Summary: An iterative, human-in-the-loop framework for SKU-level print demand forecasting, combining machine learning with expert feedback to improve forecast accuracy and decision alignment in dynamic supply environments.

HP Inc. manufactures and sells more than 18,000 print-related products in over 170 countries. Accurate forecasting of the heterogeneous and dynamic demand is vital to support supply planning decisions for manufacturing, inventory management, shipment scheduling, and ultimately customer satisfaction. Forecasting higher or lower than actual demand results in excess or shortage that reduces profitability and impacts on-time delivery to customers. Historically, the supply planning depended on (1) consensus demand forecasting approach, which requires manual collection and integration of information by the forecasting experts, and (2) statistical time-series forecasting models. The consensus forecasting approach also requires frequent corrections if some uncertainties in the demand are not accounted for when releasing the forecasting results. While traditional time-series models can work automatically without frequent correction, their forecasting performance is unsatisfactory due to oversimplified modeling inputs and assumptions. In this project, we document the process of using Machine Learning (ML) techniques across all print products at HP Inc., worldwide. Our aim is to automate the forecasting process with high accuracy and to integrate those results into a human-in-the-loop process that merges the strengths of ML, statistical, and consensus forecasting. Our tree-based (LightGBM) forecasting model reduced systematic errors in comparison to existing approaches such as the consensus and statistical forecasting approaches and was deployed as an integrated part of HP Inc.'s forecasting process. Furthermore, our ML framework establishes strong foundation for further methodological improvements in the ML algorithm. We report extensive empirical evidence guiding our methodology design and demonstrating the business implications of our project. We also share several important principles we have applied to manage team-based collaboration for an enterprise-scale project and to ensure the success of our ML-based demand forecasting.

Supermarket chains must decide whether individual stores will offer pickup, delivery, both services, or neither. While consumer preferences for e-commerce features are well studied, the firm-level and store-level decisions that drive channel offerings remain underexplored. We analyze an augmented dataset of market, demographic, spatial and e-commerce information, assembled from three proprietary sources, covering 14,275 grocery stores in the United States. We use a spatial multinomial generalized additive model (GAM) to predict the offered configuration at the store level, benchmarked against elastic net multinomial regression and random forest models, and find that spatial GAM achieved a macro ROC AUC near that of the random forest (0.873 vs. 0.896), while having the strongest balanced accuracy across configurations (0.568). The banner was the most important predictor across all model families, while other important factors include the store type and demographic variables like historical population growth, education levels, population density, household size, income levels and geospatial locations. Our analysis indicates that the local market context retains a significant secondary role that cannot be fully separated from organizational factors in determining the store’s policies. Finally, 15.2% of the stores in our dataset have different predicted configurations in models with and without the banner. This identifies stores whose predicted configurations change when banner information is omitted and may therefore warrant closer managerial review.