The Role of Artificial Intelligence in Demand Forecasting & Inventory Management

Demand forecasting plays a critical role in aligning production, inventory, and distribution decisions with market demand. However, many organizations continue to experience stockouts, excess inventory, and planning uncertainty due to insufficient forecasting accuracy. Recent advances in artificial intelligence (AI) offer new opportunities to improve predictive performance, particularly in volatile and seasonal retail environments with large product assortments. Despite these advances, the extent to which improvements in forecasting accuracy translate into tangible inventory performance gains remains insufficiently explored. This thesis investigates this gap by comparing three AI-based forecasting models—XGBoost, Long Short-Term Memory (LSTM) networks, and the Temporal Fusion Transformer (TFT)—with conventional statistical forecasting methods, focusing on both forecast accuracy and inventory performance. The study adopts a positivist and deductive research design based on quantitative modeling and empirical analysis. Forecasting models are trained, validated, and evaluated using publicly available data from the M5 Forecasting Competition under controlled experimental conditions. Forecast accuracy is assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Scaled Error (MASE). To evaluate operational impact, model forecasts are embedded within a discrete-event inventory simulation based on a continuous-review (𝑠,𝑆) replenishment policy. Inventory performance is measured using stockout rates, fill rates, and average inventory levels. The results demonstrate that AI-based models significantly outperform traditional statistical baselines, with the TFT model achieving an average RMSE reduction of 21.8% relative to the best performing statistical model. These accuracy improvements translate into measurable operational benefits, as the TFT-based forecasts reduce the average stockout rate to 6.4%, compared with 10.9% for the strongest statistical baseline. The findings contribute to the existing literature by providing empirical evidence that improvements in forecasting accuracy can generate meaningful downstream inventory performance gains. The study offers methodological insights into model selection for complex retail environments, feature engineering, and inventory simulation design. Overall, the research highlights the importance of integrating advanced forecasting models with operational evaluation frameworks to support data driven inventory planning and decision-making.