Prediction Intervals, Not Point Estimates: Forecasting For Leaders Who Decide Under Uncertainty

Ask for a forecast and you usually get a number. We will sell a thousand units. Occupancy will be seventy percent. The number feels like an answer, and it is the most dangerous kind of answer, because it tells you nothing about how much to trust it. A thousand units could mean a tight, well-understood product where the true figure will land between 900 and 1,100, or a volatile new line where it could be 400 or 1,600. Same forecast, completely different decision, and the point estimate erases the difference that matters.

Leaders who decide under real uncertainty, which in this region is most of them, need the forecast to carry its own uncertainty with it. That is what a prediction interval does, and reading one is a skill worth more than any single model.

The Range Is The Information

A prediction interval attaches a probability to a range. Rather than "1,000 units," it says "an 80 percent chance of landing between 700 and 1,400." The standard way to express it is by percentiles: P50 is the middle, the level with a 50 percent chance of being beaten; P10 is a low outcome you would beat nine times in ten; P90 is a high outcome you would beat only one time in ten. The distance from P10 to P90 is the width of your uncertainty, and it is the number most forecasts hide and most decisions turn on.

The Average Is Rarely The Right Plan

Here is the part most people miss, and it costs real money. Even with a perfect forecast, planning to the middle is usually wrong, because the two ways of being wrong do not cost the same. Run out of a product and you lose the margin on every sale you could not make. Over-order and you carry the cost of clearing the leftovers. When those two costs differ, the midpoint of the forecast is not where you should aim. You should aim higher or lower, toward the error that hurts less.

There is a clean rule for exactly how far. It comes from a classic inventory problem and it needs only your forecast and two numbers: the cost of being short by one unit and the cost of being long by one unit.

Now bring in the interval. If the demand forecast for this item is P50 of 1,000 units and P90 of 1,400, the 75th percentile sits somewhere above 1,000, roughly 1,250 on a typical spread. So the right order is closer to 1,250 than to the 1,000 the point estimate would have told you to buy. Flip the costs, an expensive perishable that spoils versus a low-margin sale, and the ratio drops below a half, telling you to stock below the average instead. The forecast did not change. The costs of error decided the plan.

Wider Later, So Re-Forecast Often

The fan chart earlier makes a second point worth holding on to. Uncertainty grows with the horizon. A forecast for next week is tight because little can intervene; a forecast for next year is wide because a season, a storm, a policy change or a competitor can all land in between. This is why a long-range plan should never be a single line committed to once. It is a fan that narrows as the future arrives, and the discipline is to re-forecast as it does, tightening the plan as the interval tightens rather than defending a number set when the range was at its widest.

A caution on the tools. Adding AI to this can sharpen the central forecast and calibrate the interval better, but it cannot remove uncertainty that is genuinely in the world. A model only knows the past it was trained on, and the storm that has not happened yet is not in that past. The honest model reports a wide interval when the future is wide; the flattering one reports a narrow interval you should not believe. When a forecast comes back suspiciously confident about something inherently uncertain, that narrowness is the warning, not the reassurance.

Frequently Asked Questions

What is the difference between a point estimate and a prediction interval?

A point estimate is a single-number forecast, such as "1,000 units". A prediction interval gives a range with a stated probability, such as "an 80 percent chance of 700 to 1,400". Two products can share a forecast of 1,000, one reliable and one a guess, and only the interval tells them apart.

What do P10, P50 and P90 mean in a forecast?

They are percentiles. P50 is the median. P10 is a low outcome you would beat nine times in ten, P90 a high outcome you would beat one time in ten. The gap between P10 and P90 is the width of the uncertainty, where most of the risk lives.

Why is planning to the average forecast often the wrong choice?

Because being wrong in one direction usually costs more than the other. If a stockout costs J$1,500 and a leftover J$500, the errors are not equal, and you should plan above the average toward the cheaper error. The right target is set by the ratio of the two costs.

What is the newsvendor critical ratio?

It is the cost of underage divided by the sum of underage and overage costs, and it gives the percentile of demand to plan to. With J$1,500 underage and J$500 overage, the ratio is 0.75, so you stock to the 75th percentile rather than the average.

How does uncertainty grow the further out you forecast?

The interval widens with the horizon. Next week is tight, next year is wide. Draw a forecast as a fan that spreads over time, and re-forecast as the horizon shortens and the interval narrows rather than committing to a single line.

Does adding AI make forecasts more certain?

AI can sharpen the central forecast and calibrate the interval, but it cannot remove uncertainty that is genuinely in the world. A storm or policy change is not in the training history. A good model reports a wide interval honestly; a suspiciously narrow one is a warning.

Related reading across the Caribbean AI network