Attention-Based Ensemble Pooling for Time Series Forecasting
This work addresses the challenge of reducing model bias in time-series forecasting for applications like chaotic systems and pandemic data, but it is incremental as it builds on existing ensemble pooling techniques with limited consistent improvements.
The paper tackled the problem of pooling ensemble forecasts in time-series forecasting by proposing an attention-based method to learn weights for weighted averaging of model forecasts. The method achieved excellent valid times for forecasting the non-stationary Lorenz '63 equation but did not consistently outperform existing ensemble pooling for COVID-19 weekly incident deaths forecasting.
A common technique to reduce model bias in time-series forecasting is to use an ensemble of predictive models and pool their output into an ensemble forecast. In cases where each predictive model has different biases, however, it is not always clear exactly how each model forecast should be weighed during this pooling. We propose a method for pooling that performs a weighted average over candidate model forecasts, where the weights are learned by an attention-based ensemble pooling model. We test this method on two time-series forecasting problems: multi-step forecasting of the dynamics of the non-stationary Lorenz `63 equation, and one-step forecasting of the weekly incident deaths due to COVID-19. We find that while our model achieves excellent valid times when forecasting the non-stationary Lorenz `63 equation, it does not consistently perform better than the existing ensemble pooling when forecasting COVID-19 weekly incident deaths.