Deterministic forecasts
A deterministic forecast states a single value for a future quantity and says nothing about how uncertain it is — “demand next week is 500 units,” with no range attached. The term is defined by what it omits rather than by how it was computed.
The standard reference’s judgment on the practice is unusually direct: “point forecasts can be of almost no value without the accompanying prediction intervals,” because without them “there is no way of telling how accurate the forecasts are.”
The information that is missing
Most forecasting methods produce a distribution and a point value is a summary of it — “the point forecast is the mean of this distribution.” So a deterministic report is usually not a different kind of forecast but the same forecast with its uncertainty discarded on the way to the reader.
Consider what three situations look like once the range is dropped.
| Full forecast | Reported deterministically | What the reader should have done differently |
|---|---|---|
| 500, 95% interval [480, 520] | 500 | Check whether plausible demand crosses the decision threshold and whether residual risk is acceptable |
| 500, 95% interval [420, 580] | 500 | Evaluate buffer costs against shortage losses and alternative actions |
| 500, 95% interval [50, 950] | 500 | Compare acting now with the value, price, and delay of further information |
All three arrive on the slide as the same number. The third case is the expensive one: a forecast whose uncertainty may change the preferred action is reported with the same authority as one that is precise, and the decision proceeds as if the middle row were the situation.
What to do when the pipeline only carries one number
Sometimes a single value genuinely is the interface — a field in a downstream table, an input to an automated order. The discipline is then to keep the distribution upstream and collapse it deliberately at the point of use, rather than producing only a point value in the first place.
Which summary to collapse to is a decision, not a default. “A forecast method that minimises the MAE will lead to forecasts of the median, while minimising the RMSE will lead to forecasts of the mean,” and for a skewed quantity those differ materially. With constant per-unit costs for shortage and excess, the expected asymmetric linear loss is minimized by the quantile set by their cost ratio. Other loss functions can require a different summary — see quantile forecasts and quantile level.
Three cheap improvements are available even inside a single-number interface. Publish the interval alongside in a separate field or a note. State the horizon with the value, since accuracy at one week and at twelve are different claims. And keep an accuracy record per horizon so that the number’s reliability is knowable even where it is not attached — see forecast error.
Two distinctions to keep clean. A deterministic forecast is not the same as a target: a number negotiated to match a plan is a commitment, and scoring it as a forecast measures the negotiation instead of the method. And it is not the same as a bad forecast — the value may be the best available estimate. What makes the reporting form a problem is that it removes the reader’s ability to tell. See point forecasts and probabilistic forecasts.
References
- Hyndman & Athanasopoulos, Forecasting: Principles and Practice (3rd ed.), §5.9 “Evaluating distributional forecast accuracy”
- §5.8 “Evaluating point forecast accuracy”
- §5.5 “Distributional forecasts and prediction intervals”.
Discover more from Insightful Data Lab
Subscribe to get the latest posts sent to your email.