Expect the UnexpectedIn a world full of uncertainty, skillful forecasting can minimize disruption due to unanticipated eventsby Seth Grimes Continued from Page 1 Speakers in another conference session, on forecasting in transportation, have had their hands full dealing with the effects of last year's terrorist attacks. Roger Schaufele of the Federal Aviation Administration works with time-based data. To allow for Sept. 11th, he switched the basis of his forecasting from yearly to quarterly and monthly. Higher frequencies better allow you to identify and isolate outliers — anomalous values that can inordinately distort a system's overall characteristics. To correct his long-range forecasts, Schaufele decided first to ignore current-year data and then to try splicing current projections and previous, pre-Sept. 11th forecasts. He looked at previous periods of terror activity to validate his revised approach, which was also recommended by Brian Sloboda of the Bureau of Transportation Statistics (BTS) in his talk. Schaufele was quite explicit about the dependence of his forecasting on assumptions about economic, business, and industry conditions. A BTS colleague, Peg Young, similarly needed to reappraise her time-series models in the wake of Sept. 11th. (A time series is a sequence of periodic observations of a value over time, often with a calendar-based frequency such as daily, weekly, monthly, quarterly, and yearly, and sometimes with an event-based frequency, as with stock trades, which occur at irregular intervals. A series consists of signal and noise, that is, meaningful information and random variation.) Young applies structural time-series techniques, which directly estimate the cyclical, trend, and seasonal components of the series signal; forecasts are created by regressing the trend and adding seasonal and cyclical adjustment factors. (Other techniques decompose series into components after estimating an initial model.) Her approach further differentiates temporary, level-shift, and long-term components in the trend, allowing for each to have different statistical characteristics. One of Young's key findings, worked out with Keith Ord of Georgetown University, is that one-step-ahead forecasts based on filtering work better in predicting individual indicators and detecting anomalous outcomes than moving-average based smoothing, which dampens variation by looking at values before and after each data point. Polishing Your Crystal BallTo sophisticated forecasters, I'll simply suggest reviewing your models for bias and preferences and testing them by introducing uncertainty in the form of significant input perturbations. Do they allow circumstances like a doubling of the price of gasoline, loss of business due to a regulatory investigation, or a weeklong transportation disruption? That is, can they predict, and what do they say about, unprecedented change and "it can't happen here" type of bad news? Do your models break down when you create unusual scenarios? If the Sept. 11th terrorism affected your organization, can you apply lessons you learned — and some of the approaches I've cited in this column — to predict the effects of similar shocks? If you've contented yourself with basic techniques — if your projections are invalidated by events like last September's — start by assessing what your forecasting limitation has or may cost you. Abrupt change may leave you with excess inventory, too little production capacity, or simple inability to do business. Do you even know what could or will result from a disruptive event? If not, evaluate ways to move beyond, for instance, linear regression of key variables. You may need to collect more data — new variables — more often and devise new performance measures, and you may also need more sophisticated software. Now, a period of stability, is the time to do it. Don't wait for a crisis. The goal in all cases is to minimize the effect of disruptions, whether foreseeable or not — in some cases, we can even profit from them — and good planning starts with the robust forecasting. Seth Grimes [grimes@altaplana.com] is a principal of Alta Plana Corp., a Washington, D.C.-based consultancy specializing in analytic computing systems and demographic and economic statistics. RESOURCESGeneral Algebraic Modeling System (also known as GAMS): www.gams.com Related Article at IntelligentEnterprise.com: "Simultaneous Equation Models," Feb. 16, 2001: www.intelligententerprise.com/010216/decision1_1.jhtml
|
 |
Most Popular This Week
IE Weekly Newsletter
Subscribe to the newsletter
| |
| |||||||||||||||||||||||||||||||
