Issue



Improving the forecasts


07/01/2000







Elizabeth Schumann Semi analyst

Whether an industry-wide outlook, regional growth projection, or product sales forecast, it seems that we're always talking about how to get better forecasts. Sometimes they're off by a few percentage points or more; sometimes they fail to accurately call the turning point; and sometimes they're flat out wrong.

It's easy to pick on the professional forecasters since their figures are usually the most visible, but the sales and marketing executives I talk to regularly also struggle with the dual issues of 1) getting an accurate forecast, and 2) getting it early enough to be useful for planning. Why is it so difficult to forecast the semiconductor capital equipment industry, and how can the forecasting be improved? I will attempt to answer those questions, as well as discuss some of the tools available for predicting growth in our market.

We are not alone

The first thing to realize is that we in the semiconductor capital equipment industry are not the only ones striving for more accurate forecasts. Most industries perennially struggle with the same issue. At the 1999 Forecasting Conference sponsored by the Association for Manufacturing Technology (AMT), I listened to speaker after speaker talk about what went wrong with last year's forecast, and yet go on to forecast the coming year.


Elizabeth Schumann Semi analyst
Click here to enlarge image

Through my affiliation with the National Association of Business Economists (NABE), I have met numerous forecasters in a variety of industries, and the cocktail party talk at NABE events invariably turns to the accuracy of forecasts and how to improve it. There have been numerous evaluations of the accuracy of forecasts at the macroeconomic level. Many of them, such as Fintzen and Stekler [1], Mills and Pepper [2], and Joutz and Stekler [3], conclude that forecasters tend to underestimate periods of growth, fail to recognize cyclical peaks until after they have occurred, identify cyclical troughs too early, and underestimate periods of decline. Sound familiar?

Fintzen and Stekler [1], in their discussion of the 1990 US economic downturn, give possible explanations for failure to forecast accurately. The same rationale can be applied to semiconductor capital equipment industry forecasts. It could be that available data are ambiguous or erroneous, leading to faulty forecasts. Another possibility is that the current market is driven by factors that have not been observed in prior cycles, and is thus unpredictable. Finally, there is the notion that the forecasting process itself contributes to forecast errors.

Problems with the data

Most forecasts begin with a model, which is basically a theory about what determines the "event" to be predicted, whether it is total semiconductor equipment annual sales or CMP tool monthly orders. A very simple model uses only the past history of the event to extrapolate the future. Sales will be the same as last quarter, for example, or orders will follow the same upward trend of the past year. With a discontinuous time series (sometimes it is trending up, sometimes down, sometimes flat), however, this technique is not usually very accurate.

Let's consider a slightly more complex model to describe changes in demand for semiconductor capital equipment. It would be reasonable to say that semiconductor manufacturers invest in capital equipment when they need more capacity and they can afford it. The need for new capacity is determined by changes in demand for the final product (semiconductors), capacity utilization levels, and yields, among other factors. New equipment investment is also determined by shifts in production technology (shrinks). This would imply that the data set required to predict semiconductor capital equipment sales would include:

  1. a measure of the industry's ability to afford new equipment (revenue, profit, market capitalization, etc.),
  2. an estimate of the change in demand for semiconductors (device sales, downstream product sales, etc.),
  3. fab capacity utilization levels,
  4. yields, and
  5. a measure of technology generation (linewidth, number of pins, etc.).

By reviewing just a few of these items, we can imagine various potential problems with forecast accuracy.

Click here to enlarge image

First — an estimate of the change in demand for semiconductors. Given the relatively short product cycles in the semiconductor industry, this would imply that a forecast of semiconductor demand should be used in our model. But if we can't get an accurate semiconductor forecast in the first place, it is easy to see that the equipment forecast won't be accurate either. Next — capacity utilization levels and yields. Semiconductor manufacturers generally consider this kind of information highly competitive, and therefore it is very difficult to get an accurate estimate of these data. Semiconductor International Capacity Statistics (SICAS) publishes a quarterly report of fab capacity utilization, but this is generally not released until two months after the end of the quarter being measured.

Given the long lead-time needed for equipment delivery, set-up, and ramp, our model would actually benefit from a forecast of capacity utilization, and this is generally not available. And yield is still an important — and missing — part of the equation. Thus, some of the possible problems with the data include inaccurate estimates, time lags in data availability, and missing data sets.

Problems with non-forecastable events

Mark Twain has said, "History doesn't repeat, but it rhymes." This is the case in the semiconductor equipment industry. The industry is characterized by cycles that generally last about five years, with three or so up years and two or so down years. The cycles are generally caused by periods of over-investment alternating with periods of under-investment.

Yet each cycle is different in terms of the key drivers, the encompassing macroeconomic situation, etc. For example, the rapid rise in Internet use and Internet-related products could be considered a non-forecast-able event. We had no history to tell us what would happen or how it would impact semiconductor demand. Another example could be seen in the rise of the Korean semiconductor industry in the mid- to late-1980s, and the subsequent rise of the Taiwanese foundry industry in the mid-1990s. In both cases, particularly in the early stages of the region's growth, the simple model outlined above could not explain the investments made in semiconductor capital equipment. Instead, something more like national strategic policy seemed to be at play.

Another type of non-forecast-able event is called exogenous shock — something that happens outside the industry, but that has an impact on the industry. A good example is the Asian Financial Crisis of 1998. When even the world's top economic forecasters were unable to call that one, how could we have included it in our semiconductor equipment forecasts? Here at the bleeding edge of technology, in an increasingly complex global market and industry, these non-forecast-able events will continue to frustrate the forecasters.

Problems inherent in the forecasting process

Most forecasts in our industry, whether a top-level outlook by professional forecasters or an individual company's product sales forecast, are prepared with a mixture of data and judgment. When dealing with discontinuous time series, the role of judgment generally increases.


Fig 1. Past sales and order tred data are the usual starting points for industry forecasters.
Click here to enlarge image

To err is human, it is said, and sometimes judgment works against us. Perhaps it is the inability to correctly interpret available information. In 1995, for example, many forecasters predicted an imminent silicon wafer shortage, and silicon suppliers began to invest heavily in new capacity. Did we fail to recognize the impact on silicon demand of the linewidth shrinks outlined in the 1994 National Technology Roadmap for Semiconductors (NTRS)? In 1997 and 1998, as semiconductor manufacturers aggressively reduced die sizes to improve productivity, device unit growth continued at a relatively steady pace, but there was a silicon wafer glut that deepened the recession for silicon manufacturers. Of course, the forecast represented by the 1994 NTRS also proved inaccurate as technology generations began to come faster and faster, but the question of properly assessing the impact of design shrinks on silicon demand is still valid.

Numerous studies have found that forecasters make relatively predictable errors. Granger [4], summarizing some of the earlier literature, pointed out that although forecasters use different techniques and models, they have a tendency to be incorrect in the same direction, and the forecasts tend to underestimate actual changes. Several studies have noted a behavioral bias — we would call it human nature — that might account for forecasting errors.

O'Connor, Remus, and Griggs [5] found that the direction of a time series makes a significant difference to forecast accuracy. Not only were downward sloping series forecasted poorly, but they also showed the greatest tendency toward dampening. Stekler [6] showed that a turning point would not be predicted if forecasters' prior probabilities about the likelihood of a downturn were low. In other words, if they weren't expecting a downturn, they didn't forecast a downturn. Schnader and Stekler [7] demonstrated that asymmetric costs associated with predicting false turns as opposed to not calling a true one (such as damage to reputation) might also produce this result.

How to improve the forecast

Up to now, I've focused on problems with forecasts and some of the reasons. But is there no hope? Since data and judgment are each used in the forecasting process, improvements in both will lead to smaller errors in the forecasts. Semi is always working to improve the breadth, depth, and timing of our market data program, and we applaud the efforts of World Semiconductor Trade Statistics (WSTS) and SICAS to continually improve the coverage and detail of their respective data programs. There are many other organizations collecting data for related industries, including AVEM, IPC, and EIA. Perhaps there are other data sets that the industry's associations should pursue. In order for these programs to result in accurate, useful data, however, the support of the industry is required.


Fig 2. Total semiconductor capital equipment quarterly sales. Based on results, 2000 is already shaping up to surpass most forecasters' expectations.
Click here to enlarge image

The industry could also benefit from more academic researchers tackling the forecasting problem and sharing their findings. Academic economics literature is rife with papers analyzing semiconductor manufacturing and trade models, but I was unable to find any references to attempts at a forecasting model for semiconductor capital equipment.

Fintzen and Stekler [1] give advice to the forecaster with regard to the forecast process:

  1. examine a wide variety of data rather than relying on a small number of indicators;
  2. always look for the possibility of a downturn, instead of believing "it's different this time";
  3. explicitly state the assumptions underlying the forecast and give its probability; and
  4. track, question, and revise.

Tools for the forecaster

The forecaster in the semiconductor capital equipment industry has numerous data sets at his or her disposal. Several information products available from Semi are helpful for predicting future trends, including the Worldwide Semiconductor Equipment Market Statistics (SEMS), a monthly report tracking sales and orders for various detailed categories of semiconductor capital equipment in six market regions. While anyone can purchase this data series through a subscription service, companies that participate in the program receive all reports free.

Semi also publishes a Consensus Forecast in June and in November. The forecast results are an average of the outlook of Semi member companies for the next three years. Our Consensus Forecast is subject to all of the forecast problems discussed above, and is no more (or less) accurate than other available forecasts. The value of the Consensus Forecast is in the process of thoughtfully reflecting on the future market trends. Rather than using the final results as the basis for new business plans, we recommend that the Consensus Forecast be used as a gauge.

Finally, for tracking new fab activity, the International Fabs on Disk, a spreadsheet listing of existing and planned fabs, is also available. And look for a new digest of economic indicators of interest to our industry coming soon from Semi. While none of these tools by themselves can pinpoint the next quarter's order volume, the more data we have to help us understand the past, the better our chances of more accurately forecasting the future.

References

  1. D. Fintzen, H.O. Stekler, "Why Did Forecasters Fail to Predict the 1990 Recession?" International Journal of Forecasting, 15, 309-323.
  2. T.C. Mills, G.T. Pepper, "Assessing the Forecasters: An analysis of the Forecasting Records of the Treasury, the London Business School and the National Institute," International Journal of Forecasting, 15, 247-257, 1999.
  3. F. Joutz, H.O. Stekler, "An Evaluation of the Predictions of the Federal Reserve," International Journal of Forecasting, 16, 17-38, 2000.
  4. C.W.J. Granger, "Can We Improve the Perceived Quality of Economic Forecasts?" Journal of Applied Econometrics, 11, 455-473, 1996.
  5. M. O'Connor, W. Remus, K. Griggs, "Going up - Going down: How Good Are People at Forecasting Trends and Changes In Trends?" Journal of Forecasting, 16, 165-176, 1997.
  6. H.O. Stekler, "An Analysis of Turning Point Forecasts, American Economic Review, 62, 724-729, 1972.
  7. M.H. Schnader, H.O. Stekler, "Sources of Turning Point Forecast Errors," Applied Economics Letters, 5, 519-521, 1997

Suggested reading

  • T. Ehrbeck, R. Waldmann, "Why Are Professional Forecasters Biased? Agency Versus Behavioral Explanations," Quarterly Journal of Economics, 111, 21-40, 1996.
  • P. Goodwin, "Improving the Voluntary Integration of Statistical Forecasts and Judgment, International Journal of Forecasting, 16, 85-99, 2000.
  • M. O'Connor, W. Remus, K. Griggs, "Does Updating Judgmental Forecasts Improve Forecast Accuracy?" International Journal of Forecasting, 16, 101-109, 2000.

    Elizabeth Schumann is a senior market analyst and the manager of the Materials Statistics Programs at Semi, 3081 Zanker Rd., San Jose, CA 95134; ph 408/943-7905, fax 408/943-7915, [email protected].