Sports science and performance analytics have historically been built on datasets that underrepresent female athletes. In many cases, when data showed structured variability linked to biological cycles, this was treated as a complication rather than something to incorporate into modelling. As a result, female athletes were often excluded from analysis frameworks that could not easily handle this type of variation.

As an amateur athlete and data analyst myself, I see this very clearly in practice. Many training guidelines, performance models, and recovery assumptions that are widely used today are still largely based on male dominated datasets. Female athletes are not only underrepresented in the data, but the way their performance varies over time is often not properly captured in the models themselves. This leads to a situation where the structure of female athletic performance is simplified, averaged out, or ignored, rather than being understood as part of the signal.

In this talk, I explore this problem through different modelling approaches in Python, comparing standard methods with approaches that explicitly account for structured variability, such as cyclical representations and Generalized Additive Models (GAMs). The focus is on how modelling choices shape what is considered signal versus noise in athlete performance data.

Recent research directions already point toward more inclusive modelling approaches in sports science, including cycle aware performance analysis, individualized training response models, and frameworks that better account for physiological variation in female athletes. These approaches aim to improve training and nutrition recommendations and reduce blind spots that arise when female athletes are not properly represented in the data.

The key question is no longer whether this structure exists, but how we choose to model it and what becomes possible when we do.