Estimating the Estimate : Using Machine Learning to Forecast Train Delay Reliability in Real Time

Sammanfattning

Real time ETAs (Estimated Time of Arrival/departure) published by railway opera- tors such as Trafikverket revise frequently between issuance and realisation, but the operational system does not quantify how much each revision is likely to still be wrong. This gap matters for both passengers and operators, wanting trust for decision-making and being able to reroute and manage connections, respectively. Most prior data-driven delay forecasting work either competes with the operational baseline from scratch or studies one model and one interval construction method in isolation. This thesis instead frames delay forecasting as a correction and improvement of the operational ETA and compares three model families on the same task, evaluating both point accuracy and the calibration of their native prediction intervals. A snapshot-keyed dataset was constructed by collecting Trafikverket’s API every five minutes for one month (2026-02-02 to 2026-03-02), recording all state changes to SJ long-distance ETAs along with the realised outcome. Ridge regression, Random Forest and LightGBM were trained to predict realised delay at each snapshot, with the operational ETA included as a feature, both as real time delay estimation and historical ETAs, so that improvements can be interpreted as corrections to the operational forecast. Models were then compared on common accuracy metrics (MAE, RMSE and R2) as well as on skill against the operational ETA, with results stratified by forecast horizon and delay severity, and 90% prediction intervals evaluated by empirical coverage. All three models improved on the operational ETA on RMSE, with skill ranking by complexity (+0.10 for Ridge, +0.21 for Random Forest, +0.24 for LightGBM). Skill was concentrated in long horizons and severe delays. In the on-time bin, models added no value over the baseline. ETA history features dominated importance rankings, and weather features were essentially uninformative. The three interval construction methods produced qualitatively distinct failure modes. LightGBM’s quantile regression achieved near nominal coverage (90.7%), Random Forest’s tree-prediction-percentile severely under-covered (27.6%), and Ridge’s independent quantile fits produced frequent bound crossings. The findings establish that real time API snapshots, using the operational systems in place, can both improve point predictions and communicate calibrated uncertainty, given the right model and interval construction. The snapshot-keyed schema and the skill versus baseline framing are transferable to other domains with revising ETAs. Results are conditional on a single winter month and a single operator, motivating future work in multi-season collection and conformal calibration.