Comparative Evaluation of ARIMA, LSTM, and Temporal Fusion Transformer for Daily Revenue Forecasting

Abstract

Accurate revenue forecasting is an important challenge for retail businesses due to fluctuating consumer demand and changing sales patterns. This study evaluates and compares three time series forecasting approaches, namely Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), and Temporal Fusion Transformer (TFT), for daily revenue prediction in a thrift fashion retail business. The study uses transaction data from Thriftbynz collected from January 2015 to December 2025, which consists of 14,267 transaction records and is transformed into 4,018 daily revenue observations after preprocessing. The forecasting models are evaluated using a rolling-origin expanding window strategy with forecasting horizons of 1, 7, 14, and 30 days. Performance evaluation is conducted using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Scaled Error (MASE), and Weighted Absolute Percentage Error (WAPE). The results show that TFT achieves the lowest aggregate MAE, followed closely by LSTM, while ARIMA provides a competitive statistical baseline. However, the comparison indicates that model performance is influenced by differences in input features and forecasting characteristics. This research provides an empirical evaluation of statistical and deep learning models for retail revenue forecasting and highlights the importance of fair evaluation frameworks in time series forecasting.