43: Comparative Time Series Analysis of the Temporal Fusion Transformer (TFT) and ARIMA Model

Music City Center 

Several new Transformer-based time series models have been developed in the past 5 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, there has been no research done on the robustness of these models on datasets with controlled modifications. In this paper, the temporal fusion transformer (TFT) model was compared to the classical statistical model ARIMA on simulated data with the following modifications: (1) increases in dependent variable noise, (2) addition of exogenous variables that are uncorrelated to the dependent variable, (3) reduction in training set size. The TFT and ARIMA models were compared using mean squared error (MSE) and mean absolute error (MAE) on various horizons. Results show X, Y, Z.

Time Series

Transformer

Temporal Fusion Transformer (TFT)

ARIMA

Simulated data