An Empirical Evaluation of Deep Learning and Machine Learning Methods towards Retail Sales Prediction

Abstract

The retail sales forecasting problem is difficult due to the factors of seasonality, non-linearity of demand, the contextual variables which are heterogeneous and the time series dynamics which changes over time. This work compares the various statistical, machine-learning and deep-learning methods to forecast weekly retail sales empirically and from a time-aware perspective. The Sample Superstore consists of 10,194 transactions with 21 original attributes collected between 2019 and 2022. Observations per transaction were weekly aggregated and transformed into a leakage-aware representation of history with the lags of 1 week, 2 weeks, 4 weeks, 8 weeks, 13 weeks, 26 weeks, and 52 weeks. The models were assessed on the 80:20 chronological holdout in addition to the time-series validation, and were ARIMA, Random Forest, Ridge Regression, SGD Regressor, LSTM and Transformer. The test error of Ridge Regression was the lowest with MAE=5,009.71, RMSE=6,580.95, MAPE=37.17%, and R²=0.249 while SGD Regressor and Random Forest were the second and third respectively. The weekly setting gave rise to higher errors for LSTM, ARIMA and Transformer models. The findings indicate that the more complex the architecture, the more difficult the task of forecasting is not necessarily the more accurate the forecasting. This study thus provides a repeatable comparison concentrating on the temporal representation, leakage control, the validation strategy, regularization, and the model capacity and available data.