Calibration of Numerical Weather Prediction (NWP) forecasts is needed to downscale forecasts from the NWP grid to output grids, and to correct biases in the NWP forecasts. NWP produces gridded forecasts of precipitation, with a single value for each grid cell, but actual conditions may vary considerably within a cell. To produce a well-calibrated forecast, we need to take account of this phenomenon, known as sub-grid variability. Also, the equations used in NWP models are simplifications of the actual dynamics, and this introduces biases into the forecast. Finally, we do not have any guarantees that the members of an ensemble represent a random sample, or evenly-spaced quantiles, of the true distribution. For all these reasons, calibration is needed to produce accurate probabilistic forecasts.
Calibrating rainfall forecasts is difficult because of the extremely skewed distribution: while most forecasts and observations of rainfall are either zero or very close to zero, the most important forecasts are for moderate to high rainfall. This means that many common parametric approaches do not perform well.
Our methodology is somewhat similar to the ECPoint approach recently adopted at ECMWF, while improving on some shortcomings of that method. We demonstrate that RainForests outperforms two existing benchmarks: an EMOS approach using logistic regression, and a non-parametric reliability calibration approach.

