A Shift Towards Probabilistic Seasonal Forecasts at the Weather Company, an IBM Business

Seasonal forecasts have been demanded by clients in the B2B (Business-to-Business) space for decades. These demands often require a deterministic forecast of monthly departures from normal for a respective field (i.e., Temperature or Precipitation). However, we are beginning to see a subset of companies become more analytically advanced. These companies often care about what the quantitative risk to the forecast is, having alternative outcomes. Some companies demand a higher temporal resolution than the standard monthly aggregations. While the bulk of our clients still prefer a deterministic outlook, we have begun to provide probabilistic offerings at The Weather Company/IBM to satisfy the growing need for quantitative risk assessment of seasonal forecasts. This offering is a global (~40km grid), probabilistically calibrated, six-month forecast of temperature and precipitation at daily resolution. We offer a total of 50-prototypes (time-series scenarios) and 11 percentiles, making this the first calibrated, daily resolution seasonal probabilistic forecast offering that we are aware of. We are leveraging the ECMWF-S5 model as the source of the ensemble distributions due to the model’s spatial and temporal resolution, along with a full hindcast reforecast offering. Clients often require a historical reforecast for verification analysis and training dataset. Therefore, we have expanded our efforts to reconstruct our calibrated Seasonal Probabilistic Forecast backwards to January 1981. This talk will discuss the science behind the construction of the forecasts and the calibration technique used.