Towards Dynamic Trend Filtering through Trend Points Detection with Reinforcement Learning
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: time series analysis, trend filtering, reinforcement learning, time series forecasting
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Trend filtering simplifies complex time series data by prioritizing proximity to the original data while applying smoothness to filter out noise. However, the inherent smoothness of trend filtering filters out the tail distribution of time series data, characterized as extreme values, thereby failing to reflect abrupt changes in the trend. In this paper, we introduce Trend Point Detection, a novel approach to trend filtering that directly identifies essential points that should be reflected in the trend including abrupt changes. We refer to these essential points as Dynamic Trend Points (DTPs) and extract trends from connecting these points. To identify DTPs, we formalize the Trend Point Detection problem as a Markov Decision Process (MDP). We solve the Trend Point Detection problem using Reinforcement Learning (RL) algorithms operating within a discrete action space, referred to as the Dynamic Trend Filtering network (DTF-net). DTF-net incorporates flexible noise filtering, preserving important original sub-sequences while removing noise as needed for other sub-sequences. We demonstrate that DTF-net excels at capturing abrupt changes compared to other trend filtering algorithms, using synthetic data and the Nasdaq intraday dataset. Furthermore, when we utilize DTF-net's trend as an additional feature for Time Series Forecasting (TSF) in non-stationary data, we demonstrate performance improvements, as abrupt changes are captured rather than smoothed out.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
Supplementary Material: zip
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7385
Loading