From forecast accuracy to inventory performance: Assessing machine-learning models for intermittent export demand

Abstract

Machine-learning forecasts are usually judged by accuracy, yet procurement decisions consume forecasts through an inventory policy whose lead times, review cycles, and lot sizes can overturn accuracy rankings. This paper quantified that value in replenishment decisions for an Indonesian cassia exporter's intermittent export demand. Ten approaches — the trading partner's company-described ex-ante annual plan, intermittent-demand estimators, ARIMA, and three machine-learning configurations — were embedded in a sequential simulation of the company's periodic-review policy over 55 months of grade-level demand, with a three-month lead time, lot-size constraints, and safety stocks calibrated to each method's realised errors. Each method was charged its cheapest configuration meeting a common achieved-fill requirement. Across sixteen demand units the plan and a Tweedie gradient-boosting model were not separated: the latter cost between 10% more and 4% less across four requirements, and the other eight alternatives at least 9% more, narrowing to 1% under a common warm-up. Resampling at the 0.98 requirement separated only two alternatives from the plan. Because the plan can carry the buyer's purchasing intentions, this shows no added value of history-based machine learning over such a plan in this portfolio, not a general absence of value. Separating the buffer rule from the forecast mattered more than the choice of model: replacing the fixed percentage mark-up with an error-calibrated buffer raised aggregate fill from 0.84 to 0.99 at unchanged cost, and from 0.75 to 0.99 at a fifth less. Policy simulation at a common service requirement should be the model-selection gate for intermittent-demand inventory systems.