Bayesian Model Averaging for Streamflow Prediction of Intermittent Rivers

Predicting future river flow is a difficult problem. Firstly, models are (by definition) crudely simplified versions of reality. Secondly, historical streamflow data is limited and noisy. Bayesian model averaging is theoretically a good way to cope with these difficulties, but it has not been widely used on this and similar problems. This paper uses realworld data to illustrate why. Bayesian model averaging can give a better prediction, but only if the amount of data is small - if the data is consistent with a wide range of different models (instead of unambiguously consistent with only a narrow range of near-identical models), then the weighted votes of those diverse models will give a better prediction than the single best model. In contrast, with plenty of data, only a narrow range of near-identical models will fit that data, and they all vote the same way, so there is no improvement in the prediction. But even when the data supports a diverse range of models, the improvement is far from large, but it is the direction of the improvement that can predict more accurately. Working around these caveats lets us better predict floods and similar problems, using limited or noisy data.