| Abstract [eng] |
Restaurant demand forecasting is an important part of restaurant operations management, as demand in this sector is highly volatile, difficult to predict, and closely related to limited operational capacity. Inaccurate demand estimation may result in lost revenue and excessive operating costs. In recent years, the rapid adoption of digital solutions in the restaurant industry has created opportunities to apply advanced machine learning and deep learning methods for demand forecasting. However, complex forecasting models often exhibit a “black-box” nature, making not only forecast accuracy but also the ability to explain model decisions important in practice. Although restaurant demand forecasting and interpretable machine learning methods have received increasing attention in the scientific literature, studies combining these two research directions within the context of restaurant demand forecasting remain limited. The object of the study is hourly restaurant demand forecasting models and their interpretation using interpretable machine learning methods. The aim of the study is to investigate how interpretable machine learning methods can explain the patterns identified by restaurant demand forecasting models. The study analyzes hourly demand data from two restaurants with different operational characteristics. Seasonal Naïve, XGBoost, LSTM, TFT, and NHITS models were applied for demand forecasting. Forecasting accuracy was evaluated using MAE, RMSE, and WAPE metrics. Model interpretation was performed using SHAP, TSHAP, permutation feature importance, and the internal interpretation mechanisms of the TFT model. The results showed that XGBoost was the most accurate model for both restaurants, reducing MAE from 27.005 to 17.873 for the “VLN_1” restaurant and from 11.324 to 6.011 for the “VLN_KLP” restaurant compared to the Seasonal Naïve benchmark. In contrast, the LSTM model failed to outperform the Seasonal Naïve method in both cases. Interpretable machine learning analysis revealed that the most influential forecasting factors were related to time, including hour of day, weekly cycles, historical demand, and restaurant operating schedules. The analysis also showed that different models relied on different historical time intervals when generating forecasts. The findings demonstrate that interpretable machine learning methods can be useful for improving model transparency and for identifying restaurant demand patterns, supporting managerial decision-making, and improving operational planning. Based on the results, restaurant managers are recommended to primarily rely on historical hourly and weekly demand patterns when planning staffing levels, inventory, and other operational resources, as these factors were found to be the most important drivers of demand in the analyzed restaurants. Furthermore, the integration of external data sources did not consistently improve forecasting performance, suggesting that their practical value should be evaluated individually according to the specific context of each restaurant. |
