Improved Techniques in Regression‐Based Streamflow Volume Forecasting

Abstract

Although multiple linear regression has been used for many years to predict seasonal streamflow volumes, typical practice has not realized the maximum accuracy obtainable from regression. Several techniques can help provide superior forecast accuracy using regression models: (1) Using only data known at forecast time; (2) principal components regression; (3) cross validation; and (4) systematic searching for optimal or near‐optimal combinations of variables. Using no future data requires that a separate equation be used each month that forecasts are made rather than using a single equation throughout the forecast season. Consistency of month‐to‐month forecasts can be obtained by judicious selection of variables to maintain a high degree of similarity in the monthly equations. Results for the South Fork Boise River at Anderson Ranch Dam and other basins in the West indicate that these new regression procedures can give substantial improvements in forecast accuracy over existing procedures without sacrificing month‐to‐month forecast consistency.