Integrated demand forecasting and reinforcement learning for order point optimization in inventory planning

Inventory planning in wholesale is increasingly challenging due to demand uncertainties, external disruptions, and rising capital costs. Traditional methods based on average values of historical data and experiential knowledge of planners often struggle to balance inventory holding costs with required service levels. To address this gap, we propose an integrated two-step framework that first generates probabilistic demand forecasts and then uses Deep Reinforcement Learning (DRL) to translate these forecasts into optimized reorder decisions. This separation offers practical value, as many enterprises already rely on forecasting tools, making the approach easier to integrate into existing planning processes while enabling more data-driven optimization. We evaluate the framework on real-world datasets from two wholesalers in Austria and Germany. Our results indicate that the approach can reduce inventory costs while meeting service levels at moderate availability targets, although improvements are less consistent under stricter requirements. Overall, our findings illustrate both the potential and the limitations of combining probabilistic forecasting with DRL, and they provide guidance and outline future pathways of research on when such a two-step approach is most beneficial in practice.