On the Use of Hindcast Skill for Merging NMME Seasonal Forecasts across the Western United States

Multimodel ensemble forecasts have gained widespread use over the past decade. A yet unresolved issue is whether forecast skill benefits from the use of prior skill from each model in providing a weighted combination. Here, we use the available seasonal ensemble forecasts of six models from the North American Multi-Model Ensemble (NMME) to study various aspects of prior skill–based weighting schemes and explore ways to merge multimodel forecasts. First, we postprocess each NMME model through quantile mapping and a simple spread error adjustment. Then, using an equal weighted combination as the baseline forecast, we test merging the models together through skill-based weights by varying the prior skill metric and varying how the metrics are aggregated across the different subbasins and time of year. Results confirm prior work that the combined forecasts do outperform individual models. When evaluating prior skill, equal weighting generally performed as well as or slightly better than all weighting schemes tried. The skill of the weighting scheme was not found to be strongly dependent on prior metric but did improve when aggregating all forecasted months and subbasins together to provide one overall weight to each model. Also, we found that including an offset to the prior metric that nudged the weights closer to equal weighting improves skill especially at longer leads where individual model skill is low. Results also show that the weighting schemes performed better than regression-based techniques including multiple linear regression and random forest. SIGNIFICANCE STATEMENT: Here, we test how effective the past performance of seasonal climate models can be used for generating weights to merge multiple models together using the North American Multi-Model Ensemble (NMME) for forecasting temperature and precipitation across the western United States. Our results showed there was little benefit in using prior performance compared to weighting all six NMME models equally. The performance of the weighted forecasts was not strongly dependent on the choice of performance metric or over how many months or basins performance was pooled. An important finding is that when only one or two models were used for merging, performance was reduced relative to equal weighting; however, if three or more models were used, performance was nearly the same.