Binning of Linear Models for Time Series Forecasting

Vladimir Pyltsov

Keywords: Time Series, Forecasting, Linear Model

Abstract: Linear models have been gaining attention in the time series forecasting (TSF) field, as several variations of modified linear layers have shown excellent performance across benchmarked datasets. This work extends these efforts by exploring the application of the binning technique to linear models. The conceptual rationale is based on the idea that binning can serve as a simple means of efficient learning through isolating temporal patterns. Instead of relying on temporally adjacent observations, binning adopts another perspective by exposing binned linear layers to periodically grouped data, treating temporal neighbors independently. Both conceptual analysis and empirical experimentation are conducted, with the results demonstrating that this modification can lead to improvements in prediction error. This work positions binning as a simple and effective technique, contributing to the exploration of representations structured by periodicity.

Primary Area: learning on time series and dynamical systems

Submission Number: 24156