Other meanings of Probabilistic forecasting
FORECASTING METHODS
Probabilistic forecasting represents uncertainty about future outcomes with probabilities or predictive distributions rather than a single point estimate. It is used in weather, economics, energy, public health, supply-chain planning, and many other fields where decisions depend not only on what is expected, but also on how uncertain that expectation is.
Probabilistic forecasting describes a range of plausible futures and assigns probabilities to them. A point forecast might predict that tomorrow’s electricity demand will be 10,000 megawatt-hours; a probabilistic forecast instead gives a predictive distribution, such as a median of 10,000 with specified probabilities for lower and higher values. For a binary event, the output may be a probability that a hurricane will make landfall or that a patient will experience a complication. For continuous quantities, forecasters may publish quantiles, prediction intervals, or a full distribution.
The probabilities represent uncertainty conditional on available information, including observations, models, and assumptions. They are not necessarily long-run frequencies for one unique event. A 30% probability of rain means that events with comparable forecasts should produce rain roughly 30% of the time when the forecasts are well calibrated. Probabilistic forecasts therefore support decisions with asymmetric costs, such as preparing for a rare but damaging flood.
Probabilistic forecasts can be produced by statistical models, ensembles, simulation, or combinations of several forecasting systems. In numerical weather prediction, an ensemble runs the model repeatedly with varied initial conditions or model formulations; its spread provides information about forecast uncertainty, although it may require statistical calibration.1 In time-series work, predictive distributions can be estimated from models of trends, seasonality, dependence, and changing variance. Bayesian methods represent parameter and model uncertainty through posterior predictive distributions, while machine-learning systems may estimate quantiles or conditional distributions directly.
Evaluation requires proper scoring rules that reward honest probability assignments. The logarithmic score evaluates the probability assigned to the outcome that occurs, whereas the Brier score is widely used for binary events. Calibration measures reliability; sharpness measures concentration, conditional on calibration. The continuous ranked probability score extends these ideas to distributions and ordered outcomes.2 Scores should be assessed on genuinely out-of-sample forecasts, ideally with rolling or blocked time-series validation.
Probabilistic forecasts become useful when they are connected to explicit decisions and consequences. A retailer can compare the expected cost of excess inventory with the cost of stockouts; an electricity operator can plan reserves against demand or renewable-generation uncertainty; and a public-health agency can evaluate intervention thresholds. The best probability threshold is therefore decision-dependent rather than universal.
Communication must distinguish aleatory uncertainty, arising from inherently variable outcomes, from epistemic uncertainty, arising from limited knowledge or model imperfections. Intervals should be labeled clearly, and users should know whether they describe a forecast distribution, a confidence interval for a parameter, or a range of scenarios. The fan chart, used in economic forecasting, displays widening predictive intervals over future horizons. For public weather communication, calibrated probabilities can outperform categorical labels when people understand how to interpret them.3 Poorly explained probabilities can still mislead, especially when users mistake a 20% event probability for an estimate that 20% of the affected area will experience the event.
Probabilistic forecasting is not simply the addition of error bars to a point forecast. A distribution can be systematically overconfident, miss rare events, or fail to represent dependence between related quantities. Multivariate forecasts must preserve relationships such as the joint behavior of temperature, wind, and power output; separately calibrated margins do not guarantee a coherent joint forecast.
Forecast combination is a major source of improvement because different models capture different aspects of uncertainty. Model averaging can reduce dependence on one specification, while post-processing methods such as ensemble model output statistics correct bias and dispersion in raw ensembles. Forecasts may also be coherent across hierarchical levels: regional demand forecasts should add up to national demand forecasts, and product-level inventory predictions should reconcile with totals.4 Rare-event forecasting presents a further edge case: ordinary scores can be dominated by common non-events, so evaluation often examines discrimination, calibration, reliability diagrams, and performance at decision-relevant thresholds separately.
Probabilities are meaningful only relative to a defined forecasting task, information set, outcome space, and forecast horizon.
Help improve the encyclopedia. Reports go straight to the site manager.