Ensemble reinforcement learning to forecast time series data

: a performance analysis

Abstract

This study aims to investigate the feasibility of applying Temporal Difference models to forecasting time-series data. Initially, this research offers an overview of the key concepts of Approximate Dynamic Programming, covering the basic principles of Dynamic Programming and the structure and architecture of Reinforcement Learning. Subsequently, a concise outline of related literature is presented, high lighting valuable and interesting contributions regarding related topics. The meth ods used in this analysis encompass Temporal Difference algorithms, in particular TD(0), TD(λ), and GTD2. To assess the suitability of these models in predicting time-series data, their performance is compared to that of benchmark models,including a Hodrick-Prescott filter and Auto-Regressive models, when applied to
economic indicators such as Gross Domestic Product, Private Consumption, Investment, and Exports in Portugal. The model’s performance was assessed through the analysis of three main indicators, the Mean Absolute Error, the Mean Square Error, and the Root Mean Square Error. By comparing the performance of benchmark and proposed models, the study suggests that temporal difference models indulge in higher quality predictions, proving themselves to be reliable tools to fore cast time-series data.

Date of Award	5 Jul 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Pedro Afonso Fernandes (Supervisor)