Retail Demand Forecasting Using Sequence to Sequence Long Short-Term Memory Networks

Abstract

Demand forecasting is crucial for a retail business as it can greatly affect everything ranging from promotion, pricing, product assortment and inventory. Building a reliable and useful demand forecasting model is still a challenging task. Machine learning techniques used for demand forecasting including Random Forest Regressor and Support Vector Regressor are inadequate when dealing with time series. Recent works show that Long Short-Term Memory (LSTM) networks can learn non-linear relationships and time-series specific information from retail time series data. In this paper, a methodology based on Sequence to Sequence Long Short-Term Memory (Seq2Seq LSTM) network is proposed to tackle short-term retail demand forecasting problem. The Seq2Seq architecture commonly used for language translation is adapted to retail demand forecasting to improve LSTM's ability of learning long-range temporal dependencies from retail time series data. Experiments are evaluated with different input sequence lengths on store item sales dataset with daily resolution data. Bayesian Optimization is conducted to tune models' hyperparameters and examine whether it could enhance the prediction accuracy of the models. In order to gauge the robustness of the proposed forecasting model, it is compared against Standard LSTM and Vanilla RNN.