In nearly seven years as a semiconductor industry analyst, I have thought about forecast methodologies almost every day. Most of the time, I prefer to explore bottom-up models that balance demand for high-volume applications, like PCs and cell phones, against supply as captured by semiconductor average selling prices. However, every once in a while, my training in time-series statistical modeling gets the better of me.

The earliest statistical test I did on the semiconductor industry was to see if lagged sales growth was a reasonable predictor of future growth (the statisticians in the audience will recognize this as autoregression). In other words, could one use, say, August 2004 sales to predict September 2004 sales? Though I did not expect to find a statistically significant relationship, the data proved me wrong. After correcting for quarterly and annual seasonality, the prior month actually does serve as a good predictor.

Simultaneous regression

This finding got me thinking about what other time series might matter in predicting semiconductor industry sales. To figure that out, it is worthwhile to consider both supply and demand variables. In terms of supply, fabrication facility capacity is likely to be a good candidate, as is the risk-free interest rate, since that rate affects the amount of capital available to manufacturers. Turning to demand, a good place to start is with gross domestic product.

The best way to test the significance of a set of variables is, of course, to perform regression on them simultaneously. So, let us consider the ability to predict this quarter's semiconductor industry sales growth given knowledge of last quarter's growth, last quarter's U.S. GDP growth and last quarter's capacity utilization. Using a data series that spans from CY1976Q1 to CY2004Q2, I find that although last quarter's semiconductor industry sales growth is a good predictor, GDP growth and capacity utilization are not statistically significant predictors of this quarter's semiconductor industry sales growth.

Whenever I do this kind of modeling, I'm always hoping to find the forecasting holy grail. In the back of my mind, however, is the cold reality that probably too many factors are at play and that the importance of each changes constantly. Still, that will not keep me from my quest.

Jeremey Donovan (jeremey.donovan@gartner.com) is chief analyst at Gartner Dataquest.</P>