Cycling Race Time Prediction: A Personalized Machine Learning Approach Using Route Topology and Training Load

TL;DR

This study tackles the practical problem of predicting cycling duration for a given route by replacing physics-based parameterization with a data-driven model that combines route topology features with athlete fitness derived from training-load metrics. Using an N-of-1 design on 96 rides from a single cyclist, the authors develop a progressive prediction framework and show that a Lasso model with Topology + Fitness features achieves MAE of 6.60 minutes and R^2 of 0.922, improving over topology-only predictions by about 14%. Key contributions include novel terrain-derived features (e.g., punchiness, ClimbPro-inspired climb detection), rigorous leakage-free feature engineering, and demonstration of progressive checkpoint predictions on a real MTB route (Track 101 MTB). The work highlights the practical potential for personalized, pre-ride planning and dynamic in-ride updates, while acknowledging limitations such as the single-athlete scope and absence of environmental factors, and outlines clear avenues for multi-athlete validation and integration of weather data.

Abstract

Predicting cycling duration for a given route is essential for training planning and event preparation. Existing solutions rely on physics-based models that require extensive parameterization, including aerodynamic drag coefficients and real-time wind forecasts, parameters impractical for most amateur cyclists. This work presents a machine learning approach that predicts ride duration using route topology features combined with the athlete's current fitness state derived from training load metrics. The model learns athlete-specific performance patterns from historical data, substituting complex physical measurements with historical performance proxies. We evaluate the approach using a single-athlete dataset (N=96 rides) in an N-of-1 study design. After rigorous feature engineering to eliminate data leakage, we find that Lasso regression with Topology + Fitness features achieves MAE=6.60 minutes and R2=0.922. Notably, integrating fitness metrics (CTL, ATL) reduces error by 14% compared to topology alone (MAE=7.66 min), demonstrating that physiological state meaningfully constrains performance even in self-paced efforts. Progressive checkpoint predictions enable dynamic race planning as route difficulty becomes apparent.

Figures (22)

• Figure 1: Elevation profile of La Desértica MTB race (110 km, 2443m ascent). Top: altitude with 13 detected climbs (ClimbPro: $\geq$3%, $\geq$500m) highlighted in red. Bottom: gradient distribution showing sections above 5% (orange) and 8% (red).
• Figure 2: Punchiness score analysis. Top: gradient profile with overall route punchiness $P$ (standard deviation of gradient changes). Bottom: rolling punchiness over 1km windows for visualization, identifying sections with highest gradient irregularity. The model uses the global $P$ as a single feature; the rolling view illustrates where variability concentrates.
• Figure 3: Punchiness score comparison across three routes with similar distance and elevation but different gradient variability. High punchiness (left) indicates frequent gradient changes requiring repeated power surges; low punchiness (right) shows steady grades allowing constant pacing. This metric captures difficulty beyond total elevation.
• Figure 4: Recovery distance analysis. Climbs (red, $g > 3\%$) and recovery zones (green, $|g| < 2\%$ within 500m of climb ends). Flat sections after climbs allow W$'$ reconstitution; routes with minimal recovery distance accumulate fatigue faster.
• Figure 5: Gradient distribution analysis showing three complementary views: histogram of gradient values, cumulative distance in gradient buckets, and sustained gradient over rolling 500m windows. Routes with similar total elevation can have vastly different gradient profiles, affecting physiological demand.