Machine learning sharpens airline route passenger forecasts, with clear limits

A new machine-learning study reports that gradient-boosted tree models can forecast monthly passenger traffic on individual airline routes with substantially better accuracy than simple persistence-based benchmarks, but only under specific near-term conditions. The research, published open access in Discover Artificial Intelligence on 27 October 2026, built a capacity- and competition-aware forecasting framework at the carrier-route-month level using U.S. Bureau of Transportation Statistics T-100 records, and it is unusually explicit about where its predictions work well and where they break down.

The study’s authors, Muhammad Zeeshan Younas and Muhammad Rizwan Younas, assembled twelve annual T-100 data files covering January 2015 through February 2026, yielding 134 months of observations. After removing 382,977 exact duplicates from 4,732,929 raw records, they aggregated the remaining 4,349,952 unique observations by carrier, origin, destination, and month. Requiring at least 24 months of available history and applying feature-construction criteria produced a final modeling panel of 1,388,451 observations spanning 22,713 eligible carrier-route combinations and 10,400 directional routes. Directional markets were kept distinct, so a flight pair such as Atlanta-Orlando and Orlando-Atlanta was treated as two separate markets.

A central methodological choice was exact calendar-month alignment. Lagged predictors and forecast targets were matched by calendar month rather than by shifting to the next available row, and missing months were never bridged. The atypical January 2020 through December 2021 disruption period was excluded from the primary modeling sample. Targets were defined by target dates: observations with targets through December 2022 formed the training set, 2023 targets formed the validation set, and targets from January 2024 through February 2026 formed the held-out test set. Feature selection, imputation, and tuning used training data only, and the test period was never touched during model development.

On a common aligned test sample of 222,218 observations, XGBoost delivered the strongest one-month-ahead passenger-traffic performance, with a weighted mean absolute percentage error of 11.91 percent, a root mean squared error of 1,434.92 passengers, and an R-squared of 0.9742. Compared with the last-month naive benchmark, which simply carries the current month’s traffic forward, XGBoost reduced WMAPE by 16.72 percent and RMSE by 18.35 percent. Random Forest and HistGradientBoosting followed closely with WMAPE values of 12.24 percent and 12.28 percent. A multilayer perceptron neural network reached only 14.45 percent WMAPE, slightly worse than the naive benchmark’s 14.30 percent, though the authors note the neural model used numeric predictors only and is not a fully feature-equivalent comparison.

The framework extended beyond point traffic forecasts. For next-month load-factor prediction, XGBoost achieved 8.37 percent WMAPE and an R-squared of 0.8309, reducing WMAPE by 15.91 percent relative to a current-load-factor persistence benchmark. The study also converted realized next-month load factors into three utilization-state classes: below 0.70 as low utilization, 0.70 through 0.85 as moderate, and above 0.85 as high. HistGradientBoosting and XGBoost produced effectively identical classification results, each with approximately 73.37 percent accuracy and a macro-F1 of 0.7413, outperforming Random Forest, class-weighted logistic regression, and the original neural classifier. Most classification errors occurred between adjacent utilization categories, and the authors stress these are descriptive screening categories, not measures of profitability or signals for automatic schedule changes.

Feature-family ablation experiments revealed which information sources actually matter. A history-only specification produced 17.31 percent WMAPE; adding focal-carrier capacity variables cut that to 14.27 percent, the largest single improvement. Adding direct competition variables brought only a marginal further reduction to 14.20 percent, and the full specification including airport-network and co-airport variables reached 14.26 percent, slightly worse than the history-capacity-competition model. Relative to history alone, the history-capacity-competition specification reduced WMAPE by 18.00 percent and RMSE by 23.03 percent. Competition and co-airport information proved more valuable in a different task: forecasting how inbound traffic is allocated among airports within twelve multi-airport metropolitan systems, including Dallas-Fort Worth, Chicago, New York, the San Francisco Bay Area, and Miami-Fort Lauderdale. There, adding co-airport variables reduced WMAPE by 3.94 percent relative to history and capacity.

Reviewer-requested robustness analyses added important caveats. Removing contemporaneous operating information and using only strictly lagged predictors increased WMAPE by 9.39 percent on a 333,999-observation robustness sample, confirming that the primary model is a near-term end-of-month forecast that treats the current month’s completed operating record as available, not a long-lead schedule forecast issued before the month begins. Rolling-origin backtesting showed regime-dependent performance: the strictly lagged model was weaker than last-month persistence in the 2022 recovery window and in 2023, roughly tied in 2024, and superior only in the January 2025 through February 2026 window. On a common sample of 251,644 observations evaluated across all horizons, WMAPE rose monotonically from 14.46 percent at one month to 19.25, 22.00, and 28.89 percent at three, six, and twelve months, while R-squared declined from 0.9661 to 0.8814.

Performance also varied sharply by market size. Thin-market errors were numerically smaller in absolute passengers but proportionally much larger, because small absolute deviations represent a large share of observed traffic on sparse routes. The segments were defined retrospectively from realized test-period demand and used only for post hoc diagnosis. Split-conformal prediction intervals achieved near-nominal marginal coverage of 89.58 percent for nominal 90 percent intervals and 94.79 percent for nominal 95 percent intervals, but coverage was not uniform: high-volume markets were undercovered at 81.23 and 89.93 percent respectively, while thin and medium markets were conservatively covered.

The authors are careful about interpretation. T-100 passenger counts represent realized traffic under the capacity actually supplied, not unconstrained latent demand, and the records omit fares, bookings, itinerary connections, traveler purpose, macroeconomic variables, fuel prices, spill, and revenue-management controls. The traffic-forecasting task is conditional on the carrier-route being observed as active in the exact target month; the framework does not separately predict route entry, suspension, or exit, and a continuity diagnostic found that among 1,388,451 origins, 1,221,144 had an active exact next-month target while 113,948 were classified as temporary disappearance or intermittent service and 21,042 as possible suspension or exit. An illustrative asymmetric cost analysis showed the model’s value flips with assumed loss asymmetry: HistGradientBoosting reduced mean cost by 3.78 percent relative to persistence under symmetric costs and by 14.34 percent when overforecasting carried twice the penalty, but increased cost by 7.03 percent and 18.10 percent when underforecasting was twice and five times as costly. These are sensitivity calculations, not realized financial savings.

The study’s overall conclusion is deliberately bounded. Calendar-aligned gradient boosting is presented primarily as a near-term passenger-traffic forecasting and utilization-screening tool rather than a universal demand or long-range planning solution. Capacity variables provided the largest improvement over traffic history, competition and co-airport variables produced smaller, task-specific gains, and forecast accuracy deteriorates with horizon even when the evaluation population is held fixed. The authors suggest future work should model route continuation jointly with traffic, incorporate booking, fare, planned-schedule, macroeconomic, and revenue-management information, and develop segment-conditional or adaptive probabilistic intervals, alongside sparse-route methods and stronger local explainability for deployment.

Subject of Research: Machine-learning forecasting of monthly passenger traffic, load factors, and utilization states on airline routes using U.S. T-100 data

Article Title: Capacity and competition aware machine learning for airline route passenger traffic forecasting and operational decision support

Article References: Younas, M. Z., & Younas, M. R. (2026). Capacity and competition aware machine learning for airline route passenger traffic forecasting and operational decision support. Discover Artificial Intelligence, 6(1), Article 1660. https://doi.org/10.1007/s44163-026-02300-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02300-1

Keywords: airline forecasting, machine learning, XGBoost, passenger traffic, load factor, route-level prediction, T-100 data, gradient boosting, forecast horizons, prediction intervals, aviation analytics, utilization classification