AI Glossary / Statistics and probability / Intermediate

Probabilistic forecasting

Definition

Probabilistic forecasting describes a future outcome using a probability distribution or useful parts of that distribution, such as quantiles and prediction intervals. It communicates which outcomes are more plausible and how uncertainty is distributed. A point forecast gives one number; a probabilistic forecast can support decisions that depend on unusually high, unusually low or costly outcomes. Its probabilities are conditional on the information, model and assumptions used when the forecast was issued.

What it means

A demand forecast of 100 units does not tell a manager whether 102 or 180 units would be surprising. A forecast distribution adds that missing context. It can be discrete for counts, continuous for measurements, or represented by simulated future paths. Different summaries answer different questions: the mean concerns average outcomes, the median divides probability in half, and an upper quantile describes a specified tail. None guarantees what will happen on a particular occasion.

How it works

Specify the target, forecast horizon and information available at the issue time. Fit a model capable of producing predictive uncertainty, then generate probabilities, quantiles or simulations without using future observations. Evaluate forecasts across a sequence of historical forecast origins. Check calibration, including interval coverage, and sharpness, meaning how concentrated useful forecasts are. Use proper scoring rules that reward honest probability estimates. For several future periods, preserve relevant dependence when simulating paths; separate marginal intervals do not describe every possible joint outcome.

Three practical examples

Example 1: A retailer uses the upper demand quantile to compare stocking policies when missed sales cost more than leftover inventory. Example 2: A renewable-energy operator forecasts the distribution of tomorrow's output to assess reserve requirements. Example 3: A service center estimates the probability that hourly arrivals will exceed staffing capacity. Each application combines forecast uncertainty with a decision rule. A single confidence-looking percentage attached to a point estimate is not a substitute for a defined predictive distribution.

Think of it this way

Think of a forecast as a map of possible routes through tomorrow. A point forecast marks one destination, while the distribution shows how much probability lies along other routes. The map can be useful without predicting the exact road eventually taken. Its reliability depends on whether similar maps have represented uncertainty honestly and whether tomorrow still resembles the conditions used to build them.

Worked example

Suppose an illustrative demand model assigns a normal distribution with mean 100 and standard deviation 20. The probability of demand at most 120 is the standard-normal probability below 1, approximately 0.8413. The 90th percentile is 100 + 1.28155 × 20 = 125.63 units. This continuous approximation may need adjustment for actual integer demand. Separately, evaluate an 80% interval [80, 120] when demand is 135. Its interval score is width 40 plus 2/0.20 times the upper miss of 15, giving 190. If demand were 100, the score would be 40. Lower scores favor narrow intervals while penalizing misses; merely widening every interval is not a free improvement.

Illustrative case study

Fictional case study: Harbor Supplies previously ordered according to a single expected-demand number. Its analyst adds quantiles and compares historical decisions using only information available at each ordering date. Managers discover that uncertainty grows during promotional weeks, so identical point forecasts can justify different orders. They evaluate shortages, waste and interval scores together. The exercise is illustrative and claims no measured commercial result. The firm keeps a separate scenario for unprecedented disruptions that its historical forecast model cannot credibly quantify.

Common misconceptions and mistakes

Common mistakes include treating an 80% interval as a promise, evaluating one interval in isolation, and confusing uncertainty about an estimated mean with uncertainty about a future observation. High coverage alone can be achieved with unhelpfully wide intervals. Good average coverage can conceal poor coverage for particular horizons or customer groups. Adding marginal quantiles across products generally does not give the corresponding quantile of total demand. Dependence and the information available at forecast time matter.

Business use

Business use: Probabilistic forecasts support inventory, capacity, budgeting and operational contingency decisions when different errors have different costs. State the forecast horizon, target and relevant assumptions next to the numbers. Compare alternatives on the same historical forecast origins and retain simple baselines. Review tail performance where decisions depend on extreme outcomes. Translate distributions into actions with explicit costs and constraints, and monitor whether changes in the business invalidate the relationships on which the forecasts depend.

Check your understanding

Two forecasts have the same mean, but one places much more probability above capacity. Are they operationally equivalent?

Answer: No. They imply different overload risks and may support different staffing decisions.

Why include an interval miss penalty?

Answer: It prevents a forecast from appearing excellent simply by issuing narrow intervals that frequently exclude the outcome.

Learn next

Build rolling-origin forecasts for a simple series and compare empirical coverage at several horizons. Add an interval score and inspect unusually large misses. Then examine how joint simulations change the uncertainty of a weekly total compared with adding daily quantiles. Keep model uncertainty, ordinary outcome variation and unprecedented events conceptually distinct when communicating the results.

Sources

Sources: Hyndman and Athanasopoulos, Forecasting: Principles and Practice, distributional forecasts and prediction intervals: https://otexts.com/fpp3/prediction-intervals.html ; evaluating distributional forecast accuracy: https://otexts.com/fpp3/distaccuracy.html . These author-maintained chapters support forecast distributions, interval interpretation and scoring. The demand calculation, business examples and fictional case above are original illustrations.

Previous: Probabilistic database Next: Probabilistic Graphical Model