A Sequential Tour-Based Mode Choice Framework with GPS-based Data: Integrating Machine-Learning-Derived Features into Mixed Logit and MNL Models

This study develops a sequential, tour-based mode choice framework using smartphone-based revealed-preference travel data from Sydney, Australia, modeling the complete daily travel chain rather than isolated trips. Building on earlier machine-learning-based feature engineering for the first trip of the day, we extend that logic to four interdependent components: the first trip of the first tour (FT1), the first trip of non-first tours (FTn), immediate previous-tour dependence (IPT), and within-tour trips (WTT). For each, machine-learning-informed variables are embedded in interpretable Multinomial Logit (MNL) and Mixed Logit specifications. A central contribution is the explicit treatment of the estimation–forecasting gap. To examine how upstream uncertainty affects downstream prediction, three information structures are evaluated: Observed–Observed (OO), Simulated–Simulated (SS), and Simulated–Observed (SO). This design assesses model-level predictive performance and error propagation across daily sequences. Mixed Logit gains differ markedly across components: modest for FT1 (+ 0.6 pp) and WTT (< 0.1 pp), but substantial for FTn and IPT (+ 5.3 and + 5.2 pp), where inter-tour decisions involve genuine behavioral flexibility. Sequential lag parameters are universally non-significant, indicating mode choices are governed by stable habitual preferences rather than tour-to-tour switching. The framework is highly stable across information structures: OO–SS accuracy differences are at most 1.7 pp, and complete-sequence accuracy is marginally higher under SS than OO (MNL: OO 61%, SS 63%; Mixed Logit: OO 63%, SS 67%). Practically, agencies and modellers can estimate sequential tour-based models directly on simulated upstream inputs without sacrificing accuracy, simplifying the models that underpin travel demand and policy analysis.