BIZ2121 Production & Operations Management
Demand Forecasting
Sung Joo Bae, Associate Professor
Yonsei University School of Business Unilever
Customer Demand Planning (CDP) System
Statistical information: shipment history, current order information
Demand-planning system with promotional demand increase, and other detailed information (external market research, internal sales projection)
Forecast information is relayed to different distribution channel and other units
Connecting to POS (point-of-sales) data and comparing it to forecast data is a very valuable ways to update the system
Results: reduced inventory, better customer service Forecasting
Forecasts are critical inputs to business plans, annual plans, and budgets Finance, human resources, marketing, operations, and supply chain managers need forecasts to plan: ◦ output levels ◦ purchases of services and materials ◦ workforce and output schedules ◦ inventories ◦ long-term capacities
Forecasting
Forecasts are made on many different variables ◦ Uncertain variables: competitor strategies, regulatory changes, technological changes, processing times, supplier lead times, quality losses ◦ Different methods are used Judgment, opinions of knowledgeable people, average of experience, regression, and time-series techniques ◦ No forecast is perfect Constant updating of plans is important Forecasts are important to managing both processes and supply chains ◦ Demand forecast information can be used for coordinating the supply chain inputs, and design of the internal processes (especially the inventory level)
Demand Patterns
Forecastable factors: things that can be forecasted (e.g. surge in demand for lawn fertilizers in the spring and summer) Uncontrollable factors (weekly demand change due to the weather) Demand Patterns
A time series is the repeated observations of demand for a service or product in their order of occurrence There are five basic time series patterns ◦ Horizontal ◦ Trend ◦ Seasonal ◦ Cyclical ◦ Random Demand Patterns
Figure 13.1 – Patterns of Demand
Quantity
Time
(a) Horizontal: Data cluster about a horizontal line Demand Patterns
Figure 13.1 – Patterns of Demand
Quantity
Time
(b) Trend: Data consistently increase or decrease Demand Patterns Figure 13.1 – Patterns of Demand
Year 1 Quantity
Year 2
| | | | | | | | | | | | J F M A M J J A S O N D Months (c) Seasonal: Data consistently show peaks and valleys Demand Patterns
Figure 13.1 – Patterns of Demand
Quantity
| | | | | 1 2 3 4 5 Years (d) Cyclical: Data reveal gradual increases and decreases over extended periods
Causes: 1. Business Cycle (recession to expansion) – unpredictable 2. Product Life Cycle – somewhat predictable Demand Patterns
Figure 13.1 – Patterns of Demand
Quantity
| | | | | | 1 2 3 4 5 6 Years (d) Random: The unforecastable variation in demand However, these random effects are embedded in any demand time series Key Decisions
What to forecast
What forecasting system to use
What forecasting technique to use
Key Decisions Deciding what to forecast ◦ Level of aggregation Most companies are successful in predicting the annual total demand for all services and products (error is less than 5% normally) Aggregation: making forecast on families of services or goods that have similar demand requirements and common processing, labor, and material requirements Aggregation is followed by forecast for individual item (SKUs – Stock Keeping Units: an individual item with an identifying code for inventory and tracking in the supply chain) ◦ Units of measure Product or service units (such as SKUs) are better than the dollar amount Key Decisions Choosing a forecasting system – Wal-Mart’s CPFR (Collaborative Planning, Forecasting, and Replenishment) system o (p.487): Wal-Mart uses CPFR and the internet to improve demand planning performance o What is Wal-Mart’s new approach for forecast? o What are the system’s benefits to Wal-Mart (rtl), Warner-Lambert (mfr)?
o Increased sales and reductions in inventory costs (stock-out reduced from 15% to 2%) o Stock-out: Demand that cannot be fulfilled by the current level of inventory Key Decisions Choosing the type of forecasting technique Judgment and qualitative methods - Other methods (casual and time-series) require an adequate history file, which might not be available - Useful when contextual knowledge from experience matters - Can be used for modifying the quantitative analysis - Cause-and-effect relationships, environmental cues, and organizational information - Executive opinion, expert opinions (demand and technological forecasting in semiconductor), market research (consumer surveys), salesforce estimates Causal methods - Use historical data on independent variables (e.g. promotional campaigns, economic conditions, competitor’s actions) Time-series analysis - Statistical approach to predict future demand using the historical demand data Delphi Method
Delphi method is a process of gaining consensus from a group of experts while maintaining their anonymity Anonymity is important to avoid “bandwagon effect” or “halo effect” ◦ a cognitive bias whereby the perception of one trait (i.e. a characteristic of a person or object) is influenced by the perception of another trait Initiated as a military project to forecast the impact of technology on war at the beginning of the Cold War
Coordinator Sending out aggregates questions the results
Keep sending out another round of surveys to narrow down the gaps ◦ Iterate the process with pre-determined criteria or until the consensus is reached Useful when no historical data are available Can be used to develop long-range forecasts and technological forecasting Key factor in choosing the proper forecasting approach is the time horizon for the decision requiring forecasts Causal Method: Linear Regression
A dependent variable is related to one or more independent variables by a linear equation The independent variables are assumed to “cause” the results observed in the past Simple linear regression model is a straight line
Y = a + bX
where Y = dependent variable X = independent variable a = Y-intercept of the line b = slope of the line Linear Regression
Deviation, Regression Y or error equation: Y = a + bX Estimate of
Y from regression equation Actual value
of Y Dependent Dependent variable Value of X used to estimate Y
X Independent variable
Figure 13.2 – Linear Regression Line Relative to Actual Data Linear Regression Linear Regression Linear Regression
The sample correlation coefficient, r ◦ Measures the direction and strength of the relationship between the independent variable and the dependent variable. ◦ The value of r can range from –1.00 ≤ r ≤ 1.00
The sample coefficient of determination, r2 Measures the amount of variation in the dependent variable about its mean that is explained by the regression line The values of r2 range from 0.00 ≤ r2 ≤ 1.00
The standard error of the estimate, syx Measures how closely the data on the dependent variable cluster around the regression line Using Linear Regression EXAMPLE 13.1 The supply chain manager seeks a better way to forecast the demand for door hinges and believes that the demand is related to advertising expenditures. The following are sales and advertising data for the past 5 months:
Sales (thousands Advertising Month of units) (thousands of $) 1 264 2.5 2 116 1.3 r = 0.980 2 3 165 1.4 r = 0.960 4 101 1.0 syx = 15.603 5 209 2.0
The company will spend $1,750 next month on advertising for the product. Use linear regression to develop an equation and a forecast for this product. Using Linear Regression SOLUTION We used the previous formula to determine the best values of a, b, the correlation coefficient, and the coefficient of determination, and the standard error of the estimate are already given. a = –8.135 r = b = 109.229X r2 = r = 0.980 syx = r2 = 0.960
syx = 15.603
The regression equation is Y = –8.135 + 109.229X Using Linear Regression The regression line is shown in Figure 13.3. The r of 0.98 suggests an unusually strong positive relationship between sales and advertising expenditures. The coefficient of determination, r2, implies that 96 percent of the variation in sales is explained by advertising expenditures.
Brass Door Hinge
250 – X
200 – X
150 – X X Data 100 – X X
Sales (000 units) (000 Sales Forecasts 50 –
0 – | | 1.0 2.0 Advertising ($000) Figure 13.3 – Linear Regression Line for the Sales and Advertising Data Application: Using Linear Regression
The manager at Tuscani Pizza seeks a better way to forecast the demand for Calzone pizza at a certain region and believes that the demand is related to the promotion expenditures. The following are sales and promotion data for the past 5 months:
Sales (thousands Promotion Cost Month of units) (thousands of $) 1 156 3.2 2 132 2.4 r2 = 0.850 3 144 3.1 4 201 3.9 5 194 3.5
The company will spend $2,900 next month on advertising for the product. Use linear regression to develop an equation and a forecast for this product. Time Series Methods
In a naive forecast the forecast for the next period equals the demand for the current period (Forecast = Dt) Estimating the average: simple moving averages ◦ Used to estimate the average of a demand time series and thereby remove the effects of random fluctuation ◦ Most useful when demand has no pronounced trend or seasonal influences ◦ The stability of the demand series generally determines how many periods to include Time Series Methods
450 –
430 –
410 –
390 –
370 – Patient arrivals Patient 350 –
| | | | | | 0 5 10 15 20 25 30 Week Figure 13.4 – Weekly Patient Arrivals at a Medical Clinic Simple Moving Averages
Specifically, the forecast for period t + 1 can be calculated at the end of period t (after the actual demand for period t is known) as
Sum of last n demands Dt + Dt-1 + Dt-2 + … + Dt-n+1 F = = t+1 n n
where
Dt = actual demand in period t n = total number of periods in the average
Ft+1 = forecast for period t + 1 Simple Moving Averages
For any forecasting method, it is important to measure the accuracy of its forecasts. Forecast error is simply the difference found by subtracting the forecast from actual demand for a given period, or
Et = Dt – Ft
where
Et = forecast error for period t Dt = actual demand in period t Ft = forecast for period t Using the Moving Average Method
EXAMPLE 13.2 a. Compute a three-week moving average forecast for the arrival of medical clinic patients in week 4. The numbers of arrivals for the past three weeks were as follows:
Week Patient Arrivals 1 400 2 380 3 411 b. If the actual number of patient arrivals in week 4 is 415, what is the forecast error for week 4? c. What is the forecast for week 5? Using the Moving Average Method
Week Patient Arrivals SOLUTION 1 400 a. The moving average forecast at the end 2 380 of week 3 is 3 411
411 + 380 + 400 F = = 397.0 4 3 b. The forecast error for week 4 is
E4 = D4 – F4 = 415 – 397 = 18 c. The forecast for week 5 requires the actual arrivals from weeks 2 through 4, the three most recent weeks of data
415 + 411 + 380 F = = 402.0 5 3 Application 13.1a
Estimating with Simple Moving Average using the following customer-arrival data
Month Customer arrival 1 800 2 740 3 810 4 790
Use a three-month moving average to forecast customer arrivals for month 5
D4 + D3 + D2 790 + 810 + 740 F = = = 780 5 3 3
Forecast for month 5 is 780 customer arrivals Application 13.1a
If the actual number of arrivals in month 5 is 805, what is the forecast for month 6?
Month Customer arrival 1 800 2 740 3 810 4 790
D5 + D4 + D3 805 + 790 + 810 F = = = 801.667 6 3 3
Forecast for month 6 is 802 customer arrivals Application 13.1a
Forecast error is simply the difference found by subtracting the forecast from actual demand for a given period, or
Et = Dt – Ft
Given the three-month moving average forecast for month 5, and the number of patients that actually arrived (805), what is the forecast error?
E5 = 805 – 780 = 25
Forecast error for month 5 is 25 Weighted Moving Averages
In the weighted moving average method, each historical demand in the average can have its own weight, provided that the sum of the weights equals 1.0. The average is obtained by multiplying the weight of each period by the actual demand for that period, and then adding the products together:
Ft+1 = W1D1 + W2D2 + … + WnDt-n+1
A three-period weighted moving average model with the most recent period weight of 0.50, the second most recent weight of 0.30, and the third most recent might be weight of 0.20
Ft+1 = 0.50Dt + 0.30Dt–1 + 0.20Dt–2 Application 13.1b
Revisiting the customer arrival data in Application 13.1a. Let W1 = 0.50, W2 = 0.30, and W3 = 0.20. Use the weighted moving average method to forecast arrivals for month 5.
F5 = W1D4 + W2D3 + W3D2
= 0.50(790) + 0.30(810) + 0.20(740) = 786
Forecast for month 5 is 786 customer arrivals
Given the number of patients that actually arrived (805), what is the forecast error?
E5 = 805 – 786 = 19
Forecast error for month 5 is 19 Application 13.1b
If the actual number of arrivals in month 5 is 805, compute the forecast for month 6
F6 = W1D5 + W2D4 + W3D3
= 0.50(805) + 0.30(790) + 0.20(810) = 801.5
Forecast for month 6 is 802 customer arrivals Exponential Smoothing
A sophisticated weighted moving average that calculates the average of a time series by giving recent demands more weight than earlier demands Requires only three items of data ◦ The last period’s forecast ◦ The demand for this period ◦ A smoothing parameter, alpha (α), where 0 ≤ α ≤ 1.0 The equation for the forecast is
Ft+1 = α(Demand this period) + (1 – α)(Forecast calculated last period)
= αDt + (1 – α)Ft or the equivalent
Ft+1 = Ft + α(Dt – Ft) Exponential Smoothing
The emphasis given to the most recent demand levels can be adjusted by changing the smoothing parameter Larger α values emphasize recent levels of demand and result in forecasts more responsive to changes in the underlying average Smaller α values treat past demand more uniformly and result in more stable forecasts Exponential smoothing is simple and requires minimal data When the underlying average is changing, results will lag actual changes Exponential Smoothing and Moving Average
450 – 3-week MA 6-week MA forecast forecast 430 –
410 –
390 –
370 – Exponential Patient Patient arrivals smoothing = 0.10
| | | | | | 0 5 10 15 20 25 30 Week Using Exponential Smoothing
EXAMPLE 13.3 a. Reconsider the patient arrival data in Example 13.2. It is now the end of week 3. Using α = 0.10, calculate the exponential smoothing forecast for week 4. b. What was the forecast error for week 4 if the actual demand turned out to be 415? c. What is the forecast for week 5?
Week Patient Arrivals 1 400 2 380 3 411 4 415 Using Exponential Smoothing SOLUTION a. The exponential smoothing method requires an initial forecast. Suppose that we take the demand data for the first two weeks and average them, obtaining (400 + 380)/2 = 390 as an initial forecast. To obtain the forecast for week 4, using exponential smoothing with and the initial forecast of 390, we calculate the average at the end of week 3 as
F4 = 0.10(411) + 0.90(390) = 392.1
Thus, the forecast for week 4 would be 392 patients. Using Exponential Smoothing b. The forecast error for week 4 is
E4 = 415 – 392 = 23 c. The new forecast for week 5 would be
F5 = 0.10(415) + 0.90(392.1) = 394.4
or 394 patients. Note that we used F4, not the integer-value forecast for week 4, in the computation for F5. In general, we round off (when it is appropriate) only the final result to maintain as much accuracy as possible in the calculations. Application 13.1c
Suppose the value of the customer arrival series average in month 3 was
783 customers (let it be F4). The actual arrival data for month 4 was 790. Use exponential smoothing with α = 0.20 to compute the forecast for month 5.
Ft+1 = Ft + α(Dt – Ft) = 783 + 0.20(790 – 783) = 784.4
Forecast for month 5 is 784 customer arrivals
Given the number of patients that actually arrived (805), what is the forecast error?
E5 = 805 – 784 = 21
Forecast error for month 5 is 21 Application 13.1c
Given the actual number of arrivals in month 5, what is the forecast for month 6?
Ft+1 = Ft + α(Dt – Ft) = 784.4 + 0.20(805 – 784.4) = 788.52
Forecast for month 6 is 789 customer arrivals Seasonal Patterns Seasonal patterns are regularly repeated upward or downward movements in demand measured in periods of less than one year Account for seasonal effects by using one of the techniques already described but to limit the data in the time series to those periods in the same season This approach accounts for seasonal effects but discards considerable information on past demand Multiplicative Seasonal Method
Multiplicative seasonal method, whereby seasonal factors are multiplied by an estimate of the average demand to arrive at a seasonal forecast
1. For each year, calculate the average demand for each season by dividing annual demand by the number of seasons per year 2. For each year, divide the actual demand for each season by the average demand per season, resulting in a seasonal index for each season 3. Calculate the average seasonal index for each season using the results from Step 2 4. Calculate each season’s forecast for next year Using the Multiplicative Seasonal Method EXAMPLE 13.5 The manager of the Stanley Steemer carpet cleaning company needs a quarterly forecast of the number of customers expected next year. The carpet cleaning business is seasonal, with a peak in the third quarter and a trough in the first quarter. Following are the quarterly demand data from the past 4 years: Quarter Year 1 Year 2 Year 3 Year 4 1 45 70 100 100 2 335 370 585 725 3 520 590 830 1160 4 100 170 285 215 Total 1000 1200 1800 2200 The manager wants to forecast customer demand for each quarter of year 5, based on an estimate of total year 5 demand of 2,600 customers Forecasting as a Process
A typical forecasting process Step 1: Adjust history file Step 2: Prepare initial forecasts Step 3: Consensus meetings and collaboration Step 4: Revise forecasts Step 5: Review by operating committee Step 6: Finalize and communicate Forecasting is not a stand-alone activity, but part of a larger process Forecasting as a Process
Adjust Prepare Consensus history initial meetings and file forecasts collaboration 1 2 3
Finalize Review by Revise and Operating forecasts communicate Committee 4 6 5 Forecasting Principles
TABLE 13.2 | SOME PRINCIPLES FOR THE FORECASTING PROCESS
. Better processes yield better forecasts . Demand forecasting is being done in virtually every company, either formally or informally. The challenge is to do it well—better than the competition . Better forecasts result in better customer service and lower costs, as well as better relationships with suppliers and customers . The forecast can and must make sense based on the big picture, economic outlook, market share, and so on . The best way to improve forecast accuracy is to focus on reducing forecast error . Bias is the worst kind of forecast error; strive for zero bias . Whenever possible, forecast at more aggregate levels. Forecast in detail only where necessary . Far more can be gained by people collaborating and communicating well than by using the most advanced forecasting technique or model Linear Regression * SSE= Sum of Squared Error Linear Regression