Improved Demand Forecasting Using Artificial Neural Networks: Incorporating Economy Indicators Through Feature Construction

Demand forecasting is critical across various sectors, particularly during highly volatile periods such as the post-COVID-19 era and concurrent economic and energy crises. While extensive machine learning models have shown high accuracy in demand forecasting, limited studies have investigated the effects of feature construction guided by domain knowledge. This study focuses on developing a forecasting model based on Artificial Neural Networks (ANNs) while incorporating economic indicators through feature construction with Genetic programming. The algorithm is leveraged to construct higher-level features based on economic indicator categories, including macroeconomic and government policy, energy and resource prices, and industrial metals prices. The best ANN hyperparameters are selected in a trial-and-error manner. Empirical comparisons are conducted using a 6-year dataset from the power tool industry, with the performance of ANNs being compared to popular multiple linear regression models. The results strongly indicate that feature construction leads to higher accuracy. The accuracy of ANNs with and without feature construction is 86% and 71%, respectively. A significant contribution of this study is the incorporation of economic indicators in feature construction, expanding the potential to enhance the predictive capabilities of ANNs. This improvement facilitates more precise and informed decision-making in various practical applications.