Sales Forecasting Based on Transformer-LSTM Model

Abstract

As a matter of fact, sales forecasting becomes a hot topic in recent years. The LSTM model is a classic model in the field of sales forecasting. Recently, Transformer has gained a lot of popularity, and many researchers are applying it to time series forecasting. To be specific, this study proposes a novel Transformer-LSTM framework for the one-step sales prediction task. This study retained the classical Transformer structure, using it as the model's encoder, and then combined attention mechanisms with the LSTM model in the decoder section. This hybrid approach allows the model to leverage the advantages of Transformer in handling long sequences while benefiting from LSTM's unique strengths in processing time series data. Ultimately, the model's performance was evaluated using prediction curve comparisons and RMSE as validation metrics, achieving significant results. In conclusion, these results present a novel model for sales forecasting and time series analysis, contributing to improved prediction accuracy in these domains.