What Is Hindcasting?

Hindcasting is the practice of running a predictive model over a past period, withholding the observations that were recorded during that period, and comparing the model output against them to estimate how the model will perform on future cases. The resulting output is called a hindcast, a reforecast, or a retrospective forecast. Because the true outcome is already known, hindcasting converts a forecasting problem into a measurable statistical one: each retrospective run yields a paired sample of prediction and observation, and a long series of such pairs supports probability statements about model error, bias, and reliability. The method is central to weather and climate prediction, ocean and coastal engineering, hydrology, and any other discipline where a model must be trusted before the event it predicts occurs.

The idea draws on probability theory and on the resampling logic of cross-validation in statistics. A single successful retrospective run proves little, since a flexible model can reproduce one historical episode by chance. Confidence comes from the size and independence of the retrospective sample, which is why operational centers generate hindcasts spanning decades rather than seasons.

Retrospective Model Runs

A hindcast begins by initializing the model with the analysis or reanalysis fields that describe the atmosphere, ocean, or system state at a chosen past date, then integrating forward exactly as an operational forecast would. Initialization must use only information available at that date, otherwise the exercise leaks future knowledge and inflates apparent skill. Operational practice fixes the retrospective period by convention so that different models can be compared on identical cases. The North American Multi-Model Ensemble maintained by NOAA's Climate Prediction Center requires each participating model to supply a complete set of retrospective forecasts covering 1982 to 2010. In the European Centre for Medium-Range Weather Forecasts configuration introduced with model cycle 48r1, re-forecasts were produced twice a week for the preceding twenty years using an eleven-member ensemble, and they are regenerated whenever the forecast system changes.

Verification and Skill Measurement

The paired hindcast and observation series feeds a verification framework built on scoring rules. Anomaly correlation, root mean square error, and the Brier and continuous ranked probability scores each summarize a different aspect of performance, and skill scores express those quantities relative to a reference such as climatology or persistence. This reference matters because a model can appear accurate simply by reproducing the seasonal cycle. Verification of deterministic skill in multi-model ENSO predictions illustrates the usual finding that skill decays with lead time and varies strongly by season, region, and variable, so a single headline score rarely describes a system adequately.

Calibration and Bias Correction

Hindcasts are not used only to grade a model. They also supply the statistics needed to correct it. Systematic drift, ensemble underdispersion, and conditional bias can all be estimated from the retrospective record and removed from real-time output through quantile mapping, model output statistics, or ensemble regression. Sub-seasonal and seasonal forecasts depend on this step so heavily that the hindcast archive is treated as part of the forecast product rather than as documentation. The requirement creates a practical constraint on model development, since every change to the forecast system invalidates the existing calibration and forces the retrospective runs to be repeated at considerable computational cost.

Applications

Hindcasting has applications across a range of technical fields, including:

• Numerical weather prediction, for calibrating and verifying operational forecast systems
• Seasonal and climate projection, including assessment of El Nino and monsoon prediction
• Ocean wave and storm surge modeling for offshore structure and coastal defense design
• Hydrology and flood forecasting, where long retrospective flow records set design return periods
• Wind and solar resource assessment for power system planning
• Epidemiological forecasting, for evaluating outbreak models against past seasons
• Air quality and dispersion modeling, including reconstruction of past pollution episodes