Recurrent Neural Filters: Learning Independent Bayesian Filtering Steps for Time Series Prediction
This work addresses the challenge of accurate and interpretable time series forecasting for applications in domains like finance or weather, offering a novel architectural approach that is incremental in its improvement over existing methods.
The paper tackled the problem of improving time series prediction by decoupling Bayesian filtering steps, introducing the Recurrent Neural Filter (RNF) to learn distinct representations for state transition and update steps. The result showed improved accuracy in one-step-ahead forecasts with realistic uncertainty estimates and facilitated multistep prediction on three real-world datasets.
Despite the recent popularity of deep generative state space models, few comparisons have been made between network architectures and the inference steps of the Bayesian filtering framework -- with most models simultaneously approximating both state transition and update steps with a single recurrent neural network (RNN). In this paper, we introduce the Recurrent Neural Filter (RNF), a novel recurrent autoencoder architecture that learns distinct representations for each Bayesian filtering step, captured by a series of encoders and decoders. Testing this on three real-world time series datasets, we demonstrate that the decoupled representations learnt not only improve the accuracy of one-step-ahead forecasts while providing realistic uncertainty estimates, but also facilitate multistep prediction through the separation of encoder stages.