Probabilistic Forecasting of Airline Cockpit Crew Reserve Demand Across Multiple Planning Horizons

Abstract

Reserve cockpit crew allocation at major airlines is typically determined by scaling historical averages to planned network demand, without uncertainty quantification or systematic incorporation of operational drivers such as crew illness rates, seasonal absence patterns, and calendar effects. Within the broader workforce forecasting literature, uncertainty quantification is largely absent, and moreover, only two prior studies have addressed reserve cockpit crew demand forecasting directly, neither of which covers the European short-haul fleet or produces probabilistic forecasts across multiple planning horizons.

This paper presents a probabilistic forecasting framework for cockpit crew reserve demand, evaluated on a cockpit division at KLM Royal Dutch Airlines across forecast horizons of 30, 90, 180, and 360 days. A direct multi-step forecasting strategy is applied, with dedicated models trained per horizon. A diverse set of classical, machine learning (Random Forest, LightGBM), and deep learning models (DeepAR, NHiTS, Temporal Fusion Transformers) is benchmarked on point accuracy and probabilistic calibration, with quantile regression and level-set forecasting (LSF) as uncertainty quantification methods.

The experimental results show that the LightGBM model, optimized via quantile regression, achieves the lowest mean pinball loss at every horizon, leading the next-best model by up to 38.0%, but exhibits under-coverage relative to the empirical 90% target at forecasting horizons of 30, 90, and 180 days. Despite its strong probabilistic performance, quantile LightGBM’s point accuracy is modest: its mean absolute error (MAE) is only 4.5% lower than a naive mean baseline at its best horizon, underscoring how little exploitable signal a point forecast alone can extract from this series. The best-performing point model overall, Random Forest, improves on the mean baseline by 9.1% MAE, still a modest margin that highlights the intrinsic difficulty of the forecasting problem and motivates a probabilistic approach.

Feature importance and ablation studies show that autoregressive lags and incidental illness rates dominate short-horizon predictions, while planned network demand becomes the primary driver at H = 360; weather alarm indicators are uninformative across all horizons. The framework is extended to planned reserves, voluntary reserves, and incidental illness as separate targets: voluntary reserve usage is up to roughly four times harder to forecast than planned usage on a normalized scale, while incidental illness is the most predictable sub-target.

The results establish a proof of concept for data-driven probabilistic reserve planning on the European short-haul fleet and provide a basis for an explainable decision support tool for operational crew planning. The probabilistic forecasting framework reveals that reserve demand predictability differs across forecast horizons, quantile levels, and reserve types, information that a single point-forecast approach cannot provide.