Long short-term memory neural networks applied in demand forecast in the retail market

Resumo

Sales forecasting is an indispensable component of the retail industry, underpinning strategic decision-making and operational planning, and is crucial for maintaining financial stability and facilitating business growth. While traditional statistical methods have been widely utilized to address issues in time series analysis, they often fall short when confronted with high-dimensional, complex, or dynamic non-linear relationships between variables.

In this context, Long Short-Term Memory (LSTM) networks offer significant advantages due to their capability to retain information over prolonged periods. These qualities make LSTMs particularly suited for handling scenarios characterized by complexity and dynamic interactions within the data.

This study evaluates the efficacy of three distinct LSTM architecturesVanilla LSTM, ConvLSTM, and CNN-LSTM Multiscalein forecasting future sales. For a comprehensive analysis, an Average model is also employed as a baseline for comparison. The performance of these models is assessed using several metrics, including Mean Average Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Scaled Error (MASE). This comparative analysis aims to elucidate the relative strengths of each LSTM model in the realm of time series forecasting.

The findings suggest that the ConvLSTM architecture generally surpasses the other models across most evaluation metrics. Furthermore, this study concludes that LSTM-based models are adept at navigating the complexities inherent in time series data, identifying intricate patterns over extended durations. This capability is important for effective forecasting in various practical, real-world scenarios, reinforcing the utility of LSTM networks in advanced analytical applications in the retail sector.