SMU Data Science Review

A Comparative Time Series Analysis of the ARIMA and Temporal Fusion Transformer (TFT) Models

Catherine Ticzon, Southern Methodist UniversityFollow
Aaron Abromowitz, Southern Methodist UniversityFollow
Bivin Sadler, Southern Methodist UniversityFollow

Abstract

Several new transformer-based time series models have been developed in the past five years and research has provided evidence of these models’ superior performance compared to classic statistical models such as ARIMA. While transformer-based models show impressive performance on baseline datasets, no research has been done on the robustness of these models on datasets with controlled modifications and in a replicable manner. In this paper, the Temporal Fusion Transformer (TFT) model was compared to the classical statistical model ARIMA on simulated data using multiple horizons. Data were simulated using a linear combination of exogenous variables; in total, 50 realizations of 52,704 observations were simulated. The comparison tests, which introduced controlled modifications, included 1) a baseline comparison on the simulated data, 2) simulated data with reduced noise, 3) simulated data with increased noise, 4) a reduction of training data, and 5) a nonlinear combination of the target variable. The TFT and ARIMA models were compared using mean squared error (MSE) and mean absolute error (MAE) on various horizons. Results showed that the ARIMA produced lower average error metrics than the TFT across all horizons and under all modified conditions.

Recommended Citation

Ticzon, Catherine; Abromowitz, Aaron; and Sadler, Bivin (2025) "A Comparative Time Series Analysis of the ARIMA and Temporal Fusion Transformer (TFT) Models," SMU Data Science Review: Vol. 9: No. 2, Article 2.
Available at: https://scholar.smu.edu/datasciencereview/vol9/iss2/2