SEASONAL SALES PREDICTION FOR INVENTORY OPTIMIZATION IN OUTDOOR RETAIL

This thesis proposes a machine learning framework utilizing the Light Gradient Boosting Machine (LightGBM) algorithm to predict seasonal demand for specific product category and store combinations across 41 brick-and-mortar stores operated by a Czech outdoor apparel retailer. Addressing the challenge of extreme data sparsity identified at the daily brand level, the research establishes a robust weekly forecasting pipeline leveraging Quantile Regression to generate probabilistic demand scenarios (Conservative, Median, Optimistic) over 4, 8, and 12-week horizons. The model integrates almost three years of SAP ERP transactional data with novel feature engineering techniques, including effective inventory weighting and store-normalized Google Analytics 4 (GA4) traffic signals. Validated on a chronological hold-out set comprising the final 15% of the dataset, the model achieved a Weighted Absolute Percentage Error (WAPE) of 22.1% for the 12-week strategic horizon. This performance represents a dramatic improvement over the industry-standard Seasonal Naive baseline (50.5% WAPE), effectively cutting the forecast error by more than half and doubling the precision of supply chain planning. The thesis further translates these technical gains into a comprehensive business strategy. It outlines a "risk-adjusted" inventory policy that protects working capital on seasonal goods while maximizing availability for core assortments. Additionally, a predictive "Smart Recall" mechanism is proposed to optimize reverse logistics, supporting the company’s international e-commerce expansion without cannibalizing physical retail performance. The final solution is architected on Google Cloud Platform, utilizing a custom Power BI dashboard to translate complex probabilistic forecasts into actionable insights for daily inventory management.