Investigator
Robin Roundy (Cornell University)
Industrial Liaisons
Sarah J. Hood, IBM Corporation
Mani Janakiram, Intel Corporation
Anticipated Primary Result
* First, an algorithm that intelligently combines demand estimates from clients, firm orders received, and forecasts computed from historical demand, into a single demand forecast. Secondly, tools for quantifying the risks and benefits of building specific amounts of inventory before firm orders are received. This data feeds into production scheduling applications. 2002 Annual Report
Background
The first aspect of the research aims to improve the accuracy of forecasts when loading semiconductor fabrication facilities, simultaneously reducing the need for human input in forecasting. The researchers forecast 0-6 months into the future, by part-number.
The main sources of information for forecasting are historical demands, forecasts from clients, and firm orders received. The relative importance of this information is situation-specific. For example, firm orders received are more effective in forecasting demand for month 1 than for month 6. Historical demand is very useful for mature products, but less so for new products. Manually determining which data to use in forecasting, by part number by forecast horizon, takes time. Their optimization-based approach will automatically adjust as products progress through their life cycles and the business climate changes.
Secondly, Semiconductor companies prefer to make to order. However they build some inventory without firm orders in hand when clients press the manufacturer to fill orders on a lead time that is shorter than the manufacturing lead time, or when demand temporarily dips below the fabs productive capacity. Benefits include smoothing out the load on fabs with minimal risk, and to shortening lead times for clients. The main risk is making product that cannot be sold immediately, or must be discounted.
The two main sources of uncertainty that the researchers capture are customer demand and the randomness in product yields. To understand the demand-related risks we need an estimate of the variance of the error in demand forecasts, derived from the first aspect of their research.
Description
The researchers discuss these topics with several SRC members, and obtain disguised data. Conversations are in progress. One company has provided data.
The authors will formulate and test several parameter-driven schemes to use the data to generate forecasts. For example they could forecast demand using the formula F = aX + bY + cZ, where a,b,c are non-negative parameters to be estimated, X is the forecast provided by the client, Y is a Simple Moving Average forecast, and Z is the firm orders that have been placed for the product for month 2. The parameters are chosen to optimize a statistical measure of forecast quality. Because the business climate constantly changes there is a limited amount of relevant data. This fact has implications: First, they only consider models with a small number of parameters. Second, to increase statistical power they will link parameter estimation for different time horizons for a given part, and may link parameter estimation for similar products. Third, regressions will be constrained to make the results consistent with the context. The second and third considerations will require the development of new techniques.
The researchers will develop probabilistic models for predicting the marginal risk and benefit associated with producing a product. Benefits are measured in dollars of revenue and the customer service provided. Risk is the cost of holding, discounting or discarding unwanted inventory. These models will be inventory-theoretic. The authors will develop approaches for capturing the risk of eroding profit margins. Tests will quantify the benefits that this research provides.
The researchers will survey SRC members on short-term demand forecasting - forecasts that are generated in sufficient detail to drive production. The goal will be to understand current practices and facilitate benchmarking.