Forecasting ATE sales at Teradyne, Inc.

May 13, 2004 Kapil Dev Singh, Torben Thurow, and Truman Bradley

1 Agenda

• Teradyne’s business • Problem statement • Process comparison and analysis – Teradyne’s forecasting process – System Dynamics process • Conclusions and insights •Next steps

2 Teradyne

• Manufactures and sells equipment that automatically tests semiconductors – Used in sort operations and – Final testing after packaging • Major customers include , Motorola, , , TSMC

Semiconductor Capex

2

1.5

1

0.5

0 3 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Problem Statement

• Teradyne sees dramatic cyclicality in its orders and struggles to efficiently adjust production to meet demand.

4 Teradyne’s forecasting process

• Size of ATE market correlated with total semiconductor market size • Historically, Constant buy rate – ATE market = 2.5% * semi market – Size of semiconductor market based on external forecasts • Recent data departs from historical trends • Sales team provides input for market share estimates and short term forecasting based on customer input 5 Reference Mode Breakdown

• Growth in market

t

e Hope k

size Mar r o ct u d con i

• Oscillation in m e Fear S market size 1990 now 2020 • Increasing

t e k

amplitude of r a Hope oscillation in market tor M nduc o mic Se size Fear

1990 now 2020

6 Momentum Policies

•Internal – Temporary employment – Expandable and contractable capacity – Long customer lead times (order to receipt) – Surge capacity – Shorten component lead times • External – Talk to customers to forecast better

Even with perfect forecasts, operational improvements will drive performance. However the effects of particular policies may not be obvious

7 Causal Loops – Process Insights

Consumer + Electronics Demand Semiconductor Content of Devices • Quick way to improve + + - Electronic Manufacturers Semiconductor Demand

B + Semiconductor understanding of the + Prices + Desired Semiconductor - Production + Semiconductor industry dynamics Capacity Gap Investments in - Process Technology + B + Semiconductor + Capacity + Additions to • Great way to capture Semiconductor + Semiconductor Production Capacity Expected Future Required + Capacity Utilization Semiconductor Testing Functionality of Semiconductors causal relationships Target Capacity + Utilization + - + - Desired ATE + Actual ATE - Semiconductor Capacity Utilization Capacity utilization B Semiconductor ATE + Revenue ATE Testing Design R&D - Obsolescence + + Capacity + + • Expands viewpoint to big B ATE Capacity + Utilization Gap Total ATE + + Deliveries picture + Other ATE - Deliveries Teradyne's + Other ATE Orders - Installed Base + - R • Valuable by themselves Relative Machine Demand for test + Teradyne's Flexibility equipment + - Teradyne's Orders Deliveries + + B + + Teradyne's Relative Relative Cost per Attractiveness + Test Desired Capacity + • Easy to understand and + - Production + + Relative Expected Forecasted Orders Future Performance Teradyne's + Teradyne work with Capacity Additions to Relative Lead B Capacity Time - + + Teradyne's Lead + Teradyne's Time + Capacity Gap

8 Causal Loops – System Insights

• Growth in market size

– Due to growth in consumer Testing market IDM's perception + volatility of risk electronics demand + IDM's production + - capacity – Due to growth in IDM's desire to Subcontractor's share eliminate risk + of testing market Excess testing capacity + semiconductor content in + + IDM's desire for flexible capacity risk aversion electronics - excess capacity attractiveness of + Subcontractoring outsourcing capacity + + • Oscillation in market size Value capturing + + + IDM's demand for Go Fabless IDM's demands for Subcontractors' + subcontractor to add outsourcing desire to get big – Due to forecasting methods Fabless companies capacity + and delays in capacity + market power + - + Land grabbing adjustment Financial success of subcontractors Money grubbing Market power of Competitiveness of - + + IDM subcontractors + Attractiveness of Consolidation subcontracting + Number of subcontractors • Increasing amplitude of - + oscillation in market size Consolidation Price wars – Due to change in industry structure – increasing use of subcontractors

9 Model Construction

Testing Yield Rate

Required Semiconductor Tests

ATE under ATE Testing ATE Testing ATE Equipment Construction Capacity ATE Test Capacity Capacity Gap Construction Starts ATE Testing Capacity Installed Obsolescence Desired ATE C apacity Utilization ATE Contruction Average ATE Test Time Equipment Life Adjustment Time ATE Cap. under Target ATE Capacity ATE Capacity Const. Gap Under Construction Awareness of ATE under Construction

Perceived ATE Cap. under Const. Gap

Desired ATE Cap. Adjustments (additions)

• Forces deep thought on each link of chain, hence detailed understanding of contributing factors • Leads to revision of dynamic hypotheses • Time consuming, hence costly

10 Conclusions and Insights

11 When to use small policy models?

• Small system dynamics models are better suited to studying internal systems than forecasting external events • A forecasting model is really only useful if better than current forecasting approach – Numerical accuracy is important • A small forecasting model may lead to a better understanding of exogenous industry structure, but most firms have few high leverage policies available to influence industry structure.

12 Forecasting drives oscillations

• Oscillations are largely driven by the individual players in the industry trying to predict demand – Production decisions throughout the industry are based on demand forecasts due to long production lead times • Reducing oscillations in the industry will require more accurate forecasting among all players – Decreasing forecast horizon improves accuracy and reduces oscillation – Decreasing time of historical trend increases responsiveness, but increases magnification – Sharing information between firms may also improve forecasting • Responding quickly to changes in required capacity does not significantly affect oscillation magnitude unless forecast horizons are changed

13 Setting Customer Lead Times

• Longer delivery lead times increase volatility in customer orders – Requires customers to make longer term forecasts about equipment requirements – Results in less accurate ordering - cancelled orders and pushbacks may become more common • Balance increased risks from volatility against reduced inventory risks from forcing customers to commit early prior to inventory investments

14 Oscillations Aren’t All Bad

• Industry cyclicality is good for total sales. • Oscillating demand for ATE results in more ATE sales • Testing capacity is driven by peak demand

15 Forecasting Isn’t Everything

• Even a perfect forecast won't solve Teradyne’s problems • Problems stem from the rapid oscillations of demand relative to speed Teradyne can adjust inventory and capacity • Competitive pressure makes reduction in production hard despite knowing that current growth is not sustainable – Customer lock-in increases the risks associated with limiting capacity – Large percentage of total sales are made during booms

16 Bullwhip effect is severe and worsening

• Disaggregation of industry leads to increased volatility • Increasing numbers of firms in supply chain increases forecasting errors – Less information sharing – More steps in the value chain with forecasts at each step • IDMs and subcontractors respond to different market signals and probably set ATE capacity targets differently – Forecasting may improve by considering IDM and subcontractor purchase decisions differently

17 Next steps

• Integrate insights into forecasting efforts – Investigate regression model including growth and size of the semiconductor market – Consider time delays in systems – use regression models with time lags • Choose how to use System Dynamics in the future – Continue with current forecasting approach without System Dynamics – Continue policy model level efforts internally to improve system level understanding and improve current methodology – Consider value of investing in fully calibrated System Dynamics model for forecasting

18