Högskolan i Skövde

Accurate sales forecasting is critical for effective planning and decision-making in retail, yet conventional statistical and machine learning models often overlook the value of human judgment. This thesis investigates the integration of human forecasts (HF) with model-based approaches in the context of weekly, multi-SKU sales data characterized by heterogeneous data availability. The study addresses two research questions: (1) to compare human forecasts with statistical and machine learning models trained on routinely available data, and (2) to assess whether simple, transparent human-in-the-loop (HITL) ensembles improve forecast accuracy relative to either humans or models alone.To this end, seven base models Exponential Smoothing (ETS), ARIMA, polynomial regression, Random Forest, Extra Trees, Histogram-based Gradient Boosting, and XGBoost were trained across three feature sets: (A) sales and derived features, (B) sales plus promotional information and prognosis leads, and (C) all features including human forecasts. Ensembles were designed by combining subsets of models with HF, incorporating promotion-gating rules and quantile-based weighting. Model and ensemble performance was evaluated using rolling-origin cross-validation with a strict temporal split into 26-week holdout, validation, and out-of-sample blocks. Metrics included MASE, RMSSE, MAE, RMSE, MAPE, ME, and pinball loss.The findings demonstrate that human forecasts provide valuable information but are not significantly biased across SKUs. HITL ensembles consistently reduced forecast error, particularly for products with sparse or irregular promotional histories, while preserving transparency and interpretability. This work contributes to the literature on judgmental forecasting by empirically showing when and how human judgment can be effectively integrated with data-driven methods, highlighting the practical potential of lightweight HITL designs for retail forecasting.