[Paper Review] Modelling the effect of training on performance in road cycling: estimation of the Banister model parameters using field data

Phil Scarf, Mansour Shrahili+3 more|arXiv (Cornell University)|Feb 6, 2019

Sports Performance and Training41 references4 citations

TL;DR

This study models the effect of training on road cycling performance using field-collected power and heart rate data, estimating parameters of the Banister training-performance model. It finds that model parameters vary significantly across individuals and performance metrics, indicating that personalized, performance-specific models are essential—rendering generic models unsuitable for optimizing training schedules in advance of competition.

ABSTRACT

We suppose that performance is a random variable whose expectation is related to training inputs, and we study four performance measures in a statistical model that relates performance to training. Our aim is to carry out a robust statistical analysis of the training-performance models that are used in proprietary software to plan training, and thereby put them on a firmer footing. The performance measures we consider are calculated using power output and heart rate data collected in the field by road cyclists. We find that parameter estimates in the training-performance models that we study differ across riders and across performance measures within riders. We conclude therefore that models and their estimates must be specific, both to the individual and to the quality (e.g. speed or endurance) that the individual seeks to train. While the parameter estimates we obtain may be useful for comparing given training programmes, we show that the underlying models themselves are not appropriate for the optimisation of a training schedule in advance of competition.

Motivation & Objective

• To conduct a robust statistical analysis of training-performance models used in proprietary training software.
• To estimate Banister model parameters using real-world field data from road cyclists.
• To assess whether generic training models are suitable for optimizing competition-ready training schedules.
• To evaluate how performance measures (e.g., power output, heart rate) influence model parameter estimates.
• To determine whether model parameters are consistent across cyclists or vary by individual and performance quality.

Proposed method

• The study models performance as a random variable whose expectation depends on training inputs, using field-collected power and heart rate data from road cyclists.
• Four distinct performance measures are defined and analyzed to assess their relationship with training load.
• Statistical modeling is applied to estimate parameters of the Banister training-performance model from empirical data.
• Parameter estimation accounts for individual variability and performance-specific training goals (e.g., speed vs. endurance).
• The analysis uses a statistical framework to test model adequacy and assess the suitability of the model for training schedule optimization.
• Comparative analysis of parameter estimates across cyclists and performance measures is conducted to evaluate model generalizability.

Experimental results

Research questions

• RQ1 How do Banister model parameters vary across individual road cyclists when estimated from field data?
• RQ2 Do parameter estimates differ depending on the performance measure (e.g., power output vs. heart rate) within the same cyclist?
• RQ3 To what extent are the underlying training-performance models suitable for optimizing training schedules in advance of competition?
• RQ4 Are the model parameters consistent across different performance qualities (e.g., speed, endurance) within a single athlete?
• RQ5 Can the statistical model reliably estimate performance outcomes based on training inputs in real-world cycling conditions?

Key findings

• Parameter estimates in the Banister model vary significantly across individual cyclists, indicating that one-size-fits-all models are inadequate.
• Within the same cyclist, parameter estimates differ across performance measures, suggesting that training effects are not uniformly distributed across performance types.
• The study concludes that training-performance models must be personalized to both the individual and the specific performance quality being targeted (e.g., speed or endurance).
• Generic models derived from aggregated data are not suitable for optimizing training schedules in advance of competition.
• The statistical analysis confirms that model parameters are sensitive to both individual physiology and the choice of performance metric.
• The findings underscore the need for individualized, performance-specific modeling rather than reliance on standardized training models.