Question 1 of 6: Components of Demand, Adaptive Forecasting and Econometric Forecasting

Question 1: Components of Demand, Adaptive Forecasting and Econometric Forecasting (20 marks)

Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.

Part (a) — the components of demand. An operations forecasting system does not try to explain demand; it tries to decompose the observed demand series into a small number of behaviours that can each be projected forward, because each behaviour needs a different mathematical treatment and each has a different planning consequence. Six components are conventionally separated.

The average, or base level, is the mean rate of demand over the horizon of interest. It is what sets the size of the facility and the standing workforce, and it is the only component a naive or simple-average forecast captures at all. The trend is the systematic drift of that base level upward or downward over successive periods; it is what drives capacity expansion decisions, and a model that omits it lags a growing series permanently, because every smoothing average is a weighted mean of the past and the past is always lower than the present when demand is rising. The seasonal element is the repeating within-year (or within-week, or within-day) pattern tied to the calendar — heating loads in winter, agricultural equipment before seeding, retail in December. It is the component that drives inventory build-ahead and temporary staffing, and because it is multiplicative in most industrial series it is normally handled as a set of indices rather than as an additive offset.

The cyclical element is the longer, irregular wave associated with the business cycle, with commodity-price cycles, or with the replacement life of the installed base. It has no fixed period, so it cannot be indexed the way a season can; it is usually picked up through causal or econometric models rather than through a time-series model. Autocorrelation is the persistence of demand from one period into the next: a high month tends to be followed by a high month, so the observations are not statistically independent. This is why exponential smoothing works at all — the recent past really does carry information about the next period — and it is also why the residuals of a forecasting model must be checked for pattern, since correlated errors mean the model has left signal on the table. Finally, random variation is the residue after the other five have been removed. It cannot be forecast by definition; it is instead measured, through the mean absolute deviation or standard error, and its size determines the safety stock and the capacity cushion the plan must carry.

Two practical additions belong in any real operations forecasting system. Demand driven by promotion, price change, a new-product launch or a single large tender is an event rather than a component, and must be carried separately so it does not contaminate the seasonal indices for every subsequent year. And the system must forecast demand, not shipments: when an order is lost to a stock-out the shipment history under-records what the customer actually wanted, and a system fed on shipments will forecast its own past shortages forward.

Part (b) — the general structure of an adaptive forecasting system. An adaptive forecasting system is an ordinary forecasting model placed inside a closed loop that monitors its own error and re-tunes its own parameters, so that no analyst has to intervene when the demand process changes. The structure has five elements, shown in the figure below.

The forecast generator is the model proper — typically single, double or triple (Winters') exponential smoothing, with smoothing constants alpha for the level, beta for the trend and gamma for the seasonal indices. It produces the forecast for the next period from the current state of the model. The comparator waits one period, receives the actual demand, and forms the forecast error. The error accumulator maintains two running statistics from that error: the running sum of forecast errors, which measures bias, and the smoothed mean absolute deviation, which measures scatter. The tracking signal is their ratio, and it is the diagnostic that closes the loop: as long as the model is unbiased the positive and negative errors cancel, the running sum stays near zero, and the tracking signal wanders inside a control band of roughly plus or minus four MAD. When the demand process shifts — a step change in level, the onset of a trend, a new seasonal shape — the errors stop cancelling, the running sum walks away from zero, and the tracking signal leaves the band. The parameter adjuster then acts on that signal, either by stepping the smoothing constants up so the model discounts old data faster and re-converges quickly, or by re-initialising the level and trend outright.

Trigg and Leach's adaptive-response-rate scheme is the compact version of the same idea: instead of testing the tracking signal against a limit and switching, it sets the smoothing constant equal to the absolute value of the tracking signal at every period, so a model that is tracking well runs with a small alpha and heavy smoothing, and a model that has just been surprised runs with an alpha near unity and almost no smoothing. The price of adaptivity is nervousness: a run of ordinary random errors in the same direction can trip the loop and make the system chase noise, so real implementations damp the response, impose a floor and a ceiling on the smoothing constants, and require the tracking signal to stay out of band for two or three consecutive periods before re-tuning.

Part (c) — economic forecasting versus regression analysis. The two are not competing techniques on the same footing; they sit at different levels. Regression analysis is a statistical estimation method. It fits a specified functional relationship between a dependent variable and one or more independent variables by minimising the sum of squared residuals, and it delivers coefficients, standard errors and a coefficient of determination. It is silent about causality: a regression will happily report a strong relationship between two series that merely share a trend, and it says nothing at all about where the values of the independent variables will come from next year.

Economic (econometric) forecasting is a modelling programme that uses regression as one of its tools. It begins from economic theory, which specifies which variables belong in which equations and with what sign; it typically involves a system of simultaneous equations rather than a single fitted line, because in an economy the dependent variable of one relation is an explanatory variable in another; it distinguishes leading, coincident and lagging indicators, so that the drivers used to forecast are ones that are observable before the quantity being forecast moves; and it must solve the forecasting problem for its own independent variables, either by projecting them, by taking them from a published macroeconomic outlook, or by treating them as scenario inputs. In Canadian practice those inputs come from sources such as the Bank of Canada's projections, Statistics Canada's industrial indicators and provincial capital-expenditure surveys.

The operational differences that follow are worth stating plainly. A regression is validated on fit and residual behaviour; an econometric forecast is validated on out-of-sample forecasting accuracy and on whether its coefficients carry the sign theory demands. A regression can be built by anyone with the data; an econometric model requires a defensible causal structure, and a statistically excellent equation with a theoretically wrong sign is rejected. And a regression's forecast horizon is limited by the availability of the driver variables, whereas an econometric system generates those drivers internally, which is precisely why it can be run several years out for capacity planning while a bare regression usually cannot.