An artificial intelligence approach to forecasting when there are structural breaks: a reinforcement learning-based framework for fast switching

Economic forecasting during structural breaks is challenging due to the possible systematic failure of existent models. Robust forecast devices are able to provide unbiased forecasts just after structural change but at the cost of higher variance in normal times. Therefore, there is a trade-off between bias and variance when we intend to forecast a variable under the possibility of structural breaks. In order to choose the best model for each case scenario, we propose a novel algorithm based on the Reinforcement Learning method. Our method is able to gather history performance from all the tested models and choose the one with best performance depending on the "state" of data as soon as the effects of this change are perceived. Hence, our method is able to adapt to the changes of the structural break very fast and change back to a model with less variance as soon as those effects vanish. We provide evidence based on an extensive and rigorous empirical test with Monte Carlo and real data forecasting exercises that this algorithm can improve forecast performance in a scenario of structural change without losing significant performance under normal times.