Better forecasts alone do not solve inventory problems: conservative policy and cost calibration matter more for fill rates than marginal gains from advanced ML forecasting. In tests on a high-volume U.S. SKU, simple exponential smoothing matched or outperformed an ML point model, while LightGBM quantiles added value through uncertainty estimates but did not by themselves reduce stockouts under realistic (s,S) policies.

Main Finding

The paper develops an end-to-end, explainable framework that links demand forecasting to inventory control and shows that inventory outcomes (costs, fill rates, stockouts) depend far more on inventory policy design, cost calibration, and how demand uncertainty is treated than on marginal improvements in point-forecast accuracy. In the demonstrative US retail SKU case, a simple exponential smoothing model produced the best point forecasts, while LightGBM quantile regression provided useful uncertainty estimates; however, conservative inventory policies driven by simple forecasts achieved high service levels, and policy/cost choices dominated the operational impact.

Key Points

• The research frames forecasting as one element of a decision pipeline: data → feature engineering → forecasting (point + probabilistic) → (s, S) policy → simulation-based evaluation → explainability.
• Forecasting methods compared: naive baselines, seasonal naive, exponential smoothing (classical), LightGBM point forecasts, and LightGBM quantile regression (probabilistic).
• Exponential smoothing gave the lowest point forecast error for the selected high-volume SKU, outperforming naive baselines and LightGBM point model.
• LightGBM quantile regression had higher point-error but provided actionable demand uncertainty ranges (probabilistic forecasts).
• Inventory simulations embedded each forecasting approach in an (s, S) replenishment policy optimized by simulation; evaluation metrics included inventory cost, fill rate, stockout counts, and robustness under stress scenarios.
• Small gains in point-forecast accuracy had limited effect on inventory performance; inventory policy parameters and cost assumptions had larger influence on service levels and stockouts.
• Conservative policy settings (larger safety buffers) maintained high fill rates even when using simple forecasts.
• Explainability (SHAP) showed recent demand lags and seasonal/calendar features dominated ML model predictions; a linear surrogate model captured most of the ML behavior, enhancing interpretability for managers.
• The authors argue that probabilistic forecasts and explicit uncertainty treatment are often more operationally valuable than marginal point-accuracy improvements.
• The paper emphasizes reproducibility and presents the approach as a demonstrative case for a generalizable decision-support pipeline, rather than a broad empirical sweep across many SKUs.

Data & Methods

• Data: publicly available US retail data (sales/demand), pricing, and calendar information; analysis focused on a selected high-volume SKU as a demonstrative case.
• Feature engineering: calendar features, price/promotional signals, recent-demand lags, seasonality indicators, etc., to feed forecasting models.
• Forecasting models:
  • Classical/statistical: naive, seasonal naive, exponential smoothing.
  • Machine learning: LightGBM point forecasting.
  • Probabilistic: LightGBM quantile regression (to estimate demand quantiles and uncertainty bands).
• Evaluation framework:
  • Rolling-origin backtesting for realistic out-of-sample assessment.
  • Forecast metrics (point-error measures such as MAE/RMSE referenced) and assessment of probabilistic predictive intervals.
  • Inventory policy: single-item (s, S) reorder policy; policy parameters optimized via simulation.
  • Performance metrics: inventory cost (holding + ordering + stockout penalties as modeled), fill rate, stockout counts.
  • Robustness/stress tests: simulated demand shocks and other exogenous disturbances to test resilience of policies.
  • Explainability: SHAP values to identify dominant features; construction of a simpler linear surrogate model to demonstrate interpretability and reproduce model behavior.
• Implementation: integration of forecast outputs directly into inventory simulation to evaluate operational consequences rather than only statistical accuracy.

Implications for AI Economics

• Economic value of forecasting AI: Gains in point-forecast accuracy alone may yield limited economic benefits; firms realize greater returns by investing in policy design, cost calibration, and uncertainty-aware decision rules.
• Investment priorities: Organizations should balance spend between forecast-model complexity and resources for policy optimization, uncertainty quantification (probabilistic forecasts), simulation-based decision testing, and explainability infrastructure.
• Adoption and governance: Explainable models and surrogate rules increase managerial trust and the likelihood of deployment—thereby improving realized economic impact versus black-box forecasters that are unused or overridden.
• Risk management and resilience: Probabilistic forecasts and robustness testing reduce downside risk (severe stockouts) that can dominate average-cost metrics—important for firms operating at scale where tail events are costly.
• Diminishing returns to accuracy: From an economic-efficiency viewpoint, small forecast improvements can have diminishing marginal value relative to improved inventory rules; this suggests reorienting analytics KPIs from pure accuracy metrics (MAE, RMSE) to decision-focused KPIs (cost, fill rate, stockout risk).
• Labor and organizational implications: Demand for skills likely shifts toward integrative roles—people who combine forecasting, simulation, inventory optimization, and explainability—rather than pure modelers focused solely on marginal accuracy gains.
• Policy and market-level effects: Widespread adoption of uncertainty-aware, policy-optimized inventory systems could reduce aggregate stockouts and waste (economic efficiency) but may also change competitive dynamics (firms that better align forecasting with control realize disproportionate operational advantages).