Dynamic time series trend filtering through structural breaks detection with reinforcement learning

Abstract

Trend filtering simplifies the complex patterns of time series data. However, smoothness of trend filtering assumes constantness with a fixed criterion for the entire sequence. Because of constantness, noise filtering occurs in sub-sequences that should be fully reflected in the trend, such as abrupt changes. In this thesis, we propose a Dynamic Trend Filtering network (DTF-net) based on simple Reinforcement Learning (RL). Our model starts with the hypothesis that noise has two types: important events as structural breakpoints and simple white noise. The goal is to include an important event in the trend that existing trend filtering algorithms remove as noise. In DTF-net, discrete action defines Dynamic Trend Points (DTP), which we define as Trend Point Detection. The trend extracted from DTPs uses flexible noise filters, thus conserving important original sub-sequences while removing noise for other sub-sequences as much as necessary. By using forecasting MSE as a reward, we can also capture the temporal dependency before and after the action event when extracting the trend. We demonstrate that DTF-net reflects important events more effectively than other trend filtering methods by analyzing synthetic data. Furthermore, the noise-robust property of DTF-net enhances forecasting performance by utilizing the trends obtained from DTF-net as an additional feature.