Short-Term Demand Forecast of E-Commerce Platform Based on ConvLSTM Network

Area of Science:

• Data Science
• Machine Learning
• Retail Analytics

Background:

• Accurate sales forecasting is crucial for retail inventory management and operational efficiency.
• Traditional models may not fully capture complex temporal dependencies in sales data.

Purpose of the Study:

• To construct and compare LightGBM (LGBM) and Long Short-Term Memory (LSTM) sales prediction models.
• To evaluate the performance of a convolutional LSTM (ConvLSTM) network for short-term (T+3 days) product sales forecasting.

Main Methods:

• Utilized real sales data for model construction.
• Implemented and compared LGBM, LSTM, and ConvLSTM models.
• Employed Mean Absolute Percentage Error (MAPE) as the primary evaluation metric.

Main Results:

• The ConvLSTM model exhibited superior prediction accuracy compared to LSTM and LGBM.
• ConvLSTM achieved a lower MAPE by 0.42% versus LSTM and 0.68% versus LGBM at the store level.
• Optimal forecasting methods varied by commodity category, with ConvLSTM excelling in time-series applicable categories.

Conclusions:

• ConvLSTM is a highly effective model for short-term product sales forecasting in retail.
• Model selection should consider the specific characteristics of commodity categories.
• The findings provide valuable insights for optimizing retail sales prediction strategies.