This dissertation investigates how well different forecasting methods can predict retail demand to support inventory and replenishment decisions. It uses the Rossmann Store Sales dataset and focus on Store 1, with daily Sales as the target variable. The goal is to compare three model families under the same, fair setup: a simple baseline (Simple Exponential Smoothing – SES), a classical time-series model (ARIMA(1,1,1)), and a modern machine learning model (XGBoost).

It splits the data in time order: the models are trained on the earlier history and evaluated on a holdout test set of the last 30 observed open days, using MAE and RMSE to measure forecast accuracy. SES and ARIMA mainly rely on past sales patterns, while XGBoost is trained with additional explanatory features such as DayOfWeek, Month/Year, Promo, SchoolHoliday, and StateHoliday, allowing it to learn how demand changes with promotions and calendar effects.

The results show a clear ranking: XGBoost performs best, followed by ARIMA, then SES, meaning the machine learning approach produces the most accurate forecasts in this retail case. Practically, this suggests that when demand is influenced by factors like promotions and holidays, feature-based ML models can reduce forecasting errors and help companies avoid stockouts and overstock. The dissertation also discusses limitations (single-store focus, one test window, limited tuning) and proposes future work such as multi-store validation, rolling evaluation, and testing additional models.