Customer Forecasting Capability in a Process Industry Supply Chain: : A DEA-ML Framework for Evaluation and Prediction

Sammanfattning

Customer-provided demand forecasts constitute a central input to production and logistics planning in supply chain management, particularly in process industry contexts where production flexibility is limited. Forecast evaluation is, however, often conducted reactively and primarily through accuracy metrics, despite the operational relevance of forecast instability and systematic bias. This study develops and evaluates a tripartite analytical framework for assessing customers' forecasting capability in a process industry supply chain context. The framework combines Data Envelopment Analysis, slack-contribution analysis, and eXtreme gradient boosting classification, evaluating forecasting capability at both customer-stock-keeping-unit level and customer-level. The Data Envelopment Analysis model evaluates undesirable forecasting behaviours related to revision magnitude, revision timing, revision variability, directional inconsistency, and upward revision tendency, in relation to forecast accuracy across multiple forecast horizons. Slack-contribution analysis identifies the dominant drivers of poor forecasting capability, and the classifier translates historical capability patterns into proactive forecast accuracy assessment. The framework is empirically tested using forecast and delivery data from a case company in the pulp and paper industry. The results show substantial heterogeneity in forecasting capability, with upward revision tendency emerging as the dominant driver of poor forecasting capability. The classifier outperforms a majority-class baseline and improves progressively across the forecast horizon as additional revisions become available. These findings suggest that systematic forecasting behaviours, rather than isolated forecast errors, are the more informative basis for evaluating customer forecasting capability, and that capability patterns observed historically carry predictive value for forthcoming forecasts. The study contributes by shifting forecast evaluation from individual forecast errors toward multidimensional customer forecasting capability, and by demonstrating that capability assessment can be operationalised proactively rather than reactively. Practically, the framework provides a structured basis for differentiated customer engagement and earlier identification of forecasts likely to deviate from actual demand.