Generating families of synthetic forecasts of different skills from an existing forecast product

An emerging literature evaluates the water management benefits of hydroclimatic forecasts with timescales of a few days to several months ahead. These studies rely on existing forecast products, which they compare to one another and to baseline scenarios, such as perfect forecast or usual climate or streamflow conditions. Results compare the different products and baselines, both in terms of forecast skill and in terms of value for water management. Yet, the means to systematically explore the link between forecast skill and value (e.g., in terms of water supply reliability or hydropower production) are hampered by the lack of techniques to generate synthetic forecasts that 1) are realistic in that they present similar statistical properties to existing products, and 2) foster productive two-ways conversations between the analysts and academics who propose new products and those who use them to inform decision-making, so they can determine where to focus further product development efforts.

This work proposes a methodology for generating forecasts from an existing product and existing hydroclimatic records (rainfall, temperature, streamflow…). It perfects and extends a recent synthetic forecast generation technique that deterministically generates a forecast for a point in the future with the desired bias and accuracy, using a linear combination of the quantity to predict with a predictor. It perfects it by proposing a methodology to generate a family of forecasts with desired skill and bias, for several of the most common skill measures, including mean absolute error and (root) mean square error. Generated synthetic forecasts are therefore based on existing products and retain their statistical properties while presenting improved skill. The skill improvement can apply to the whole forecast or only to targeted conditions, e.g., drought or flood conditions, or forecasts during and for a certain period of the year. This opens the doors to systematic exploration of the benefits of marginal forecast improvements. The technique is also extended to ensemble (or probabilistic) forecasts, to allow for generating synthetic ensembles with targeted improvements to the continuous ranked probability skill score (CRPSS).