Multi-Location Demand Forecasting in FMCG via Deep Reinforcement Learning

In this study, we propose a unified model for forecasting the daily demand of Fast-Moving Consumer Goods (FMCG) across multiple restaurant locations. Unlike traditional machine learning approaches that require prior segmentation of restaurants and products or separate forecasting models for each combination, our approach enables a single model to predict sales for multiple products and locations simultaneously. To achieve this, we trained and evaluated reinforcement learning (RL) models using key features such as pricing, holidays, weather conditions, and USD exchange rates. The study utilized daily sales data spanning from January 1, 2022, to October 14, 2024, covering three restaurants and two products. We experimented with several RL-based models, including Deep Q-Network (DQN), Convolutional Deep Q-Network (CDQN), Long Short-Term Memory (LSTM)-based RL, and Recurrent Neural Networks (RNN)-based RL, comparing their performance using Mean Absolute Percentage Error (MAPE) and Mean Squared Error (MSE) as evaluation metrics. Experimental results indicate that the DQN model achieved the highest predictive accuracy, outperforming other approaches. The proposed forecasting model can significantly contribute to price optimization, inventory management, and strategic decision-making, offering businesses a more efficient way to anticipate demand without the need for extensive segmentation or multiple independent models.