Forecasting from Counterfactual Simulator Rollouts: A Sim2Real Evaluation
Simulator-trained forecasters beat historical-data baselines in two inventory-control deployments.
The paper tests cold-start forecasting for new decision policies by training on counterfactual simulator rollouts, then checking transfer against real past deployments. In Study 1, the simulator-trained model cut point-estimate MAPE by 1.2-3.1 percentage points versus the same architecture trained on historical real data. In Study 2, the reduction was larger, at 12.5-18.7 points. Early real observations after deployment allowed lightweight calibration that reduced error by up to another 2.5 points. Source: ArXiv · AI/CL/LG's note.
The paper tests cold-start forecasting for new decision policies by training on counterfactual simulator rollouts, then checking transfer against real past deployments. In Study 1, the simulator-trained model cut point-estimate MAPE by 1.2-3.1 percentage points versus the same architecture trained on historical real data. In Study 2, the reduction was larger, at 12.5-18.7 points. Early real observations after deployment allowed lightweight calibration that reduced error by up to another 2.5 points. Source: ArXiv · AI/CL/LG's note.
score 4