Self-evolutionary sibling models to forecast railway arrivals using reservation data

摘要

Accurate forecasts of daily arrivals are of essential to allocate seat resources for transportation companies. Most studies addressed the issue from conventional time series aspects to retrieve historical arrival patterns and project future numbers. This study aims to utilize railway reservation records instead of arrival data to construct self-evolutionary advanced booking models and compare with three benchmarks. In addition, the proposed model involved the spirit of one prototype with multiple versions to pursue accuracy improvement. A family of eight sibling versions based on the curve similarity model, differentiating from the evaluation of similarities among booking curves, was established. The results showed that the constructed sibling versions perform differently with respect to individual data series. In other words, the way of similarity evaluation did affect the predictive performance. Although there was no single version outperforming the others, the selection based on the lowest validation errors was verified to be a good strategy to attain promising out-of-sample performance. Overall speaking, maintaining the family of sibling models for booking data with distinctive characteristics can achieve at least 4.5% and at most 23% improvement of accuracy if comparing with one specific version to all data series. In addition, the proposed sibling models can also outperform popular advanced booking benchmarks such as pick up, regression, and conventional curve similarity approach up to 36%, 32%, and 35%, respectively.