A diffusion model approach to time series forecasting

Abstract

This thesis explores the utilization of diffusion models for time series forecasting and investigating how model architectures affect the accuracy of forecasts and if these diffusion models are capable of outperforming industry standard methods such as Autoregressive (AR) and Autoregressive Moving Average (ARMA) models. Two models with different architectures were developed for time series forecasting, the Universal Approximation Theorem (UAT) model and the enhanced diffusion model. The UAT model emerged as superior, outperforming the enhanced model in both mean squared error (MSE) and runtime. Comparisons with traditional AR and ARMA models, revealed that the UAT model achieved lower MSEs compared to the AR and ARMA models, showing its robustness in forecasting non-stationary time series. Despite discrepancies in MSE values of state-of-the-art models, the UAT model showed competitive performance, inference times and seemed to perform comparatively to state-of-the-art models through visual inspection. This suggests that while some complexity in model architecture design is beneficial, it may not be necessary for effective time series forecasting.