Forecast Range Calibration

A demand forecast gives a range of 900 to 1,100 units and labels it an 80% interval. If actual demand falls inside such ranges only half the time across many comparable forecasts, the ranges are too narrow or the model is poorly calibrated, and forecast range calibration is the test of whether stated uncertainty matches later outcomes.

Forecasting: Principles and Practice explains prediction intervals and how uncertainty grows with horizon, and Skforecast documents empirical interval calibration, but business data and decisions determine the right grouping and response. Define the target, since sales, cash receipts and warehouse demand have different patterns and should not be pooled into one calibration score without reason.

Freeze forecasts by keeping the lower bound, upper bound, confidence label, issue date and target period before the actual value arrives. Use comparable horizons: a one-day-ahead range should be tested with other one-day-ahead predictions, while a six-month horizon usually needs wider uncertainty.

Check coverage by counting how often actual outcomes fall inside the stated interval and comparing observed coverage with the advertised level over a suitable sample. Check width too, because a range that covers every result by being enormous may be well covered but useless, so report interval width alongside coverage.

Review misses below the lower bound and above the upper bound, which may reveal systematic bias or skewed uncertainty. Segment by conditions, since promotions, holiday weeks, new products and stable recurring contracts may need separate calibration, and test at the level decisions use, because a portfolio total can look well calibrated while individual products repeatedly miss.

Watch small samples, as ten forecasts cannot establish an 80% coverage rate precisely, so show counts and uncertainty. Inspect data quality as well, because a late actuals feed or changed SKU code can create false misses.

Use appropriate methods and label how intervals were built, since historical residuals, statistical models or expert scenarios can all produce them, and a lower 10th-percentile and upper 90th-percentile forecast can form an 80% central interval under the stated convention. Compare the point forecast as well, because a well-calibrated interval can still surround a biased central estimate.

Review width by horizon, since one-month and one-year bands that are equally narrow suggest uncertainty growth is not represented, and consider asymmetric risk, because demand cannot fall below zero while upside may be large and symmetric bands can give impossible lower bounds. An 80% range intentionally misses some outcomes, so cash or safety decisions may need an additional severe downside scenario, and a calibrated lower cash bound can inform liquidity reserve discussions without authorising borrowing or spending.

Update carefully, because widening ranges after every miss may create a useless blanket interval, so diagnose changing volatility or model bias first, retest on future data, mark structural breaks such as new competitors, pricing changes or supply constraints, and keep ranges aligned with lead time so they arrive early enough to act. For an owner, range calibration is an honesty check on uncertainty, and a plain statement such as "historically, about eight of ten actuals fell inside this kind of band" turns a decorative low-high band into a useful planning signal.