.42) (0.20) (0.14) (0.12) 1.47*** 1.59*** 5.36*** 3.18*** 1.93*** 1.32*** 1.14*** Mobil (0.22) (0.25) (0.56) (0.38) (0.22) (0.15) (0.19) 0.86*** 0.83*** 1.37*** 0.44 0.36 1.01*** 1.05*** Gull (0.17) (0.17) (0.49) (0.34) (0.22) (0.18) (0.20) -0.43 -0.37 -0.25 -1.34** -0.88* -0.14 -0.16 Woolworth (0.30) (0.31) (0.48) (0.58) (0.44) (0.35) (0.36) 0.67*** 0.81*** 0.99** -0.13 -0.00 0.49*** 0.46*** Minor Brands (0.10) (0.12) (0.39) (0.31) (0.18) (0.11) (0.12) 40.78*** 40.89*** 42.69*** 45.44*** 46.11*** 44.38*** 44.07*** Constant (1.05) (1.02) (1.12) (0.94) (0.88) (0.99) (1.01) # of cycles 37 37 37 37 37 32 30 # of obs. 7,284 7,265 7,268 7,266 7,268 6,264 5,848 R-squared 0.0029 0.0502 0.118 0.0453 0.0154 0.0048 0.0033 Note: Day -1 is the last day of the previous cycle or the day immediately before the first day of the current cycle. The constant represents Peak’s retail LPG prices. In parentheses are robust standard errors (clustered by cycle). *** Significant at the 1% level; ** significant at the 5% level; * significant at the 10% level.

The normalized prices of the three leader firms in Figure 8 increase on each of the first three cycle days. This does not mean that brands increase the price of individual stations three times during the restoration phase. Individual stations essentially always hike price in a single jump.

The normalized prices averaged across cycles increase on consecutive days for two reasons.

First, a brand’s behavior may differ across cycles. A large brand may be a leader in some cycles, but a second- or third-day follower in other cycles. Once averaged across cycles, price will

28 increase on consecutive days. Second, intrabrand synchronization is not perfect. A small proportion of a brand’s stations may not hike price on the same day as the brand’s other stations.

4.5. Comparison with Gasoline Price Cycles

The LPG price cycle dynamics documented in previous subsections are similar to those of gasoline price cycles studied in the existing literature (e.g., Wang 2009). However, there is a major difference between LPG cycles and gasoline cycles in the Perth area: LPG price cycles are much longer and more asymmetric than gasoline price cycles.

From January 2002 to October 2003, there were a total of 76 full gasoline price cycles, which is more than twice the number of LPG price cycles during the same period. Hence, the average length of the gasoline price cycles, which is 8.6 days, is less than half the average length of the LPG price cycles, which is 18 days. (The length of the gasoline price cycles range from 5 days to 14 days, with a median of 8 days.) For the gasoline price cycles, the average length of the restoration phase and the non-restoration phase is 2.8 and 5.8 days, respectively. Hence, gasoline price cycles are less asymmetric than LPG price cycles; the ratio of the length of the non-restoration phase to the length of the restoration phase is 2.2 for gasoline price cycles and

3.4 for LPG price cycles. Given that gasoline price cycles differ from LPG price cycles in terms of length and degree of asymmetry, it is not surprising that these two cycles, as those in Figure 3, appear to be independent from each other.

Why are the LPG price cycles much longer and more asymmetric? These findings are consistent with Noel’s (2008) prediction that price cycles in a market with greater aggregate demand elasticity are longer and more asymmetric. The suppliers and the market structure of the

LPG and gasoline markets are essentially the same. The main difference between these two markets lies in the demand side; the demand for LPG is much more elastic than the demand for

29 gasoline because most LPG vehicles can also use gasoline. Noel (2008) also finds that cycles become longer and more asymmetric when firms face tighter capacity constraints. However, we are not aware of any evidence that suggests capacity constraints, if they exist at all, are tighter in the LPG market than in the gasoline market.

5. Conclusion

This paper contributes to the literature on gasoline price cycles by documenting the LPG price cycles in the Perth area of Western Australia and comparing the LPG price cycles with the gasoline price cycles in the same area. We find that the LPG price cycles, similar to the gasoline price cycles, are well characterized by the Edgeworth price cycle model. In particular, we have emphasized the role of intrabrand synchronization and price leadership in generating the regular

LPG price cycles. We also find that LPG price cycles are much longer and more asymmetric than gasoline price cycles, which is consistent with the fact that the aggregate demand for LPG is more elastic and that Edgeworth cycles are predicted to be longer and more asymmetric when demand is more elastic. Our paper is the first to compare regular price cycles in two products, and our findings provide empirical evidence on how demand elasticity affects Edgworth price cycles.

References

Atkinson, Benjamin. 2009. “Retail Gasoline Price Cycles: Evidence from Guelph, Ontario Using Bi-Hourly, Station-Specific Retail Price Data.” Energy Journal 30(1): 85-109.

Australian Department of the Environment, Water, Heritage and the Arts. 2010. “Review of LPG (Autogas) Fuel Quality Standard.” Draft Discussion Paper. Available at http://www.environment.gov.au/atmosphere/fuelquality/publications/lpg-fuel-quality- standard-discussion-paper.html

Australian Senate Standing Committee on Rural and Regional Affairs and Transport. 2007. “Australia’s Future Oil Supply and Alternative Transport Fuels.” Available at http://www.aph.gov.au/Parliamentary_Business/Committees/Senate_Committees?url=rra t_ctte/completed_inquiries/2004-07/oil_supply/index.htm

Australian Transport Council and Environment Protection and Heritage Council. 2008. “Vehicle Fuel Efficiency: Potential Measures to Encourage the Uptake of More Fuel Efficient, Low Carbon Emission Vehicles.” Public Discussion Paper. Available at http://www.thepep.org/ClearingHouse/docfiles/Vehicle.Fuel.Efficiency.pdf

Beer, Tom, Tim Grant, Harry Watson, and Doina Olaru. 2004. “Life-Cycle Emissions Analysis of Fuels for Light Vehicles: A Report to the Australian Greenhouse Office.” Available at http://113.105.0.85:8076/resnet/admin/upload/news/1002074/lightvehicles.pdf

Byrne, David and Roger Ware. 2011. “Price Cycles and Price Leadership in Gasoline Markets: Evidence from Rural and Urban Canada.” Working Paper. Available at http://ssrn.com/abstract=1903558

Doyle, Joseph, Erich Muehlegger, and Krislert Samphantharak. 2010. "Edgeworth Cycles Revisited." Energy Economics 32(3): 651-60.

Eckert, Andrew. 2003. "Retail Price Cycles and the Presence of Small Firms." International Journal of Industrial Organization 21(2): 151-70.

Eckert, Andrew, and Douglas S. West. 2004. "Retail Gasoline Price Cycles across Spatially Dispersed Gasoline Stations." The Journal of Law and Economics 47(1): 245-73.

Lewis, Matthew S. 2012. “Price Leadership and Coordination in Retail Gasoline Markets with Price Cycles.” International Journal of Industrial Organization 30(4): 342-351.

Lewis, Matthew S., Michael D. Noel. 2011. “The Speed of Gasoline Price Response in Markets with and without Edgeworth Cycles.” Review of Economics and Statistics 92(3): 672- 682.

Maskin, Eric, and Jean Tirole. 1988. "A Theory of Dynamic Oligopoly, II: Price Competition, Kinked Demand Curves, and Edgeworth Cycles." Econometrica 56(3): 571-99.

Mehrara, Mohsen, and Saeedeh. 2011. “The Estimation of the Automotive Fuel Demand in Iran: Almost Ideal Demand System Approach.” Australian Journal of Business and Management Research 1(7): 72-77.

Morley, David. 2006. “LPG: Too Good to Be True?” The Age, September 27. Available at http://www.drive.com.au/editorial/articledetail.aspx?ArticleID=20340&IsPgd=0

Noel, Michael D. 2007a. “Edgeworth Price Cycles, Cost-based Pricing and Sticky Pricing in Retail Gasoline Markets”. Review of Economics and Statistics 89(2): 324-334.

Noel, Michael D. 2007b. "Edgeworth Price Cycles: Evidence from the Toronto Retail Gasoline Market." Journal of Industrial Economics 55(1): 69-92.

Noel, Michael D. 2008. "Edgeworth Price Cycles and Focal Prices: Computational Dynamic Markov Equilibria." Journal of Economics & Management Strategy 17(2): 345-77.

Noel, Michael D. 2011a. "Edgeworth Price Cycles." The New Palgrave Dictionary of Economics, Online Edition. Eds. Steven N. Durlauf and Lawrence E. Brume. Palgrave Macmillan.

Noel, Michael D. 2011b. “Edgeworth Price Cycles and Intertemporal Price Discrimination.” Energy Economics 34(4): 942-954.

Western Australia Select Committee on Petroleum Products Pricing. 2000. “Getting a Fair Deal for Western Australian Motorists”. Available at http://www.parliament.wa.gov.au/Parliament/Commit.nsf/(Report+Lookup+by+Com+ID )/8BB6CD2152C1F86C48257831003E94B7/$file/RPT+PETROL+PRICING.pdf

Wang, Zhongmin. 2008. “Collusive Communication and Pricing Coordination in a Retail Gasoline Market.” Review of Industrial Organization 32(1): 35-52.

Wang, Zhongmin. 2009. "(Mixed) Strategy in Oligopoly Pricing: Evidence from Gasoline Price Cycles Before and Under a Timing Regulation." Journal of Political Economy 117(6): 987-1030.

Zhang, Xiaoquan. 2005. “Finding Edgeworth Cycles in Online Advertising Auctions.” MIT Sloan School of Management Working Paper.

Zimmerman, Paul, John M. Yun, and Christopher T. Taylor. 2011. “Edgeworth Price Cycles in Gasoline: Evidence from the U.S.” FTC Working Paper No. 303.

Chapter 2: Price Cycles and the Level of Margin in Retail Fuel Markets

1. Introduction

The purpose of this paper is to examine the level of retail competition in a market that exhibited Edgeworth price cycles and then transitioned to a more stable equilibrium. Using data from the Fuelwatch website (see chapter 1 for market details), I am able to calculate the price- cost margin for each station on each day. I use these calculated price-cost margins and station prices to estimate a difference-in-difference like estimator to investigate the average change in retail margin between the LPG market, in which Edgeworth price cycles disappeared, and the gasoline market, in which they did not.

I find evidence that average posted price and margin increase after the cessation of the price cycles. These findings are consistent with theoretical findings in the literature suggesting that Edgeworth price cycles represent an intermediate level of competitiveness between the competitive equilibrium and focal point tacit collusion.

2. Literature Review

Edgeworth price cycles can be framed in one of two ways: first, one can think of price cycles as a form of tacit collusion, allowing firms to raise average price above competitive levels; second, they can be considered a form of continually repeating price wars, as seen in

Castanias and Johnson (1993), who discovered and analyzed gasoline price cycles in the Los

Angeles metropolitan area. In the second interpretation, a price leader will attempt to “reset” the market at a higher price level; however, since the incentive to undercut is too large, a price war ensues until one competitor relents. Other papers, such as Eckert (2002), also uncover some evidence that markets exhibiting price cycles tend to be more competitive than other, non- cycling fuel markets.

This paper builds on the literature investigating the welfare effects of Edgeworth price cycles. Since their empirical discovery by Castanias and Johnson (1993), the welfare effect of price cycles has been debated. Consumers, often noting the swift and market-wide increase in prices, tend to believe that price cycles are evidence of harmful collusion among fuel retailers.

Castanias and Johnson (1993), however, characterized these cycles as periodic price wars, highlighting the fact that stations would be better off if they were able to maintain a higher price level. Other papers support this view, with Eckert (2002) finding that under a theoretical model of Edgeworth price cycles, focal point coordination would yield higher average profits than cycling. Similarly, Doyle (2010) finds that markets in the U.S. that exhibit cycles are neither concentrated nor too competitive, implying price cycles are an intermediate level of competition that is more competitive than many fuel markets not experiencing price cycles. Lewis (2010) finds that pass through rates in midwestern U.S. cities that cycle are faster than in those that do not, suggesting that the lack of price rigidity in cycling markets might improve the market inefficiencies caused by price rigidities. Finally, Zimmerman, Yun, and Taylor (2013) study

MSA-level retail gasoline prices from 1996 to 2010. The unprecedented length of this time series in the U.S. allowed the authors to observe changes in market price during the transition from rigid prices to price cycles. They find that market prices in MSAs that began cycling are lower. These papers all suggest that price cycles may be more competitive than they otherwise appear.

To detect cycles in a manner that is consistent and reproducible, I utilize the characteristics of price cycles that have been described in the price cycle literature. For example,

Eckert (2002) notes that price cycles (consistent with their theoretical formulation) are very asymmetric, with the undercutting (falling) phase lasting significantly longer than the restoration

(rising) phase. Atkinson (2009), who uses bi-hourly prices in Guelph, Ontario, finds retail price decreases 4.3 times more often than it increases in this market. Lewis (2010) exploits these characteristics by calculating the median daily price change and finds cities with negative median price changes are also cities that cycle. This paper follows Lewis (2010) by using median price changes to identify periods of cycling and non-cycling.

3. Data

3.1. The Fuelwatch Data

I collect this paper’s data from a public website called Fuelwatch, run by Western

Australia’s Department of Commerce. This is a rich and accurate station-level panel dataset enabled by a law colloquially called the 24-hour rule. The 24-hour rule requires all stations to report their retail prices to the government and then maintain the reported price for a 24-hour period: 6 a.m. to 6 a.m. the following day. With this publically mandated and maintained data, I am able to observe the full set of market prices for both LPG and gasoline from January 1, 2003 to December 31, 2012 in the Perth metropolitan area. This data also contains the station name, brand and street address.

In addition to the price data, Fuelwatch collects and releases a set of benchmark wholesale prices in an effort to increase the transparency of the market. Using these wholesale prices, I construct a proxy for the marginal cost of providing LPG and gasoline which I then use to calculate retail margin. To construct the marginal cost proxy for gasoline, I calculate the average daily terminal price for regular unleaded gasoline. For LPG, I use the Saudi Aramco contract price, which is the benchmark wholesale price for the region. Due to the long timeframe

35 of my data, I convert all pre-tax prices into real 2012 Australian dollars. I report summary statistics for LPG in Table 1 and ULP (gasoline) in Table 2.

Table 1 LPG Summary Price and Margin Statistics All variables are pre-tax in real 2013 AUS cents Variable N Mean Std. Dev Min Max 2003 price 121770 60.37 15.77 38.36 128.17 margin 121770 28.98 14.36 9.88 100.46 2004 price 125934 57.76 14.84 33.92 121.27 margin 129534 26.21 14.04 6.13 98.75 2005 price 124290 64.29 13.01 41.26 111.98 margin 124290 28.14 12 7.91 80.33 2006 price 120375 71.7 12.81 52.88 121.32 margin 120375 29.32 12.67 11.78 83.74 2007 price 116001 71.36 13.47 43.39 125.93 margin 116001 27.57 12.94 -2.86 84.71 2008 price 118445 80.57 13.95 50.12 144.5 margin 118445 29.55 13.87 8.06 102.44 2009 price 119305 69.79 15.11 44.89 129.8 margin 119305 32.55 14.88 4.54 98.97 2010 price 120721 75.16 13.97 47.89 131.58 margin 120721 32 13.77 1.79 94.74 2011 price 120999 77.44 13.53 55.98 127.73 margin 120999 33.43 13.37 9.07 87.97 2012 price 120829 85.8 13.88 49.95 146.95 margin 120829 39.22 13.13 7.45 96.55

Table 2 ULP Summary Price and Margin Statistics All variables are pre-tax in real 2013 AUS cents Variable N Mean Std. Dev Min Max 2003 price 223026 122.26 9.65 100.72 160.38 margin 223026 7.1 8.24 -11.98 53.84 2004 price 230483 128.41 10.8 103 171.46 margin 230483 5.51 9.28 -12.17 53.72 2005 price 222968 140.83 15.23 107.96 206.8 margin 222968 4.99 9.69 -16.22 56.17 2006 price 216437 152.59 13.66 124.5 206.59 margin 216437 6.36 9.8 -11.44 71.46 2007 price 206034 149.15 11.78 114.36 204.34 margin 206034 7.8 9.54 -17.48 64.08 2008 price 203601 162.04 19.4 103.66 229.39 margin 203601 9.42 11.57 -10.47 108.98 2009 price 198908 133.14 11.77 101.57 188.21 margin 198908 6.84 10.93 -18.93 80.69 2010 price 198898 136.9 9.34 115.47 181.79 margin 198898 7.59 9.72 -10.38 59.19 2011 price 196662 147.53 9.27 102.2 204.92 margin 196662 8.46 8.57 -35.75 67.73 2012 price 196924 147.8 10.38 118.1 207.5 margin 196924 8.91 9.65 -14.94 74.77

3.2. Cycle Detection

Although Edgeworth price cycles are noticeable to consumers, the task of identifying the markets and time periods where cycles exist is not straightforward. Numerous papers, such as

Lewis and Noel (2010) and Eckert (2002), explore methods to identify price cycles. I identify price cycles in two ways. First, I manually indicate which time periods exhibit price cycles; and second, I follow Lewis (2010) and Doyle et al. (2010) by using negative median price changes to identify cycling periods.

Edgeworth price cycles have a very distinctive shape and are easily identified by looking at the series of mean market retail prices. Eckert (2002) calls this the “eyeball test” and notes that, while effective, its primary downside is that the result is non-reproducible. I use this method to confirm the results from the median price change measure. Figure 1 shows the average market retail price and wholesale costs from January 2004 to February 2004 in the market for LPG when cycles are present, while Figure 2 presents the same variables from

January 2012 to February 2012, when cycles are not present.

Figure 1: Average Market Price and Wholesale Cost of LPG January 2004 – March 2004 When Cycles are Present

Figure 2: Average Market Price and Wholesale Cost of LPG January 2012 – March 2012 When Cycles are not Present

The two price series in Figures 1 and 2 are easily distinguishable from one another, allowing for easy identification of cycle locations and periods. This method, while simple, is not preferred due to its subjective nature and lack of reproducibility. Consequently, I use another simple (but non-subjective) method to identify cycling periods.

As Lewis (2010) notes, most stations will infrequently change their price if wholesale costs are constant. This makes the market's median price change is useful in detecting price cycles. If we tabulate the stations' daily price change in a non-cycling market, we expect stations with constant marginal costs to have a median daily price change of zero, absent any other market changes. In contrast, when a market exhibits price cycles, the stations' price changes are motivated not by a change in wholesale costs or market demand, but the market's location in the price cycle. Due to the asymmetric nature of the price cycle, where market restorations occur quickly and the undercutting phase takes place over a longer period, stations spend most of their time undercutting each other. As a result, most price changes will be negative, and the median price change in the market will tend to be negative.

I calculate the median price change by brand by month. Brands with a median monthly price change of less than -0.1 cents are considered to be cycling that month. This method of

39 detection works well to filter the periods of cycling and non-cycling in a reproducible manner.

Figure 3 shows the market average price for Shell in October and November 2003. Shell’s monthly median price change for these months is -0.336 and -0.402, respectively. Figure 4 again shows the market average price for Shell stations, but this time for October and November 2004.

In this case, cycles are clearly not present, and the median monthly price change for both months is 0. For the remainder of this paper, I use the monthly median brand price change method to identify periods of cycling and non-cycling. I calculate the median price change for each brand for each month in my data. The average median price change by year can be seen in Figure 5.

This figure provides a broad overview of brand cycling behavior over time. Recall, I use a median price change of -0.1 to indicate cycling over a specific time period. Using this figure, we can visually see when brands exit the price cycle. I find that Shell is the first to exit the price cycle in April 2004. BP and Caltex maintain regular cycles until 2009, at which point cycles become irregular, but they do not completely disappear until the end of 2011.

Figure 3: Average Price of Shell-Branded Stations October 2003 – December 2003 Prior to Cycle Exit

Figure 4: Average Price of Shell-Branded Stations October 2004 – December 2004 after Conversion to Coles Express Stations

Figure 5: Average Yearly Median Price Change by Brand

Note: Average yearly median price changes greater than -0.1 indicate non-cycling periods.

4. Implications of Edgeworth Price Cycles for Consumers

Due to the cyclical nature of price cycles, patient and well-informed consumers can intertemporally shift their purchases to periods of relatively low prices. This practice of

“gaming” price cycles is encouraged by government officials in Western Australia, and consumers in this market are aided by their access to future prices on the Fuelwatch website,17 which even sends out e-mail announcements of impending price hikes, making it easier for consumers to identify the bottom of each price cycle.

17 See < http://www.fuelwatch.wa.gov.au/fuelwatch/pages/public/contentholder.jspx?key=priceCycle.html >. Fuelwatch estimates drivers can save up to 20 cents a liter on average. 41

The predictability of the price cycles and the availability of some information about future prices mean that one cannot just compare the average level of prices during cycles and after cycles. Ideally, the transaction-weighted price would be used to determine whether consumers are paying more or less for fuel. Unfortunately, since I do not observe the quantity sold at each price (or any quantity at all), I can only calculate the average price over a period of time. Consequently, my results must be interpreted carefully.

We can speak to the welfare effects of Edgeworth price cycles under some reasonable assumptions. Let us assume there are two basic types of consumers, those who are patient and able to adjust the timing of their purchases strategically, and those who are impatient and purchase fuel non-strategically. Further assume that stations’ costs are a constant c, and the price cycle is stable as in the original theoretical formulation of Maskin and Tirole (1988). Cycles have a constant amplitude of pmax – pmin, and average market price (over time) is pavg. If we allow the purchases of impatient consumers to be uniformly distributed among all days of each cycle, we can expect them to purchase fuel at the average transaction price of pavg. Patient consumers should pay on average a price between pmin and pavg, as they are able to shift their purchases towards days that tend to be cheaper. If patient consumers can control their purchases completely (for example, costless storage), then they will only pay pmin.

If we take the market described above and disrupt the equilibrium so that stations coordinate around a focal price equal to pavg, the two classes of consumers will be affected differently. Impatient consumers will not, on average, be affected since they were already paying an average of pavg for fuel. Patient consumers, however, will be strictly worse off because they can no longer shift consumption to days when stations sell cheaper fuel. The implication of this simplistic analysis is that provided the cycles do not cause the average market

42 price to increase, consumers are unambiguously better off if at least some portion of consumers is able to strategically shift their purchases to minimize purchase cost.

5. The Collapse of Edgeworth Price Cycles

5.1. Industry Changes and Early Cycle Disruption

On March 15, 2004, the Coles Group purchased pricing control rights to the majority of

Shell stations in the Perth metropolitan area. The intent of this purchase was to use fuel as a loss-leader for grocery sales.18 Before pricing control was transferred to the Coles Group, Shell stations often acted as a price leader as well as a quick and reliable price follower.19 After the conversion to Coles Express retailers, these stations not only left the cycle, but also disrupted the entire cycle itself. Wang (2013) finds that in the short-run, these Coles Express stations price aggressively, although their prices eventually increase in the long-run.

The increase in fuel-grocery bundling was a major industry change in Australia from the

1990s until last year, when the ACCC (Australian Competition and Consumer Commission) came to a voluntary agreement with retailers to limit the size of these discounts. It is due to this large industry-level change that gasoline retailers are an ideal control group for LPG, enabling me to isolate the effect of the change in cycling behavior from the overall changes in the fuel market occurring during this time period.

5.2. LPG Retailers Permanently Exit the Cycle in January 2012

Figure 6 presents LPG and gasoline prices. The sets of series are very different, despite the fact that the sellers participate in both markets. I compare January 2012 to December 2012.

18 See Wang (2013) 19 See Wang (2009) and Chapter 1 43

Figure 6: Comparison of Average LPG and ULP Market Prices January 2012 – January 2013

Prior to January 2012, both LPG and gasoline exhibit the sawtooth-edged pattern in margins that are expected in price changes driven by Edgeworth price cycles and not changes in cost. This contrasts with the same period a year later after cycles have collapsed. From May 1,

2012 to June 1, 2012, gasoline margins continue to exhibit the sawtooth pattern characteristic of price cycles, while the margins for LPG do not. Similarly, I also compare LPG and gasoline price and margins in January 2012, the month cycles collapse permanently in the LPG market. It is clear that while both fuels experience similar trends in the evolution of wholesale marginal costs, the pricing strategies in the two markets diverge significantly.

Although it is not clear what triggered the collapse of price cycles in LPG, there was likely some change in the retail fuel market that affected its level of competitiveness. News reports have also carried anecdotal evidence supporting the idea that the retail market for LPG was not very competitive in the Perth area. For example, in March 2009, when price cycles in

LPG were temporarily disrupted, WAtoday (a Western Australian newspaper) wrote,20 “… Perth retailers were also using FuelWatch to signal each other on price rises, with other service stations

20 http://www.watoday.com.au/wa-news/perth-motorists-pay-the-most-for-lpg-20090309-8swb.html

‘quickly’ following Caltex’s move.”21 Since the LPG market had been transitioning between cycling and non-cycling for years prior to January 2012, it is likely even a small change affecting competition could tip the market out of the price cycles.

6. Empirical Analysis of Cycling Behavior on Average Price-cost Margins

This section analyzes the effects of two periods of cycle destabilization – the first in

March 2004 when the majority of Shell stations were converted to Coles Express stations, and the second in January 2012 when all LPG retailers exited the cycle. As a robustness check, I carry out this analysis in multiple timeframes and find similar results regardless of the timeframe used. My analysis proceeds as follows: I first consider the transition of the majority of Shell stations to Coles Express stations, which then stop participating in the cycle; I then perform a simple analysis comparing the changes in LPG prices over time; lastly, I use a difference-in- difference framework to control for industry-wide market changes.

Figure 7: Average Price and Monthly Price Change of Shell-branded Stations During Conversion to Coles Express Stations

21 This bears resemblance to the outcome of Albaek and Møllgaard’s (1997) paper entitled, “Government-Assisted Ologopoly Coordination? A Concrete Case” 45

6.1. Changes in Coles Express Station Pricing after the Conversion from Shell Stations

While it is not clear what market change caused the permanent destabilization of price cycles in January 2012, the temporary collapse of price cycles in 2004 is instigated by the conversion of

Shell stations into Coles Express stations. Since the purpose of this conversion was to use the new Coles Express stations as loss-leaders for grocery sales,22 we would expect that this conversion to cause a decrease in the price charged by these stations. All else equal, we would also expect to observe a similar pattern in competing stations in response. However, because the

Coles Express stations were previously price leaders, the Edgeworth price cycle was temporarily destabilized. To analyze the short-term effect of the conversion of Shell stations to Coles

Express stations and the subsequent brief cessation of cycles, I fit the following model at the station level for the period of June 2003 to the end of May 2004:

퐵 퐽=퐵

푝푟푖푐푒푖푡 = 훽0 + 훽1푎푓푡푒푟푡 + ∑ 훽푏 퐵푟푎푛푑퐷푢푚푚푦푖 + ∑ 훽푗퐵푟푎푛푑퐷푢푚푚푦푖 ∗ 푎푓푡푒푟푡 푏=1 푗=0

푀 푌

+ ∑ 훽푚푀표푛푡ℎ퐷푢푚푚푦푚 + ∑ 훽푦푌푒푎푟퐷푢푚푚푦푦 + 푢푖푡 푚=1 푦=1

퐵 퐽=퐵

푚푎푟푔푖푛푖푡 = 훽0 + 훽1푎푓푡푒푟푡 + ∑ 훽푏 퐵푟푎푛푑퐷푢푚푚푦푖 + ∑ 훽푗퐵푟푎푛푑퐷푢푚푚푦푖 ∗ 푎푓푡푒푟푡 푏=1 푗=0

푀 푌

+ ∑ 훽푚푀표푛푡ℎ퐷푢푚푚푦푚 + ∑ 훽푦푌푒푎푟퐷푢푚푚푦푦 + 푢푖푡 푚=1 푦=1

I define four specifications. Specifications (1) and (2) do not have any station or time controls, while station and month fixed effects are added for specifications (3) and (4). The results can be found in Table 3. This specification allows an examination of how the change in pricing strategy

22 See http://www.colesexpress.com.au/about-us.aspx 46 affects both the Coles Express stations and their competitors. Of particular interest are the coefficients on the interaction term between after and Shell, Coles Express. I find that the effect of the conversion and price cycle collapse affected Shell stations differently depending on whether they were converted into Coles Express stations or continued to be solely Shell branded.

The interaction terms in the model designate the brand-specific effect of the conversion after

March 15, 2004. The stations that rebranded as Coles Express on March 15, 2004 have lower price-cost margins in all specifications.23 After controlling for time and station effects, this decrease in margin is quite large. This is in contrast to the group of Shell stations that did not rebrand, whose price-cost margins increased by an average of 7.1 cents in specification (2) but showed no statistically significant change in specification (4). Overall, margins decrease after the converted Shell stations disrupt the cycle by approximately 6 cents, and in the full specification, both BP and Caltex do not experience any statistically significant change in margins. This is consistent with the use of Coles Express branded stations as a loss-leader for groceries.24

Despite the fact that converted stations charge lower prices and enjoy lower price-cost margins, a similar pattern is not seen in competing stations. Despite facing a new, relatively low-price competitor, these competitors enjoy price-cost margins on average. This suggests market participants are better off under a regime of focal price coordination despite the appearance of an aggressive competitor.

23 Note that this is relative to the price-cost margins of the same group of stations, not relative to all Shell stations. 24 It is likely the average transaction margin was even lower, as the grocery-fuel discounts are not reflected in the data. 47

Table 3 Regression Results June 2003 - June 2004 (1) (2) (3) (4) Station FE No No Yes Yes Time FE No No Yes Yes price margin price margin

after -3.266*** -2.577*** -6.063*** -6.063*** (-26.87) (-21.25) (-94.89) (-94.89) Brand Fixed Effects BP 3.958*** 3.971*** 8.257 8.602 (41.25) (41.18) (0.00) (0.01)

Caltex 3.766*** 3.787*** -1.183 -0.960 (40.78) (40.80) (0.00) (-0.00)

Shell 7.179*** 7.186*** 19.31 19.53 (52.76) (52.38) (0.00) (0.02)

Shell/Coles Express 5.732*** 5.739*** 21.97 22.19 (39.24) (39.13) (0.00) (0.02)

BP*after 0.478* 0.470* 0.234*** 0.234*** (2.23) (2.19) (4.33) (4.33)

Caltex*after 0.957*** 0.940*** 0.362*** 0.362*** (4.60) (4.52) (6.54) (6.54)

Shell*after 0.695 0.696 0.436*** 0.436*** (2.33) (2.32) (4.87) (4.87)

Coles*after -0.750** -0.748** -23.85*** -23.85*** (-2.27) (-2.27) (-56.12) (-56.12)

Constant 40.75*** 19.20*** 37.94 17.82 (722.64) (338.10) (0.00) (0.02)

N 126631 126631 126631 126631 R-sq 0.044 0.040 0.938 0.938 F 642.7 579.7 4422.4 4429.2 Robust t-statistics in parentheses. *p<0. 05 **p<0.01 ***p<0.001

6.2. Changes in Market Pricing after the End of Edgeworth Price Cycles

I perform a simple analysis of the change in station pricing behavior and observed price-cost margin after all stations exit the cycle in January 2012. I fit the following models comparing the cycling periods of 2004 to 2012, 2004 to 2006 and 2012, and June 2011 to May 2012.25

푀 푌

푝푟푖푐푒푖푡 = 훽0 + 훽1푎푓푡푒푟푡 + ∑ 훽푚푀표푛푡ℎ퐷푢푚푚푦푚 + ∑ 훽푦푌푒푎푟퐷푢푚푚푦푦 푚=1 푦=1

푆

+ ∑ 푆푡푎푡푖표푛퐹퐸푠푖 + 푢푖푡 푠=1

푀 푌

푚푎푟푔푖푛푖푡 = 훽0 + 훽1푎푓푡푒푟푡 + ∑ 훽푚푀표푛푡ℎ퐷푢푚푚푦푚 + ∑ 훽푦푌푒푎푟퐷푢푚푚푦푦 푚=1 푦=1

푆

+ ∑ 푆푡푎푡푖표푛퐹퐸푠푖 + 푢푖푡 푠=1

The results can be found below in Table 4. Specifications (1) – (6) include only LPG data, while specifications (7) – (12) replicate the previous specifications using ULP data.

25 For robustness, I also fit models using just 2005 and 2006 individually as the before period. The results are qualitatively similar. 49

Figure 8: Average LPG Market Price July 2004 – July 2012

Qualitatively, all the LPG specifications (1) to (6) show that price and margin increase in the post cycling period relative to the cycling period. My results show that the price-cost margins in 2012 tend to be over 6 cents per liter higher than they were during cycling periods in

2012 real Australian cents. Specifications (5) and (6) cover only eleven months, a much shorter period than (1) – (4). Seven of those months exhibit cycles while four do not. In this case, my results indicate that the average price-cost margin increased by 16.63 cents, even after accounting for time and station fixed effects. In the related gasoline market, stations also appear to enjoy an increase in average price-cost margins in 2012 relative to 2004 through 2006, but it is difficult to interpret the magnitude of the change, since the marginal cost indicators in the data for LPG and gasoline originate from different portions of the value-chain. My results indicate that gasoline sellers posted on average lower prices with lower price-cost margins than before price cycles ended in the LPG market.

While these results support the hypothesis that the end of price cycles in LPG marked a movement towards a less competitive equilibrium (despite the absence of the obvious tacit collusion in the form of price cycles), it is important to note that these results only include general time and station fixed effects. While the time fixed effects may capture some market

50 change that can affect price and margins in principle because the same stations selling LPG also sell gasoline, gasoline prices can be used to isolate the effect of cycles on stations separate from other events over time that may affect competition between stations, which sell both fuels. To utilize this unique control, I use a difference-in-difference framework.

6.3. Employing a Difference-in-difference Model to Control for Market Changes Over

Time

While my preliminary analysis indicates that margins rise at an absolute level, it is not clear what is driving that increase. As I only observe price and wholesale benchmark data, it is likely that there are many other factors that affect the price-cost margin of posted prices that I do not observe. For example, there could be additional cost changes outside wholesale prices that are location specific, or other factors such as grocery discounts, convenience stores, and repair centers that are affecting competition in addition to the change in pricing strategies.

Table 4 Regression Results: Price and Margin Changes after Price Cycles End Fuel Type LPG ULP Fixed Effects Time and Station Period CY 2004 | CY 2012 CY 2004-06 | CY 2012 June 2011 - May 2012 CY 2004 | CY 2012 CY 2004-06 | CY 2012 June 2011 - May 2012 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Dep. Var. price margin price margin price margin price margin price margin price margin

After 28.34*** 6.025*** 28.28*** 5.961*** 9.680*** 16.63*** 28.96*** 2.709*** 26.24*** 1.111*** -0.0721 -1.331*** -451.29 -95.93 -440.69 -92.89 -36.94 -63.46 -713.56 -66.93 -582.27 -24.59 (-1.87) (-33.98)

Constant 48.28*** 18.90*** 48.28*** 18.90*** 85.21*** 41.16*** 116.0*** 1.970*** 116.0*** 1.934*** 143.1*** 5.210*** -111.03 -43.46 -111.06 -43.47 -466.74 -225.47 -807.33 -15.11 -809.33 -14.88 -1327.91 -49.45

N 257421 257421 502086 502086 121591 121591 447230 447230 886635 886635 197698 197698 R-sq 0.945 0.902 0.936 0.902 0.931 0.919 0.921 0.841 0.934 0.848 0.859 0.847 Note: t statistics in parentheses * p<0.05; ** p<0.01; *** p<0.001

To control for these unobservable changes, I take advantage of the fact that the same stations selling LPG also sell gasoline and that the cycle continues in the gasoline market. This

51 allows me to compare prices and margins between these two products, which operate in almost identical environments.

Price and price-cost margin for LPG and gasoline are in Australian cents per liter. Since

LPG and gasoline are sold in liters to consumers in this market, these prices and margins represent unit prices and price-cost margins. While a liter of LPG is not equivalent to a liter of gasoline due to the lower energy content of LPG,26 I do not scale the LPG data because it would misrepresent the price-cost margin results, as these already incorporate the wholesale costs of

LPG, which tend to be much lower than those of gasoline. That a price differential between LPG and gasoline exists is not surprising and not the focus of this study. I instead focus on the relative changes in price and margins. Even this is a conservative approach due to gasoline's higher average price, as proportionally equal increases in price and price-cost margin will be larger in the gasoline data in absolute terms. This effectively biases downward the coefficient of interest.

There are a few downsides to using the difference-in-difference framework in this context. First, given the dynamic nature of price cycles and the widespread public knowledge of them, it is likely that at least some consumers strategically time their fuel purchases when prices are predictably lowest. In this case, the benefit to consumers may not be lower prices on average

(since minimum cycle prices may be lower than the new market equilibrium), but the fact that consumers no longer incur costs to synchronize their buying patterns with the cycle. This issue cannot be resolved without quantity data, although it is discussed below.

Another issue is that the propensity to cycle in a market is undoubtedly endogenous. If the markets for gasoline and LPG are so similar (to the point where they are largely serviced by

26 See Chapter 1 for a discussion on this. 52 the same retailers), why would cycles cease in one market but not the other? If the reason is correlated with LPG margins but not gasoline margins, there is an endogeneity problem.

Although it is not possible to control for this with the data available, I can examine the reasons why LPG margins alone are affected, and make a determination on the likelihood of each scenario affecting my results. One major difference between gasoline and LPG is that, while it was common for retailers to offer gasoline discounts for grocery bundles (as described in section

4.1), the availability of these discounts for LPG was inconsistent. Wang (2013) notes that where retailers are not horizontally integrated, a pricing externality occurs between the two. This could bias downwards the difference between gasoline and LPG, which if extreme enough would cause a type 1 error. As seen in the below results, I am still able to detect a statistically significant difference in prices despite this potential bias.

I estimate several variations of the following models:

푝푟푖푐푒푖푡 = 훽0 + 훽1퐿푃퐺퐷푢푚푚푦푖 + 훽2퐴푓푡푒푟푡 + 훽3퐴푓푡푒푟 ∗ 퐿푃퐺퐷푢푚푚푦푖푡

푀 푌 푆

+ ∑ 훽1푚푀표푛푡ℎ퐷푢푚푚푦푚 + ∑ 훽2푦푌푒푎푟퐷푢푚푚푦푦 + ∑ 푆푡푎푡푖표푛퐹퐸푠푖 + 푢푖푡 푚=1 푦=1 푠=1

푚푎푟푔푖푛푖푡 = 훽0 + 훽1퐿푃퐺퐷푢푚푚푦푖 + 훽2퐴푓푡푒푟푡 + 훽3퐴푓푡푒푟 ∗ 퐿푃퐺퐷푢푚푚푦푖푡

푀 푌 푆

+ ∑ 훽1푚푀표푛푡ℎ퐷푢푚푚푦푚 + ∑ 훽2푦푌푒푎푟퐷푢푚푚푦푦 + ∑ 푆푡푎푡푖표푛퐹퐸푠푖 + 푢푖푡 푚=1 푦=1 푠=1

Where priceit and marginit is the price and margin, respectively, of station i at time t.

LPGDummyi is a binary variable that takes the value of 1 if the observation is associated with the

LPG market and 0 if it is associated with the gasoline market. Aftert takes the value of 1 if cycles are not present in in the LPG market at time t, and Aftert*LPGDummy is an interaction term between LPGDummyi and Aftert. I also include monthly and yearly fixed effects to account for

53 short- and long-term unobserved demand fluctuations. The interpretation of this model is that B1 will represent the time-invariant difference between LPG and gasoline, B2 will identify the product-agnostic difference in price and margin between the two periods (cycling and non- cycling), and B3 will be the change in LPG offer price and margin relative to gasoline. A negative value suggests the change in cycling behavior caused the LPG retail market to become more competitive while a positive value suggests the opposite.

To ensure the robustness of my results, I fit the above model for three separate periods, the results of which can be found in Table 5. Specifications (1) and (2) compare the years 2004 and 2012, specifications (3) and (4) compare the years 2004, 2005, and 2006 with 2012, and finally specifications (5) and (6) compare the time period of June 1, 2011 to December 31, 2011 to the time period of January 1, 2012 to June 1, 2012. Qualitatively, the results for all six specifications are similar. Since cycles were unstable in 2011, the value of the variable I am interested in, After*LPG_Dummy, is biased downwards because June 1, 2011 to December 31,

2011 contains periods of cycling and non-cycling behavior. Despite this, there is still a statistically significant difference in the price-cost margin between the second half of 2011, during which cycles were present (but unstable), and the first half of 2012, during which they were not present at all. For the remainder of this analysis, I focus on specifications (3) and (4).

As in the previous specifications, I find that LPG tends to have significantly lower prices than gasoline. Despite this, margins on LPG tend to be much higher, by 24 cents on average, even before the collapse of the fuel price cycle. I find that for both fuels, prices and margins increase in 2012 after accounting for changes in taxes and inflation. However, the increase in price and margin for LPG is 13.22 and 7.6 real 2012 Australian cents greater than gasoline, respectively.

Consumers in this case are made unambiguously worse off. Not only are average prices higher,

54 but the cyclical nature of price cycles allowed informed and patient consumers to change the timing of their purchases. Without cycles, consumers no longer have the ability to time their purchases to minimize their fuel costs.

These findings suggest price cycles made the buyers of LPG in this market better off despite public belief of the opposite. Maskin and Tirole’s (1988) model allows for both focal point and price cycle coordination. Although they do not make any comment on which equilibrium is more likely, it seems that the price cycle equilibrium can, under some circumstances, be preferred to the focal point equilibrium.

Table 5 Regression Results: Difference-in-Difference Price and Margin Changes in LPG Relative to ULP Before-After Period CY 2004 | CY 2012 CY 2004-06 | CY 2012 6/1 to 12/31/11 | 1/1 to 6/1/12 Fixed Effects Time and Station (1) (2) (3) (4) (5) (6) Dependent Variable price margin price margin price margin

After 27.69*** 3.465*** 25.82*** 4.002*** -1.295*** -2.902*** -588.91 -76.63 -485.56 -87.49 (-30.49) (-65.67)

LPG_Dummy -68.30*** 22.59*** -73.62*** 24.27*** -71.88*** 25.97*** (-3349.00) -1204.6 (-4754.38) -2292.6 (-2515.40) -944.13

After*LPGDummy 7.791*** 9.196*** 13.22*** 7.579*** 10.20*** 4.739*** -266.43 -330.59 -496.34 -317.99 -277.29 -125.02

117.7*** 1.538*** 119.6*** 0.986*** 144.3*** 4.906*** Constant -760.03 -9.57 -765 -6.15 -1123.2 -37.45 N 704651 704651 1388721 1388721 319289 319289 R-sq 0.979 0.917 0.975 0.919 0.981 0.917 Note: t statistics in parentheses, calculated using Huber-White Standard Errors * p<0.05; ** p<0.01; *** p<0.001

6.4. Changes in Market Pricing After Shell’s Exit in 2004

I fit the model outlined in section 5.3 including all stations except Shell-branded stations, designating January 2003 to March 2004 as the pre-collapse period and April 2004 to the end of

2005 as the post-collapse period. The results can be found in Table 4. As expected, I find that

55 stations charge significantly less per liter for LPG than gasoline.27 The price-cost margin on

LPG increases by approximately 1 cent relative to gasoline after Shell stops participating in the price cycle. This result suggests that once Shell stopped cycling, the rest of the market benefited from an increase in price-cost margins.

7. Discussion: Potential Causes for Brands Exiting the Cycle and the Effect on Consumers

The motivation for brands to exit the cycle may differ. In the case of Shell branded stations, it appears that the conversion to Coles Express and subsequent bundling of fuel and groceries decreased the incentive to participate in the cycle. As detailed in Wang (2013), it does not appear that the behavior of these stations was to maximize profit of the fuel station itself, but to use fuel as a sort of loss-leader for their grocery operations.

The temporary disruption caused by Shell’s exit is not unexpected, but as seen in Figures

5 and 7, regular cycles were later reestablished. The question then remains, what caused the other brands to eventually abandon the regular price cycle by January 2012? Previous literature on Edgeworth price cycles provides a few hints. Noel (2008) simulates price cycles and finds them robust to many market changes. One thing that could cause the collapse of the Edgeworth price cycle is an increase in differentiation between stations. As I will argue in chapter 3, the opposite has most likely occurred in the Perth metropolitan area, and in any case, there is no reason why differentiation between stations would increase in LPG but not gasoline. Noel (2008) also investigates other market characteristics that may either cause the cycle to decay or make the label cycle seem like a misnomer. He notes that in the case of tight capacity constraints price

27 Since LPG has poorer fuel economy (per liter) than gasoline, the main advantage of using the fuel is its lower cost-per-mile. 56 cycles can be eliminated and, in the case of asymmetric capacity constraints, where a “hyper- cycle” can be created, which appears in the form of multiple focal point equilibria.

Other literature suggests other explanations for the collapse of price cycles. Maskin and

Tirole’s (1988) original formulation made the simplifying assumption that only two firms existed in the market. As the number of firms grows, the coordination problem between stations grows as well. In markets that cycle, this coordination problem has been solved in a few ways. In

Wang (2008), firms illegally communicated the timing of price hikes. In Wang (2009) and

Lewis (2012), stations are coordinated by the brands they carry, reducing the effective number of players in the market. One possible explanation for the disappearance of price cycles in the

Perth LPG market is the brands’ inability to coordinate stations carrying their fuel. This explanation seems unlikely, as these very same stations continue to cycle in the gasoline market carrying the same brands.

A decrease in the overall competitiveness of the LPG retail market may also be responsible for the disappearance of price cycles. As I observe the entire set of station prices, I also observe the entire set of brand names offering fuel in the retail market. I calculate the number of different brands28 selling fuel in each market from 2009 to 2012 and report the results in Figure 9. In 2009, the number of brands offering LPG was greater than the number offering gasoline, most likely due to the existence of small but branded LPG-only retailers. By 2012, the number of brands offering LPG falls from a peak of 17 to 14, while the number of brands offering gasoline increases. If the decrease in brands in the marketplace made it easier for stations to coordinate, this is a likely cause for the disappearance of cycles in this market.

28 In order to be considered a brand, the brand must operate at least two stations. Independents (operators with only one station) are considered one brand as a group. 57

Figure 9: Number of Brands Offering LPG and Gasoline by Year 2009 – 2012

8. Conclusion

This paper examines a change in the market equilibrium of LPG in the Perth metropolitan area. My results indicate that the strategy shift from Edgeworth price cycles to some stable price-point (that only drastically changes when marginal input costs change) leads to higher overall price-cost margins, making consumers unambiguously worse off since they not only face higher average price-cost margins, but are no longer able to engage in intertemporal substitution to take advantage of the trough of each cycle.

While it is not clear what caused the shift in strategies, by using gasoline as a control group along with a series of time period dummies, I am able to account for much variation in market conditions, isolating the effect that caused the shift in strategies from the change in price- cost margins caused by the shift in strategies. This paper adds to the literature supporting the idea that Edgeworth price cycles are an intermediate form of competition, and that a shift to (or from) price cycles alone cannot indicate whether the market has become more or less competitive. As a result, regulatory authorities must be careful to analyze the effect price cycles have on variables such as price-cost margin to determine whether the shift indicates an increase or decrease in competition, despite the fact that Edgeworth price cycles are clearly a form of tacit

58 collusion. In the case of the Perth retail fuel market, while I am unable to determine the cause of the shift in behavior, I can say that it appears that the LPG market became less competitive, and anecdotal evidence suggests stations were utilizing the regulatory mechanisms (the 24-hour rule and Fuelwatch) to enforce cooperation. It is likely this paper underrepresents the effect on consumers, since they can no longer shift purchases to days with the lowest expected price.

Future work in this area would benefit greatly from the inclusion of quantity data and station characteristic data. One striking feature of this market is that some stations participate in the cycles while others do not. Station characteristic data would shed further light on the incentives involved in cycle participation. Quantity data would also be very valuable, particularly in regulatory settings, since it is possible that patient consumers will be better under a price cycle regime even if average prices are higher. Quantity data could also be used to calculate cross-price elasticities such as Wang (2009) and estimate what proportion of consumers are patient and what proportion are not.

References

Atkinson, Benjamin. "Retail Gasoline Price Cycles: Evidence from Guelph, Ontario Using Bi- Hourly, Station-Specific Retail Price Data." The Energy Journal 30.1 (2009): n. pag. Web.

Castanias, Rick, and Herb Johnson. "Gas Wars: Retail Gasoline Price Fluctuations." The Review of Economics and Statistics 75.1 (1993): 171-74. Print.

Doyle, Joseph, Erich Muehlegger, and Krislert Samphantharak. "Edgeworth Cycles Revisited." Energy Economics 32.3 (2010): 651-60. Print.

Eckert, Andrew. "Retail Price Cycles and Response Asymmetry." Canadian Journal of Economics/Revue Canadienne D`Economique 35.1 (2002): 52-77. Print.

Lewis, Matthew, and Michael Noel. "The Speed of Gasoline Price Response in Markets with and without Edgeworth Cycles." Review of Economics and Statistics (2010): 110301164542093. Print.

Lewis, Matthew S. "Price Leadership and Coordination in Retail Gasoline Markets with Price Cycles." International Journal of Industrial Organization 30.4 (2012): 342-51. Print.

Lewis, Matthew S. "Temporary Wholesale Gasoline Price Spikes Have Long‐Lasting Retail Effects: The Aftermath of Hurricane Rita." The Journal of Law and Economics 52.3 (2009): 581-605. Print.

Maskin, Eric, and Jean Tirole. "A Theory of Dynamic Oligopoly, II: Price Competition, Kinked Demand Curves, and Edgeworth Cycles." Econometrica 56.3 (1988): 571. Print.

Noel, Michael D. "Edgeworth Price Cycles, Cost-Based Pricing, and Sticky Pricing in Retail Gasoline Markets." Review of Economics and Statistics 89.2 (2007): 324-34. Print.

Noel, Michael. "Do Retail Gasoline Prices Respond Asymmetrically to Cost Shocks? The Influence of Edgeworth Cycles." The RAND Journal of Economics 40.3 (2009): 582-95. Print.

Wang, Zhongmin. "(Mixed) Strategy in Oligopoly Pricing: Evidence from Gasoline Price Cycles Before and Under a Timing Regulation." Journal of Political Economy 117.6 (2009): 987-1030. Print.

Wang, Zhongmin. "Station Level Gasoline Demand in an Australian Market with Regular Price Cycles." Australian Journal of Agricultural and Resource Economics 53.4 (2009): 467- 83. Print.

Wang, Zhongmin. "Supermarkets and Gasoline: An Empirical Study of Bundled Discounts." Resources for the Future. Working Paper. (n.d.): n. pag. Web. 2013.

Zimmerman, Paul R., John M. Yun, and Christopher T. Taylor. "Edgeworth Price Cycles in Gasoline: Evidence from the United States." Review of Industrial Organization (2013): 1- 24. Print.

Chapter 3: Spatial Competition and Edgeworth Price Cycles

1. Introduction

The purpose of this paper is to examine the effects of a novel pricing regulation implemented in Western Australia’s Perth metropolitan area. This law aimed to reduce price volatility and disrupt the tacit collusion (in the form of Edgeworth price cycles) between fuel retailers. To achieve these goals, the law significantly restricted retailers’ pricing behavior and decreased search costs by limiting stations’ ability to set price to once a day and by posting market fuel prices on a public website. This paper’s contribution, which, to the author’s knowledge, is unique in the literature, is to examine the effect of this regulation on the scope of spatial competition in the Perth metropolitan market for LPG29 and to investigate the role of spatial product differentiation in markets exhibiting Edgeworth price cycles.

In the past decade, the literature has seen a large increase in empirical studies on

Edgeworth price cycles in retail fuel markets. This increase is partially due to new sources of rich, high-frequency price data that allow researchers to accurately track daily and even intra- daily prices. The majority of these studies have focused on identifying where Edgeworth price cycles exist, documenting their characteristics, and investigating how the relatively unconcentrated retail fuel market can support price cycles, typically considered evidence of tacit collusion. This study’s goal is to extend the empirical literature with an investigation of how regulation action affects the scope of spatial competition and, consequently, cycling behavior.

To preview the results of this paper, I find evidence that the law, colloquially called the

“24-hour pricing rule,” increases competition between nearby stations. These results are

 I would like to thank my committee - James Dana, John Kwoka, and Zhongmin Wang - for their help and advice. I would also like to thank Gustavo Vicentini and the participants of the IO lunch for their helpful comments. 29 Wang (2009a) examines the effect of the law on price leadership allocation in the related gasoline market. 62 consistent with decreased consumer search costs associated with the 24-hour pricing rule. I use prices to construct a measure of how intensely nearby firms compete, and find that over time firms appear to be competing with firms that are geographically further away. Despite this increase in the scope of geographic competition, I find no evidence that the law reduced the volatility of prices, nor do I find any evidence of a decrease in the average amplitude of the cycle. Nominally, the regulation was unsuccessful in achieving its stated objectives of reducing inter-daily price volatility and preventing tacit collusion between fuel retailers.

1.1. Market Background

Automotive fuel in Western Australia is sold to consumers from a number of retail stations. The ownership and pricing structure of these stations vary by location, although most stations30 carry branded fuel from one of the major petroleum firms: Ampol/Caltex, Shell, Mobil,

Gull, or Peak. Stations can be independently owned, owned by a smaller chain such as

Woolworths Plus (which is also a grocery chain), or owned by one of the large fuel brands.31

With 135 stations in the prelaw data and 214 stations in the postlaw data (the latter including almost all fuel stations in the Perth metropolitan area), a high level of competition would be expected; however, price differentials between urban and rural areas and frequent market-wide fluctuations in price concerned both consumers and regulators, leading to the 24-hour rule and the Fuelwatch web service.

1.2. The 24-Hour Rule

The retail markets for LPG, gasoline, and diesel fuel in Western Australia are regulated in a fairly novel fashion. Retailers in these markets are only permitted to change their prices once

30 93.33% of stations observed in the prelaw period and 83.3% in the postlaw period 31 Many stations carrying the major brands’ fuel are franchisees. The Petroleum Retail Marketing Sites Act 1980 prevents integrated fuel producers from owning a large number of stations. 63 every 24 hours at 6 a.m. and must report the new prices to the government by 2 p.m. on the preceding day. The government posts these prices on a consumer information website called

Fuelwatch before they become valid. This website not only allows consumers to observe the set of current and future prices, but also sends out automated price reports and warnings of future price increases.

The purpose of the 24-hour rule and the Fuelwatch consumer website is to increase the transparency of retail fuel markets in Western Australia with the hope of reducing margins and eliminating intra-daily price fluctuations.32 While it is clear that the 24-hour rule will reduce intra-daily price volatility by fiat, this paper will test whether the law has any effect on inter- daily price volatility and average cycle amplitude.

2. Literature Review

This paper primarily extends the literature on previous studies of Western Australian fuel markets and the effect of the 24-hour rule. This market is notable not only for exhibiting multiple price cycles in a variety of fuels (including gasoline and LPG) but also for having accurate and comprehensive public pricing data available. Wang (2009a) examines the effect of the 24-hour rule on the retail gasoline market in the Perth metropolitan area. He finds that while the rule does not permanently disrupt the price cycles, it causes a major change in the way firms allocate price leadership. Prior to the law, one firm acted as the price leader. After the passage of the law, firms’ behavior can best be described as mixed strategies to determine price leadership, which is consistent with Maskin and Tirole’s (1988) theoretical formulation. This contrasts to Wang (2008), who finds that in Ballarat, gasoline firms solve the coordination

32 See “Background to Fuelwatch” 64 problem of price leadership via telephone calls instead of pure or mixed strategy. Finally, Wang

(2009b) finds that own price elasticities for stations selling gasoline after the 24-hour rule can be quite high, with own price elasticities at some stations reaching -20.38.33 This paper extends this literature by examining the effect of the 24-hour rule on the spatial scope of competition in a related market, the retail market for LPG.

There have also been many papers studying other examples of price cycles around the world, most notably in Canada and the Midwestern United states. Many of these papers highlight the importance of space in the price cycle equilibrium. In Canada, Noel (2007) finds that the price cycles in Toronto are consistent with Edgeworth price cycles, and Atkinson (2009) finds that spatial proximity to areas with high levels of traffic is an important factor in firms’ abilities to control market price in the restoration phase. Eckert and West (2004) find that not only are the price cycles in Vancouver best described by Edgeworth price cycles, but that the undercutting phase of the cycle occurs more rapidly in areas of low concentration, and then spreads to the rest of the market over time. In the U.S., Doyle et al. (2010) find that firms with increased geographic (or other types) of differentiation have less incentive to enter price cycles. Lewis

(2012) finds that larger chains tend to lead price restorations; implying chains must be able to coordinate price increases over relatively large geographic areas to successfully lead market prices upwards. These papers stress the fact that firms compete most intensely with nearby stations and are predominately affected by the prices set by nearby competitors.

Noel (2008) performs computer simulations using Maskin and Tirole’s (1988) framework with some extensions.34 In particular, Noel (2008) finds that while price cycles appear to be

33 The minimum he finds is -5.64. 34 Noel extends the model to accommodate fluctuating costs, capacity constraints, differentiated products, and triopoly. 65 relatively robust to market changes (both aggregate price elasticity of demand and cross-price elasticity between competitors), the shape of the cycles can be used to gain insight on underlying market characteristics. While Noel’s (2008) results do not provide a way to directly calculate variables such as margin and elasticity from cycle characteristics, they do allow a general comparison between two distinct cycle series. I use Noel’s (2008) results to interpret the effect of the 24-hour rule by observing how the characteristics of the cycles change with the passage of the law.

This paper also draws from the literature on information and consumer search. With the introduction of the internet, it has become possible to empirically calculate the effect of changes in market information on market prices and price dispersion. For example, Brown and Goolsbee

(2000) look at the effect of insurance comparison websites on the market for term life insurance.

Consistent with search models like those found in Stahl (1989), the authors find that lower search costs cause market prices to fall, but price dispersion rises. Similarly, Orlov (2011) investigates the effect of the internet on price dispersion in the airline industry, postulating that the internet lowers consumers’ search costs.35 He finds that, as expected, average fares and the dispersion of prices between carriers fall. Paradoxically, intra-carrier price dispersion increases, and average fares charged by low-cost carriers actually increase. This paper looks at a similar change in which search costs are reduced through government regulation.

Finally, to investigate the change in the scope of spatial competition caused by an exogenous change in search costs, I utilize the spatial autoregressive model framework suggested by Pinske, Slade, and Brett (2002) and adapted to the gasoline market by Lee (2008) with a few modifications to account for the unique cyclical nature of this market.

35 Dana and Orlov (2009) also find that adoption of the Internet and subsequent increase in consumer information significantly increased capacity load factors. 66

3. Data

This paper utilizes four data sources to analyze the retail market for LPG in the Perth metropolitan area. The first two datasets contain station-level price, identity, and location data, and the third dataset contains wholesale cost information. I supplement these data with demographic information from the Australian Bureau of Statistics’ 2001 census. Using these data sources, I construct a dataset that enables me to analyze spatial competition and the effect of the 24-hour rule on this market.

3.1. Data Sources

The primary data source from the period before the law (henceforth referred to as the prelaw data) is a station-level panel dataset spanning the six months preceding the 24-hour rule’s implementation (July 1, 2000 to December 21, 2000) and the six months following the law’s implementation (January 3, 2001 to June 30, 2001). This dataset was acquired from the consulting firm Informed Sources (Australia) Pty. Ltd36 and contains hourly pricing data for many LPG and gasoline retailers in the Perth area as well as the rough station address37 and brand carried by that station. I calculate the average price each station charges by day in the data. Summary statistics and observed brand distribution are presented in Table 1.

Table 2 presents summary statistics and observed brand distribution for the data after the law comes into effect (henceforth referred to as the postlaw data). I obtain daily price data for the postlaw period from the Fuelwatch consumer protection website created by the 24-hour rule and maintained by Western Australia’s Department of Commerce. This dataset includes the daily station price, accurate station address and suburb, and station brand. Employing Google

36 I would like to thank Dr. Zhongmin Wang for this data, also used in Wang (2009a). 37 For example, one Shell station’s address is listed as “Stirling Highway.”

Maps and a variety of publically available geocoding services,38 I use the station addresses to determine the latitude and longitude of each station. After the passage of the law, the price cycle equilibrium collapsed for approximately a year. As a result, I treat January 3, 2001 to December

31, 2001 as an adjustment period.39

I augment these two primary datasets with two additional sources of information. First, I observe the confidential wholesale price of LPG paid by one multi-site BP franchisee from

January 9, 2002 to October 31, 2003. I use this data as a proxy for the marginal cost faced by the market. Second, I utilize the Australian Bureau of Statistics’ 2001 census for demographic information. This data is available at the very fine suburb level, allowing for variation in demographic information across stations. Not all suburbs contain all the demographic measures

I use in this paper. For suburbs missing some of the demographic data, I substitute the market average demographic value.

In total, I observe 43 cycles - 6 prelaw and 37 postlaw. Summary statistics for the characteristics of cycles before and after the law came into effect can be found in table 10. After the passage of the law, cycles tend to be shorter with larger amplitudes. Unlike Wang (2008), who finds that cycle length becomes less predictable under the law, the cycle length standard deviation in the prelaw period is almost double that of the postlaw period.

3.2. Matching Routine

The postlaw data from Fuelwatch contains very accurate addresses that can be easily converted into latitude and longitude values for spatial analysis. Unfortunately, the data from the

38 Batchgeo (http://www.batchgeo.com), GPS Visualizer (http://www.gpsvisualizer.com/geocoder/), Google Maps (http://maps.google.com) 39 During this time, the price cycle equilibrium is unstable, with cycles often disappearing and many failed restoration attempts. Since regular price cycles resume later, this period is not representative of the market’s eventual equilibrium. 68

prelaw period often contains less accurate and more generic address information. It is not

uncommon for a station to be identified only by brand, suburb, and street. In order to properly

geocode these stations, it was necessary to match them between the two primary sets of data.

Table 1 Summary Prelaw Statistics July 1, 2000 to June 30, 2001 Unique Stations: 135 Price Statistics and Brand Distribution Std. Observations % Cumulative % Minimum Maximum Average Dev

Brand: BP 14578 37.7 37.7 37.9 59.9 50 4.28 Ampol/Caltex 12236 31.7 69.4 37.5 66.4 50.7 4.04 Shell 4109 10.6 80.1 35.9 62.9 49.9 4.05 Mobil 3091 8.0 88.1 35.9 60.4 49.6 4.27 Gull 1904 4.9 92.9 38.5 58.9 48.9 3.63 Other 2709 7.0 100 37.9 54.9 48.9 3.19 Total 38627 100

Table 2 Summary Postlaw Statistics January 9, 2002 to October 31, 2003 Unique Stations: 214 Price Statistics and Brand Distribution Observations % Cumulative % Minimum Maximum Average Std. Dev

Brand: BP 30596 23.6 23.6 30.5 80.9 44.44 6.36 Ampol/Caltex 36525 28.1 51.7 30.5 99.5 44.24 6.44 Shell 18274 14 65.8 30.7 89.9 44.62 6.57 Mobil 10914 8.4 74.9 30.7 59.9 44.51 6.55 Gull 10317 7.9 82.1 31.9 59.5 43.49 5.91 Other 23202 17.9 100 30.4 99.9 42.92 6.18 Total 129828 100

This was accomplished in a number of ways. First, I matched stations with identical accurate street addresses.40 Then I matched stations at corresponding cross streets. Finally, I matched stations by comparing brand, suburb, and price series during the overlapping period between the datasets (January 1, 2001 to June 30, 2001). I was able to match a total of 135 stations between the two periods. While I observe 214 stations in the postlaw period, without location information for the 66 remaining stations in the prelaw period data, I must exclude them from the postlaw period analysis.41

Table 3 Comparison of Matched Stations with Total Sample Variable Observations Minimum Maximum Average Std. Dev Total Sample Suburb Population 196 9 22591 7302.214 5255.943 Median Age 189 24 45 34.59259 3.863103 Suburb UR 189 3.373016 17.97707 8.579437 3.085264 Vehicles per House 189 0.64 2.38 1.55963 0.3034805 Matched Sample Suburb Population 127 9 22591 7379.976 5420.573 Median Age 121 24 45 34.91736 3.95303 Suburb UR 121 3.373016 17.97707 8.603567 3.089543 Vehicles per House 121 0.81 2.14 1.543636 0.2778399

3.3. Comparison of Matched Sample with Overall Dataset

Before proceeding with my analysis, I compare the characteristics of the matched dataset to the entire dataset to ensure that my matching procedure does not result in an unrepresentative sample. Table 3 reports selected demographic measures for the unfiltered and matched data,

40 For example, one Kleenheat station is listed at 276 Leach Highway in both datasets. I can safely match these two stations without further confirmation of identity. 41 These excluded stations are still included in the calculation of weighted prices, described below. 70 while Table 4 reports the distribution of stations’ brands in each dataset. Both sets of stations have very similar demographic values, although Table 4 shows that the distribution of stations in the filtered dataset is slightly different than the distribution in the unfiltered dataset. Specifically, the larger chains are slightly overrepresented while smaller chains and independent stations are underrepresented. This is not unexpected, as stations belonging to smaller chains and independent stations were surveyed less frequently in the prelaw data, and often contain missing attributes like name, location, and address. However, since the larger chains comprise over 80% of the observed stations in the unfiltered dataset, it is unlikely that the slight overrepresentation of larger chains will cause significant bias in the results.

Table 4 Comparison of Matched and Total Sample Station Distribution Percent of stations in dataset BP Caltex Gull Independent Mobil Peak Shell Total Stations Matched Dataset 30.37 28.89 8.15 6.67 8.15 4.44 13.33 135 Total Dataset 23.36 27.57 7.94 11.68 7.94 7.94 13.55 214

4. Estimation Model and a Model of Spatial Competition

I follow the framework in Pinske et al. (2002), adopting the modifications in Lee (2008) for products sold to end users. I also borrow, with modifications, Lee’s (2008) notation.

4.1. Theoretical Model

Suppose there is a market with n stations that play a differentiated-Bertrand game selling n product variations. Each station sells one unique variant, the main differentiating characteristic of which is its location. The market contains h consumers, each of whom have an income of 푦̃푐 and purchase 푄 = (푞1푐, … , 푞푛푐), where 푞푖푐 ≥ 0, 푖 = 1, … 푛 and 푐 = 1, … ℎ. Consumers also purchase 푄0푐 of an outside good (defined as all other goods available outside the market). Like

Pinske et al. (2002) and Lee (2008), I define consumers’ indirect utility functions as a normalized quadratic indirect utility function with consumer c’s indirect utility function being:

푛 푛 푛 푛 푃0 푉̃푐(푃0, 푃푛, 푦푐) = − [∝0푐+ ∑ ∝푖푐 푃푛 − 푃0푦푐 (훾0 + ∑ 훾푖 푃푛) + ∑ 푃푗 (∑ 푃푖 푆푖푗푐)] 푖=1 푖=1 2 푗=1 푖=1

푛 −1 Where 푃0 is the price of the outside good which I normalize to 1; 푃푛 = ∑푖=1 푃0 푃푖 and is the

−1 normalized sum of prices for the competing differentiated goods; 푦푐 = 푝0 푦̃푐 is consumer i’s normalized income; ∝0푐 and ∝푖푐 represent consumers’ preferences for the outside good and differentiated goods, respectively; and 훾0 and 훾푖 are consumer-invariant income effects. Finally,

푠푖푗푐 is a symmetric negative semidefinite matrix where the diagonal elements describe own price effects and the off-diagonal elements describe cross-price effects. Using Roy’s identity, we can recover each consumer’s demand for station i:

휕푉 푐⁄ 푛 휕푝 ∝푖푐+ ∑푗=1 푆푖푗푐푃푗 − 훾푖푦푐 푞 = 푖 = 푖,푐 휕푉 ∑푛 푐⁄ 푃0(훾0 + 푖=1 훾푖 푃푛) 휕푦푐

푛 I normalize the price index 푃0(훾0 + ∑푖=1 훾푖 푃푛) = 1. Total demand for each station i is then the sum of each consumer’s individual quantity demanded for that station:

ℎ 푛 푄푖 = ∑ 푞푖,푐 = ∝푖+ 푠푖푖푃푖 + ∑ 푠푖푗푃푗 − 훾푖푦 푐=1 푗≠푖

ℎ ℎ Where ∝푖= ∑푐=1 ∝푖푐 is the sum of consumer c’s preferences for station i; 푠푖푗 = ∑푐=1 푆푖푗푐 and is

ℎ the sum of cross-price effects between station i and all other stations; and y= ∑푐=1 푦푐, which represents aggregate income. If each station solves the problem:

max 휋푖 = (푝푖 − 푚푐푖)푄푖 − 퐹푖 푝푖

Where 푚푐푖 and 퐹푖 are station-specific marginal cost and fixed cost, respectively, then each station’s price reaction function can be found by solving the first-order condition:

푛 ∗ ∝푖 1 푠푖푗 훾푖 1 푝푖 = − − ∑ 푝푗 + 푦 + 푚푐푖 2푠푖푖 2 푗≠푖 푠푖푖 2푠푖푖 2

1 ∝푖 훾푖 To simplify estimation, I assume that − ( − 푦 − 푚푐푖) [an expression that describes 2 푠푖푖 푠푖푖 consumers’ preferences for specific stations (훼푖), consumer income effects (훾푖), and stations’ costs (푚푐푖)] is a function of observed station and market characteristics in the form 퐶 + 푋푖훽 +

푢푖, where C is a constant, 푥푖 is a vector of observed station-specific characteristics, and 푢푖 is a random error. The price reaction function can then be rewritten as:

푛 ∗ 1 푝푖 = 푋푖훽 + ∑ 푑푖푗 푝푗 + 푢푖 2 푗≠푖

푠푖푗 Where 푑푖푗 = − (the diversion ratio between station i and j) for all 푖 ≠ 푗. 푠푖푖

4.2. Empirical Specification

For estimation, I represent each station’s price reaction function as:

푛 ∗ 푝푖 = 푋푖훽 + 휆 ∑ 푤푖푗 푝푗 + 푢푖 푗≠푖

푑푖푗 1 푛 Where 푤푖푗 = 푛 ; 휆 = ∑푗≠푖 푑푖푗; 푥푖 is again a vector of observed station-specific ∑푗≠푖 푑푖푗 2 characteristics; 훽 is a vector of regression coefficients; and 푢푖 is a random error. In the literature, this form is known as a spatial autoregressive model (i.e. Ord (1975) and Pinske et al (2002)), and 휆 is called the spatial parameter. Assuming that consumers do not switch to the outside good when a specific station raises its price, the spatial parameter 휆, which I estimate via OLS and

푛 2SLS in the following sections, should approach ½ as ∑푗≠푖 푑푖푗 (the sum of competing stations’

푛 diversion ratios), approaches one, and all lost sales are accounted for. ∑푗≠푖 푤푖푗 푝푗 is then the weighted average price faced by station i, weighted by relative diversion ratio. In lieu of

푠푖푗 observed diversion ratios, the weight matrix is constructed with the assumption that 푑푖푗 = − is 푠푖푖

73 a function of a station’s distance in miles relative to its competitors. The construction of the

푛 weight matrix 푤푖푗 and weighted market price ∑푗≠푖 푤푖푗 푝푗 is discussed below. Since I define multiple weight matrices that cover different geographic distances from a station, 2 ∗ 휆 can be interpreted as the percentage of the business a station loses when engaging in a price hike that is captured by surrounding stations in that specified critical area. To select the optimal weight matrix, I follow Lee’s (2008) suggestion to use the R-square as a model selection criteria.42 I perform an exhaustive search over the non-linear weight matrix parameter, and then estimate the remaining linear parameters using a linear model. The estimates I find with this procedure are the same as those calculated using a non-linear least squares estimator, however since the objective function for this estimation is a step function in the nonlinear weighting parameter, it is difficult to numerically compute the correct standard errors using standard techniques. As a consequence, I do not calculate standard errors for the non-linear critical distance parameter.

5. Weight Matrix Construction and Calculation of Weighted Prices

One of the main differentiating characteristics between retail fuel stations is their location. Stations can be considered substitutes for each other to the extent that they sell similar fuel (real or perceived), and to the extent that consumers are willing and able to travel farther for cheaper fuel and dedicate time and effort to discovering market prices. We therefore expect that the substitutability of each pair of stations will depend upon the geographic distance between each station and the ease by which consumers can discover and compare station prices. Ignoring this fact and considering only average market price, or the price of a few nearby competitors, can

42 Lee (2008) performs Monte Carlo simulations that show the weight matrix best representing the relationship between spatial units maximizes R-square.

74 cause an omitted variable problem.43 To overcome this issue, I follow Lee (200) and Pinske et al. (2002) by generating a weight matrix that depends upon the distance between two stations, using each station’s latitude and longitude coordinates.44 I then use this weight matrix to generate the individual weighted market price faced by each station.

5.1. Weight Matrix Construction

The weight matrix is an NxN matrix, with each element equal to:

1 휔푖푗 = 1 + 푑푖푗

Where dij is the distance between station i and j if dij is less than or equal to a defined critical distance (described below) and infinity otherwise.

To detect changes in the intensity of spatial competition following the 24-hour law, I calculate multiple weight matrices, each differing by the critical distance used. It is assumed that two stations that are farther apart than the critical distance do not directly compete and cannot be considered substitutes.45

43 See Lee (2007) 44 I use the straight-line distance calculated between the two latitude and longitude coordinates. 45 This does not mean that prices are uncorrelated, but that any such correlation is due to the correlation of both stations’ prices with other 3rd part(ies). 75

Table 5 Summary of Generated Weighted Price Statistics Prelaw Period Variable Observations Minimum Maximum Average Std. Dev Market Mean 24258 40.66726 57.92029 50.05132 3.754114 Price 24258 35.9 66.4 49.99416 4.138034

Weighted Market Price Critical Distance: 0.25 miles 24258 0 62.9 10.92301 20.76064 0.50 miles 24258 0 62.9 30.33388 24.70264 1 mile 24258 37.11535 61.2854 49.97396 3.966129 1.5 miles 24258 37.11535 61.2854 49.98422 3.910405 2 miles 24258 36.84266 61.2854 49.99521 3.872802 3 miles 24258 38.65069 59.89198 50.01875 3.815271 4 miles 24258 39.15541 59.36896 50.04433 3.801886 5 miles 24258 39.12457 58.79983 50.05096 3.791282 6 miles 24258 39.10515 58.80794 50.0565 3.782323 7 miles 24258 39.10515 58.80794 50.05785 3.779026 8 miles 24258 39.09185 58.81233 50.06053 3.775866 9 miles 24258 39.18112 58.75532 50.06039 3.774378

Table 6 Summary of Generated Weighted Price Statistics Postlaw Period Variable Observations Minimum Maximum Average Std. Dev Market Mean 92168 31.7087 58.2912 44.0854 6.00266 price 92168 30.4 99.9 44.0549 6.40102

Weighted Market Price Critical Distance: 0.25 miles 92168 0 68.8954 9.49375 18.3701 0.50 miles 92168 0 99.9 24.138 22.4859 1 mile 92168 30.5 80.9 44.0712 6.26689 1.5 miles 92168 30.7 65.1987 44.08 6.17239 2 miles 92168 30.7 62.7817 44.0687 6.12716 3 miles 92168 30.7 59.9 44.0636 6.08479 4 miles 92168 30.7 59.9 44.0805 6.06223 5 miles 92168 30.8472 59.5652 44.0847 6.0491 6 miles 92168 30.8472 59.5876 44.0881 6.04208 7 miles 92168 30.8472 59.596 44.0838 6.03678 8 miles 92168 30.9332 59.4992 44.0787 6.0322 9 miles 92168 31.01 59.3915 44.0749 6.02762 Note: Only includes stations with at least 1 competitor within 1 mile.

5.2. Weighted Market Price

I construct an individual weighted market price faced by each station using the weighted spatial matrix described above. The purpose of this is to capture the fact that while each station operates within the same market, the competitive pressure each station faces is a function of its spatial position within the market and is therefore unique. The weight-adjusted price faced by each station is defined as:

푁 푤 휔푖푗 푝푖푡 = ∑ [ 푁 ] ∗ 푝푗푡 ∑ 휔푖푗 푗≠푖 푗≠푖

Where N is the number of stations in the data; and pjt is the price of station j at date t. I calculate the daily station-specific weighted market price for twelve distinct critical distances (ranging from one quarter of a mile to nine miles), and once with a critical distance of infinity (all stations compete with all other stations in the market). The resulting series of weighted market prices is similar to the geometric mean market price. I summarize these calculated statistics below in

Table 6.

The weighted prices faced by individual stations are very similar to the market price geometric mean, although the weighted prices tend to have a slightly greater standard deviation and range. This illustrates the individual market conditions each 0station faces from its unique location in the market. Using these individual prices, I estimate each station’s best-response price functions to shed light on the changes in the scope of spatial competition.

6. Empirical Results

Consistent with the hypothesis that the 24-hour rule and accompanying price change restrictions change the scope of spatial competition, I find evidence that stations compete over a

77 smaller geographic area prior to the law and respond to the decrease in search costs by competing with stations up to three miles away after the passage of the law.

6.1. OLS Results

To determine the appropriate spatial scope of competition, I first estimate stations’ price response functions using a standard OLS fixed effect model. While the weighted average price each station faces is clearly endogenous since the price set by competing stations also depends on the station’s own price, fixed effects can allow the model to be estimated consistently under some conditions.46 The weighted price faced by each station is correlated with cost factors, station characteristics, and other station prices. I include station fixed effects to control for time invariant station characteristics, as well as the mix of surrounding station characteristics. Cycle fixed effects control for changes in the cost of LPG between cycles, and any fluctuations in market demand. As a result, the spatial parameter can be estimated consistently.

Since there is a large amount of variation in daily price caused by the price cycles in this market, and since the model I attempt to fit is a static model of equilibrium price response, I calculate the average price charged by each station during the cycle. I then use the variation in average station prices between cycles to calculate stations’ best response functions. The abridged results of this model are reported in Table 7. As expected, the spatial coefficient is near

0.5 in both cases, which theory predicts should be the case. I use the R-square as the model selection criteria47 and find that the appropriate spatial scope is one mile in the prelaw period. In the postlaw period, the appropriate spatial scope is three miles. This is consistent with the

46 See Lee(2002) 47 I am unable to calculate standard errors for the non-linear critical distance parameter as I perform an exhaustive search through a discrete space. 78 hypothesis that the law increased the scope of spatial competition by decreasing consumer and station price discovery costs.

6.2. 2SLS Results

Using the appropriate spatial scope of competition, I am able to estimate consistent coefficients on station characteristics such as brand when describing the effect of weighted market price and station characteristics on a station’s price using two-stage least squares.

Kelejian and Prucha (2002) suggest using weighted market characteristics as instruments for weighted market price. I use demographic statistics from the Australian Bureau of Statistics’

2001 census, collected at the suburb level, as an instrument for weighted market price.

One drawback to using demographic information as an instrument for price is that the observed demographics do not change over the sample period. To overcome this, I again utilize cycle fixed-effects to provide price variation over time.

The results of the two-stage least squares estimation can be found in Table 8. For brevity, I suppress cycle fixed-effects and most demographics, focusing on weighted market price and station brand.

Table 8 Estimated Price Response Function Two-Stage Least Squares Result Prelaw Period Postlaw Period Weighted Price 0.426*** 0.614*** Brand Dummies -3.66 -7.51 Caltex 1.640*** 1.237*** -7.95 -18.18 BP 1.195*** 1.230*** -5.81 -18.55 Gull 0.521 0.233** -1.95 -2.68 Mobil 0.961*** 1.157*** -3.89 -12.22 Shell 1.487*** 1.261*** -6.27 -15.84 Woolworths Plus -0.272 0.128 (-0.55) -0.7 N 622 3151 R-sq 0.953 0.97 F 429.5 1699.9 Note: Cycle FE and Station characteristics suppressed. Full results in appendix. A critical distance of 1 mile is used in the prelaw period, and 3 miles in the postlaw period. * p<0.05; ** p<0.01; *** p<0.001

Similar to the OLS results and consistent with the model, the coefficients on weighted market price in both the prelaw and postlaw periods are statistically significant at the α = 0.001 and close to 0.5, which is contained within the 95% confidence interval. Of particular interest are the coefficients on station brand. The excluded category is independent and small-chain brands

(e.g., Kleanheat), and the results are similar to those found by in the first chapter using brand fixed-effects. One material difference between these two results is that, while the chapter estimates brand fixed-effects on each day of the cycle, I now estimate the average brand effect over all days in the cycle. Nonetheless, the results are qualitatively similar. I find that the major oil brands (Shell, BP, Caltex, and Mobil) charge higher average prices than their independent

80 and smaller-chain competitors in both the prelaw and postlaw period. This is consistent with the previous results in this market.

6.3. Future Extension: Estimating Using Non-Linear Least Squares

In this paper I use ordinary least squares to estimate the non-linear spatial scope parameter. One the advantages of using OLS to calculate the optimal spatial scope of competition is that the model is easy to specify, tractable to calculate on most modern workstations, and does not require the spatial scope parameter to be continuous. The primary disadvantage of this method is that while I can calculate a point estimate, I am unable to calculate standard errors. As a result, I am unable to test statistical significance of my results.

An alternative method to estimate the spatial scope parameter is to use non-linear least squares. In addition to requiring significantly more processing power, this estimation method also necessitates modifications to the system of weights described in section 4.1. The weight matrix structure defined in the previous sections is unsuitable for this type of estimation because the weight matrix is not continuous.

There are a number of potential continuous weights that can be utilized in a non-linear framework. Ideally, any weight used would capture the reality that the effect of one station’s price on another decreases with distance, as well as additional stations within the same distance.

One potential simple, continuous weight would be:

−∝ 푤푖푗 = 푑푖푗

Where alpha is an indicator of the spatial scope of competition, and wij and dij are the weight and distance between stations i and j, respectively. This simple weight system will place a smaller weight on stations that are further away. The weight on each competing station’s price can then be calculated as:

푤푖푗 푊푖푗 = ∑푗≠푖 푤푖푗

Where wij is the weight matrix element describing the spatial relationship between stations i and j, and Wij is station j’s weighted price for station i. Under this specification, larger values of alpha indicate a smaller scope of spatial competition, and smaller values of alpha indicate a larger scope of spatial competition. With the calculation of standard errors, I will then be able to test the statistical significance of the change in spatial scope.

6.4. Effect of the Law on Price Dispersion

As previously explored, the 24-hour rule affects the market both by increasing information on the consumer and station level and by constraining the timing of firms’ price changes. These market changes affect the distribution of prices in ways that are consistent with the previous literature on both Edgeworth price cycles and price dispersion. As discussed in

Wang (2009) and the first chapter, the large oil chains exert considerable control over their retail sites, even when they are independently owned. While the 24-hour rule does not enhance this control, it does enhance the ability of stations and their larger controlling firms to observe market prices. I find that the law does not change overall price dispersion between brand means, although it does slightly increase average intra-brand price dispersion.

6.5. Overall Market Price Dispersion

The 24-hour rule and implementation of the Fuelwatch website constitute a major change in consumer information. To investigate the effect this change has on the market’s price dispersion, I calculate the gini coefficient and coefficient of variance for the set of observed market prices on each day of the data. Over the entire sample period, the mean gini coefficient is

0.0198, with a standard deviation of 0.0113. I regress each day’s gini coefficient and coefficient of variation on three dummies, representing the post-law period, peak cycle days, and the last

82 day of each cycle. I also include a quadratic trend term48 in an alternative specification to account for the fact that we would expect price dispersion to change in the market over time for a number of reasons.49

Table 9 Effect of the 24-hour rule on Market Price Dispersion Measure of Variability Daily Gini Daily CV Law Dummy 0.00400*** 0.00238 0.0128*** 0.00279 (4.30) (1.19) (6.90) (0.73) Peak Day 0.00327 0.00332 0.00557 0.00592 (1.85) (1.89) (1.59) (1.76) End Day -0.00550** -0.00552** -0.00580 -0.00591 (-3.16) (-3.17) (-1.67) (-1.77) Trend 0.000000620 -0.00000174 (0.05) (-0.08) Trend^2 4.87e-09 3.84e-08 (0.43) (1.76) Constant 0.0168*** 0.0167*** 0.0360*** 0.0357*** (20.44) (13.44) (21.98) (14.97) N 826 826 826 826 R-sq 0.038 0.045 0.061 0.134 F 10.81 7.643 17.85 25.41 Note: * p<0.05; **p<0.01; ***p<0.001 In both specifications without any time trends, I find that the law is associated with a very slight increase in price dispersion. The addition of a time trend causes the coefficient on the 24-hour rule dummy to become non-significant, although the trends themselves are also non-significant.

As expected, I do find that price dispersion is at a minimum on the last day of each cycle, when stations are entrenched in the war of attrition phase, and prices are expected to be near marginal cost.

48 The trend is (1+d) and (1+d)2, where d is the number of days after July 1, 2000. 49 Consumer awareness increases, as well as internet penetration and mobile use. 83

7. Analysis of the 24-Hour Rule on Price Cycle Characteristics

Noel (2008) finds that as differentiation increases, the incentive to undercut decreases, and the price cycle equilibrium collapses to coordinated focal point prices when the extent of differentiation becomes too large (i.e. cross-price elasticity of demand becomes too low). It is not surprising, then, that the 24-hour rule did not permanently destabilize the price cycles in this market since the 24-hour rule lowers search costs and likely results in an increase in the cross- price elasticity between stations. This is consistent with the findings in Section 6.1., resulting in an increase in the scope of spatial competition due to consumers’ increased willingness and ability to switch between competing fuel stations.

Table 10 Cycle Statistics Observations Minimum Maximum Average Std. Dev Prelaw Length 6 13 35 24.66 8.26 R. Amp 5 1.46 8.49 4.29 2.9 F. Amp 6 0.75 3.96 2.18 1.27 Postlaw Length 37 10 30 17.86 4.36 R. Amp 37 1.4 10.26 6.11 2.26 F. Amp 36 1.76 18.22 6.29 3.26

Table 10 reports the characteristics of observed cycles in the prelaw and postlaw periods.

While cycle length tends to become more volatile (which is not predicted by Noel’s model), cycles also become more asymmetric after the passage of the 24-hour rule. In addition, the restoration phase is shorter in the postlaw period than in the prelaw period. This result highlights the fact that both consumers’ and stations’ information are improved by the 24-hour rule, simultaneously aiding both consumer search and station coordination.

The increased asymmetric nature of the postlaw cycles is also consistent with an increase in price elasticity.50 Although Wang (2009b) does not calculate prelaw cross-price elasticities, he shows that cross-price elasticities in the postlaw period can be quite high. We would also expect overall elasticity to increase after the passage of the law as consumers gain additional information about the given price level each day as well as future prices and the current location in the price cycle. This increase in information can aid patient and informed consumers in shifting their demand towards the end of each cycle, when average market price is lowest.

7.1. Effect of Law on Observed Undercutting Behavior

The 24-hour rule may change undercutting behavior in a few potential ways. Noel

(2008) finds that firms tend to undercut less aggressively if demand is price-elastic, and consumers’ improved information about future prices likely encourages intertemporal substitution, increasing aggregate price elasticity. The 24-hour price rule can also be interpreted as an elongation of the “current” pricing period; prior to the law, stations were able to set price numerous times a day, and after the law, firms can only change their price once every 24 hours,

In Noel’s (2008) model, this change is observed as a decrease in the discount rate, which he predicts would cause a decrease in undercutting aggressiveness.

Table 11 reports the average undercutting behavior before and after the law. To observe undercutting behavior, I only include observations for days when stations lower their price.51 I find that the law is associated with smaller, less aggressive undercuts. This is consistent with both the hypothesis that the law increased price elasticity in the market and the results from the previous section, which indicate cycles are more asymmetric after the law comes into effect.

Table 11

50 Profitability of the undercutting phase increases with the price elasticity, thus giving stations the incentive to spend relatively more time in the undercutting phase than the restoration phase. 51 Including all observations does not qualitatively affect the results. 85

Analysis of Undercutting Behavior Brand BP Shell Mobil Caltex Prelaw Average Negative Price Change 2.25 2.75 2.10 2.66 Average Lagged Undercut 1.37 0.44 1.98 0.87 Average Percent Lagged Undercut 2.50% 0.63% 4.05% 1.69% Average Actual Undercut 1.13 0.83 1.63 0.77 Average Percent Actual Undercut 2.20% 1.66% 3.37% 1.56% Postlaw Average Negative Price Change 1.19 1.42 1.3 1.17 Average Lagged Undercut 0.6 0.72 0.42 0.61 Average Percent Lagged Undercut 1.35% 1.66% 1.01% 1.40% Average Actual Undercut 0.2 0.25 -0.12 0.17 Average Percent Actual Undercut 0.45% 0.59% -0.22% 0.41%

8. Policy Recommendation

There are three distinct questions worth asking in terms of policy recommendations: first, is it advantageous for regulators to disrupt the price cycle equilibrium? Second, if it is advantageous, then what is the most effective way to destabilize price cycles? Finally, how successful was the 24-hour rule in reducing consumer fuel costs? While it is not possible to definitively answer these questions with the data at hand, I offer a discussion of each below based on previous literature and the results from this paper.

Edgeworth price cycles are a visible type of tacit collusive equilibrium. Due to the frequency of price changes, consumers can easily observe (particularly during the restoration phase) the large fluctuations in market price. As a result, many consumers and government officials assume that the existence of price cycles in a fuel market is undesirable and aim to disrupt them. While there has been no conclusive evidence confirming whether Edgeworth price cycles lead to higher average offer prices (to say nothing of actual transaction price), there has been some evidence indicating fuel markets with Edgeworth price cycles may be more

86 competitive than many markets without them. For example, Doyl et al. (2010) do not find any evidence that U.S. cities have higher (or lower) average prices than non-cycling cities, and Lewis

(2009) finds that margins in cycling cities had lower prices and margins in the weeks after

Hurricane Rita than non-cycling cities. If these results are representative, disrupting price cycles may not be a desirable regulatory outcome.

Assuming disrupting the price cycle equilibrium is desirable however, what method of disruption might be most advantageous? One key weakness in the Edgeworth price cycle equilibrium is that it breaks down quickly when there are more than two competing firms. While

Noel (2008) shows that price cycles can exist when more than two firms are present, a coordination method at the bottom of the cycle during the war of attrition phase becomes critical.

This coordination problem has primarily been solved through price leadership [i.e.Wang (2009),

Byrne and Ware (2011), and Lewis (2012)], suggesting that regulations degrading the firms’ ability to control a large number of retail sites may be successful in disrupting Edgeworth price cycles by preventing price leadership. In Australia, the Petroleum Retail Marketing Sites Act

1980 attempted to achieve this very objective, but was unsuccessful.52

The final important question is whether the 24-hour rule is beneficial to consumers, regardless of the fact that it failed to disrupt the cycling equilibrium. This paper finds that the law (1) does not have much effect on price dispersion, (2) causes the price cycles to be more asymmetric with a greater rising amplitude (implying larger price hikes), and (3) decreases the size of the average daily undercut. Without quantity data, it is not possible to calculate the average weighted cost (or margin) paid by consumers. As a result, it is not clear whether the 24- hour rule caused consumers to pay higher or lower prices. Nonetheless, because the 24-hour rule

52 See Western Australian Select Committee on Pricing of Petroleum Products (2000, p. 17) for details of the LPG wholesale arrangements that allow larger petroleum firms to control the prices of their franchisees. 87 gives consumers access to the entire set of current market prices as well as the set of future prices, it is likely that patient and informed consumers who are able to travel are better off.

Impatient consumers who do not track the price cycle may pay higher peak prices, as these results seem to imply.

While it is not possible to determine the effect on the transaction-weighted average price faced by consumers, it is possible to comment on whether the law was likely to succeed in breaking down the price cycle in the first place. Since the law acts to reduce search costs, the increase in price transparency can be interpreted as a decrease in differentiation between stations.

This increases the similarity between the Perth metropolitan LPG market and Maskin and

Tirole’s (1988) theoretical formulation, and Noel’s (2008) results also find that this will make the market more amenable to price cycles. If the primary goal of the 24-hour rule is to disrupt this equilibrium, it is clear the law is not suitable for this purpose.

9. Concluding Remarks

This paper finds that the 24-hour rule affects the scope of spatial competition as well as the characteristics of the Edgeworth price cycles found in the Perth metropolitan area’s LPG market. Consistent with decreased search costs, the scope of spatial competition, which is a measure of both station responsiveness and consumers’ willingness to substitute one station for another, increases from one mile in the prelaw period to three miles in the postlaw period. I also find evidence that cycles are more asymmetric, stations undercut less aggressively, and intrabrand price dispersion increases after the enactment of the law. These results, along with the fact that price cycles were observed in the market in the postlaw period, suggests that the 24-hour regulation was ill-suited to disrupt the Edgeworth price cycles in this market.

While the 24-hour rule does not eliminate price cycles, it does increase price transparency, which may benefit fuel consumers even if average prices are higher due to the law.

The FuelWatch website itself states: “In fact, according to a recent independent survey, people who subscribed to the email service said they have saved between $2 and $20 per week, saving as much as $1,050 a year.”53 This suggests that for patient and well-informed consumers, the regulation and the cyclical nature of Edgeworth price cycles allow them to save money optimizing the time of purchase. In order to definitively determine whether consumers are better or worse off after the law, quantity data is required. Future research is also required to determine what effect, if any, Edgeworth price cycles have on the average level of prices offered throughout the market.

53 See “How FuelWatch Works” at http://www.fuelwatch.wa.gov.au/fuelwatch/pages/public/contentholder.jspx?key=works.html 89

References

Albaek, Svend, Peter Mollgaard, and Per B. Overgaard. "Government-Assisted Oligopoly Coordination? A Concrete Case." Journal of Industrial Economics 45.4 (1997): 429-43.

Atkinson, Benjamin. "Retail Gasoline Price Cycles: Evidence from Guelph, Ontario Using Bi- Hourly, Station-Specific Retail Price Data." Energy Journal 1st ser. 30 (2009): 85-109.

"Background to FuelWatch." Background to FuelWatch. N.p., n.d. Web. 06 July 2013.

Brown, Jeffrey R., and Austan Goolsbee. "Does the Internet Make Markets More Competitive?" National Bureau of Economic Research No. W7996 (2000): n. pag.

Dana, James, and Eugene Orlov. "Internet Penetration and Capacity Utilization in the US Airline Industry." Northeastern U. College of Business Administration Research Paper 08-001 (2009): n. pag.

Doyle, Joseph, Erich Muehlegger, and Krislert Samphantharak. "Edgeworth Cycles Revisited." Energy Economics 32.3 (2010): 651-60.

Doyle, Joseph, Erich Muehlegger, and Krislert Samphantharak. "Edgeworth Cycles Revisited." Energy Economics 32.3 (2010): 652-60.

Eckert, Andrew, and Douglas S. West. "Retail Gasoline Price Cycles across Spatially Dispersed Gasoline Stations*." The Journal of Law and Economics 47.1 (2004): 245-73.

Kelejian, Harry H., and Ingmar R. Prucha. "2SLS and OLS in a Spatial Autoregressive Model with Equal Spatial Weights." Regional Science and Urban Economics 32 (2002): 691- 707.

Lee, Lung-Fei. "GMM and 2SLS Estimation of Mixed Regressive, Spatial Autoregressive Models." Journal of Econometrics 137.2 (2007): 489-514.

Lee, Lung-Fei. "Consistency And Efficiency Of Least Squares Estimation For Mixed Regressive, Spatial Autoregressive Models." Econometric Theory 18.02 (2002): 252-77.

Lee, Sang-Yeob. "Three Essays on Spatial Econometrics and Empirical Industrial Organization." The Ohio State University. Thesis Paper. (2008): n. pag. Web.

Lewis, Matthew. "Price Leadership and Coordination in Retail Gasoline Markets with Price Cycles." International Journal of Industrial Organization 30.4 (2012): 342-51.

Lewis, Matthew S. "Temporary Wholesale Gasoline Price Spikes Have Long‐Lasting Retail Effects: The Aftermath of Hurricane Rita." The Journal of Law and Economics 52.3 (2009): 581-605.

Maskin, Eric, and Jean Tirole. "A Theory of Dynamic Oligopoly, II: Price Competition, Kinked Demand Curves, and Edgeworth Cycles." Econometrica 56.3 (1988): 571-99.

Noel, Michael D. "Edgeworth Price Cycles and Focal Prices: Computational Dynamic Markov Equilibria." Journal of Economics & Management Strategy 17.2 (2008): 345-77.

Noel, Michael D. "Edgeworth Price Cycles: Evidence From The Toronto Retail Gasoline Market." Journal of Industrial Economics 55.1 (2007): 69-92.

Ord, Keith. "Estimation Methods for Models of Spatial Interaction." Journal of the American Statistical Association 70.349 (1975): 120-26. JSTOR. Web. 07 July 2013.

Orlov, Eugene. "How Does the Internet Influence Price Dispersion? Evidence from the Airline Industry." The Journal of Industrial Economics 59.1 (2011): 21-37.

Pinkse, Joris, Margaret E. Slade, and Craig Brett. "Spatial Price Competition: A Semiparametric Approach." Econometrica 70.3 (2002): 1111-153.

Stahl, Dale O. "Oligopolistic Pricing with Sequential Consumer Search." American Economic Review (1989): 700-12.

Wang, Zhongmin. "Collusive Communication and Pricing Coordination in a Retail Gasoline Market." Review of Industrial Organization 32.1 (2008): 35-52.

Wang, Zhongmin. "(Mixed) Strategy in Oligopoly Pricing: Evidence from Gasoline Price Cycles Before and Under a Timing Regulation." Journal of Political Economy 117.6 (2009): 987-1030.

Wang, Zhongmin. “Station Level Gasoline Demand in an Australian Market with Regular Price Cycles.” Australian Journal of Agricultural and Resource Economics 53.4 (2009): 467- 83.

Wang, Zhongmin. "Station Level Gasoline Demand in an Australian Market with Regular Price Cycles." Australian Journal of Agricultural and Resource Economics 53.4 (2009): 467- 83.