Revenue Recognition Date Forecasting in Enterprise Supply Chains: A Machine Learning Approach

Baker Hughes is a global energy technology company operating in over 120 countries, provid- ing equipment, services, and digital solutions across the oil and gas value chain. Within the Turbomachinery and Process Solutions (TPS) Services division, accurate forecasting of revenue recognition dates is critical for quarterly financial planning, investor reporting, and operational decision-making. The company currently relies on a deterministic algorithm to estimate when each sales-order line will generate recognised revenue. While this algorithm encodes established business rules and operational logic, it does not account for the full range of external factors that lead project managers to manually override its predictions. These manual interventions, combined with the algorithm’s systematic biases observed across different process stages, regions, and product types, indicate a persistent gap between algorithmic forecasts and actual outcomes. This thesis investigates the application of machine learning to improve revenue recognition date forecasting for spare parts orders. Two correction strategies are formulated and evaluated on seven years of weekly order-line snapshot data. Strategy 1—the PM Intervention Pipeline— decomposes the correction into a sequential classification-then-regression architecture that ex- plicitly models human oversight: a classifier predicts whether a project manager will override the algorithm, and a regressor estimates the override magnitude. Strategy 2—the Direct Al- gorithm Error Correction—trains a single regressor to predict the total deviation between the algorithmic forecast and the actual close date for all order lines. Evaluation on a held-out test set demonstrates that Strategy 2 reduces the mean absolute error from 49.04 days to 39.51 days, a 19.4% improvement over the raw algorithm baseline, and increases the within-90-day forecast accuracy from 81.1% to 88.5%. Strategy 1 achieves a more modest improvement and introduces a failure mode on the subset of orders it targets for intervention, making Strategy 2 the recom- mended approach. The trained models are planned for integration into the myQMI produc- tion platform as a weekly automated scoring pipeline augmenting the existing deterministic system.