Model Verification, Improvement, and Long Term Maintenance
HydroForecast™ is typically delivered as a software service which includes ongoing model retraining, verification, and continual improvement of the entire modeling and forecasting process. We have an active in-house research process producing state-of-the-art publications and a continual process of integrating both internal and external research into the HydroForecast™ service.
Forecast verification is integrated with Upstream Tech’s cloud forecasting infrastructure. We automatically calculate a range of accuracy metrics for each forecast horizon and store these results for easy access.
Forecast verification is done both while developing new improvements to our models (where a re-forecast is created to validate the forecasts as the model would have produced them) and in near real time as gauge observations of true inflows are recorded.
All forecast verification is computed on out-of-sample data (data that was not provided to the model during training.)
Improvement & Long Term Maintenance
The most common reason for updating a model is due to technical improvements the Upstream Tech team has developed through ongoing research and development, both internally and through collaborations with academic and government institutions. In addition, at least once a year Upstream Tech will retrain (re-calibrate) models to ensure we have used the latest historical records in calibration. Importantly, because HydroForecast™ uses many additional inputs to describe the state of the basin, many situations (such as a large forest fire) which would require re-calibration of a traditional model will be immediately observed by satellite inputs and our model will react accordingly without the need for retraining.