Beyond route-specific forecasting: An empirical test of two cross-series transfer learning strategies for airline demand with short-data constraints

Abstract

Airline demand forecasting frequently faces a “short-data” problem, where individual routes have limited historical records due to new route launches, seasonal suspensions, or external shocks such as COVID-19. While this constraint impedes the performance of forecasting models trained on single-series data, the presence of numerous parallel routes—a typical characteristic of airline networks—presents a “short-but-wide” data structure, offering a clear opportunity for cross-series transfer learning.

To exploit this opportunity, we propose and empirically validate two competing strategies: a feature-rich Fine-Tuned Global Forecasting Model (FT-GFM) and an adaptation-focused Adapted Model-Agnostic Meta-Learning (Adapted-MAML).

Our analysis uses 28 months of data for 1203 origin–destination pairs arriving at three major U.S. hubs (ATL, DFW, DEN). The feature-rich FT-GFM improved accuracy substantially, reducing SMAPE by an average of 28.45% and outperforming local models on 86.74% of routes. The data-efficient Adapted-MAML achieved even greater gains, reducing SMAPE by 45.88% and outperforming local models on 93.81% of routes, despite using only historical passenger volumes.

The results validate both strategies as effective solutions and show that, in data-constrained environments, Adapted-MAML's meta-learned initialization yields superior route-level forecasting accuracy compared with FT-GFM's feature-driven approach. These findings provide actionable guidance for airlines on selecting appropriate cross-series transfer learning strategies to mitigate the short-data problem and enhance operational resilience under uncertainty.