What a Confidence Interval Actually Means in Forecasting

A point forecast is a promise you almost never keep. The number itself will be wrong — the only real question is how wrong. That is what a confidence interval is for: it tells you the range in which the outcome is likely to land. But intervals are routinely misunderstood, and a misunderstood interval is worse than none, because it manufactures false certainty.

The Definition Nobody Reads Twice

A 95% confidence interval is a procedure, not a property of a single number. If you repeated the entire forecasting exercise many times — new data, new fit, new interval — roughly 95% of the intervals would contain the true outcome. For any single interval, the true value is either inside or it is not; the "95%" is a statement about the method, applied over many repetitions.

The practical consequence: one interval landing on the wrong side is not evidence the method is broken. A method is broken when too many intervals miss — say 30% instead of 5%.

Why Naive Bands Are Too Narrow

The cheapest way to build a band is to take the forecast error's standard deviation from the fit period and draw mean ± 1.96 * sd. This almost always understates uncertainty, for three compounding reasons:

• Estimation error is invisible. The model parameters themselves were estimated from limited data. The band treats them as exact, so it misses the uncertainty in the trend slope — which grows with every step you extrapolate.
• Errors are rarely Gaussian. Forecast errors have fat tails. Spikes, level shifts and one-off events produce outliers that a normal curve assigns near-zero probability to.
• Horizon matters nonlinearly. Uncertainty grows with the forecast horizon, and for trended series it grows faster than a straight line — the band should widen by more than sqrt(h).

A more honest approach uses the residuals instead of a formula: take the one-step-ahead errors from the fit, resample them (a simple bootstrap), add each resampled error to the point forecast, and read the 2.5th and 97.5th percentiles of the resulting distribution. That captures fat tails and irregular shapes without assuming normality.

Calibration: The Check That Keeps You Honest

You can measure whether your bands mean anything. Hold out the last N periods, produce intervals for them, and count how often the true value landed inside the 95% band. If the hit rate is close to 95%, the bands are calibrated. If it is 70%, they are overconfident and should be widened — usually by modelling the error distribution more carefully, not by adding a fudge factor.

This is the discipline that separates a forecast from a narrative. The band's width is not a weakness to apologize for; it is the honest report of what you know and do not know about the future. In KPI Master, forecast cards show the interval alongside the point estimate for exactly this reason — the interval is the forecast, the point is just its centre.