Using Weather Data to Predict Operational Metrics in Food Delivery Services

The thesis investigates the extent to which weather data can improve the accuracy of operational forecasting in food delivery services, using data from one food delivery company. Three operational metrics are examined: order volume, delivery time, and courier utilisation. The study compares multiple machine learning approaches, including gradient-boosted tree ensembles, recurrent neural networks, and Transformer-based architectures, evaluating their performance both with and without meteorological conditions as exogenous inputs. SHAP-based feature attribution is applied to interpret which weather variables exert the greatest influence on each metric. The findings are intended to provide guidance on model selection for weather-augmented operational forecasting and to inform short-term planning decisions at food delivery platform operating in weather-volatile environments.