Explainable automobile sales prediction based on temporal fusion transformer with multi-source heterogeneous data

Accurate automobile sales forecasting is the foundation for enterprises to plan production resources, optimize inventory management, and develop marketing strategies. In this study, the initial features were constructed by merging data from several sources, such as the Baidu search index, online reviews, news headlines, and macroeconomic indicators. Variance screening and mutual information approaches were then combined to select features. This study proposes a temporal fusion transformer (TFT) prediction framework based on an arithmetic optimization algorithm (AOA) to address the limited interpretability of standard forecasting models and the ineffectiveness of parameter optimization in automobile sales prediction models. The AOA was used to optimize the seven key TFT parameters in an adaptable manner. The mean absolute percentage errors for monthly sales forecasts for the Audi A6L and Mercedes-Benz E-Class are 12.97% and 23.59%, respectively, lower than those of the eight popular forecasting models. According to the proposed forecasting model’s explanatory analysis, the most important predictors of sales include consumer evaluations, Baidu search indexes, and macroeconomic indicators. The Audi A6L and Mercedes-Benz E-Class models have different lag-order attention rankings in historical data.