Project M11: Cross-Event Performance Transfer & Sparse Imputation in Competitive Swimming

1. The Sparse Repertoire Problem

In competitive swimming, complete event profiles do not exist. Unlike running or cycling — where one fitness marker carries fairly cleanly across distances — swimming is shaped by water resistance and the very different techniques of its four strokes: Freestyle, Backstroke, Breaststroke, and Butterfly.

Large audits of competition results show just how thin most race histories are:

• Over 42% of club and regional athletes never enter the 400m Freestyle, opting instead for pure sprint programmes (50m and 100m events).
• Over 75% have no recorded time at all in secondary events such as the 200m Butterfly, 200m Breaststroke, or 400m Individual Medley.
• Split histories by pool length: a swimmer may have a full record in Short Course Metres (25m pool), but only one or two swims in Long Course Metres (50m Olympic pool) from a summer championship.

For training and session-planning tools like Sesh 3.0, this thin data is a fundamental problem. Working out an athlete's Critical Swim Speed ($v_{\text{CSS}}$) needs an aerobic distance anchor (traditionally a 400m time). If a sprinter has only ever raced 50m and 100m Freestyle, older methods fall back on squad averages or crude guesswork. Project M11 was built to solve this missing-data problem properly.

2. Why Simple Rules Fail

Before building anything new, we tested the standard rules of thumb coaches and swim software have long relied on:

The Single Best Points Fallacy (Conventional Peak Rule)

The most common shortcut assumes an athlete will score a similar World Aquatics (FINA) points score in an untried event as their best in their main event. Tested on unseen races, this fails badly: a mean percentage error (MAPE) of nearly 4.0% and an average miss of over 7.3 seconds. Only 37.8% of predictions land within the standard acceptable threshold of ($\pm 2.0\%$).

The reason is physical: a swimmer with 650 FINA points in the 50m Breaststroke has specialist whip-kick timing, ankle flexibility, and explosive power. Assuming the same 650 points in the 200m Backstroke ignores that Backstroke depends on continuous flutter kick and shoulder mechanics that the breaststroker may never have developed.

The Mean Points Heuristic Trap

A tempting alternative is to average a swimmer's points score across everything they have raced. However, this approach blunts the unique strengths of specialists:

Suppose an athlete's best event is the 100m Butterfly (600 FINA points). They also swam a 100m Backstroke once at an off-season development meet, swimming poorly (380 points). Averaging the two drops their rating to 490 points. Predicting their untried 200m Butterfly from that average comes out nearly 6 seconds slower than what they are truly capable of.

The Cross-Season Timing Trap

A third common error is comparing swims from very different points in the season. A fully rested, end-of-season 200m Butterfly from a July national championship cannot be directly paired with a 400m Freestyle swum tired in heavy winter training in November. In our tests, mixing dates across the season added up to 1.8% of pure timing error.

3. How Project M11 Works

To estimate missing events without falling into these traps, Project M11 combines statistical cross-event transfer models with multi-dimensional pattern analysis across competitive swimming histories:

4. The Key Modelling Principles

1. Dynamic Form Alignment

A central challenge in sports modelling is that historical performances vary across competitive cycles. A rested taper swim from a summer championship cannot be directly compared to an un-tapered winter league swim without context.

By aligning historical benchmarks with contemporaneous meet indicators, the system maintains high predictive coverage while eliminating the seasonal drift inherent in raw historical personal bests.

2. Non-Linear Distance Transfer

Conventional scoring tables assume a universal scaling relationship across race distances. In reality, speed decays non-linearly as distance increases, and this fatigue rate differs significantly across stroke specialisations and developmental levels.

Modeling event-to-event transfer directly accounts for the reality that a 50m sprint specialist fatigues at a fundamentally different rate over 200m or 400m than a middle-distance specialist.

3. Multidimensional Performance Profiling

To avoid watering down stroke and distance specialists, Project M11 moves beyond simple distance or points averages:

The model evaluates a broad range of individual performance characteristics, including event specialisation, stroke proficiency, pacing behaviour, and competitive history. These signals are combined to establish a more complete representation of an athlete's capabilities than conventional single-event or average-performance approaches.

4. Multi-Event Stroke Inference

Multi-event competition records contain valuable indicators of athlete capabilities beyond their primary competitive events. Project M11 incorporates these additional performance signals to develop a more comprehensive understanding of individual stroke proficiency and competitive potential, providing useful performance indicators even for events an athlete has never contested as a standalone race.

5. Validation & Results

Project M11 underwent an extensive iterative development and validation process. Successive research stages evaluated different approaches to performance transfer, athlete specialisation, historical calibration, and cross-event relationships. These investigations informed the final proprietary modelling framework, which demonstrated substantial improvements over conventional performance estimation methods.

The model was evaluated against an independent athlete-level holdout test population comprising over 180,000 verified competition swims from more than 10,000 athletes, with no swimmers shared between the training and evaluation cohorts. Target events were completely masked on both short-course and long-course formats to simulate true zero-history prediction.

Performance Across Athlete Profiles

Across the independent holdout population, Project M11 achieves an overall average percentage error of approximately 1.99% (median error: 1.42%), with nearly two-thirds of predictions landing within ±2.0% of actual race times.

Accuracy systematically improves as an athlete's competitive race history deepens. For active club competitors with core multi-event track records, average error drops below 1.8%, and reaches under 1.4% for experienced regional and national-calibre competitors where pacing and technical consistency are most established.

Middle and Distance Event Transfer

The framework demonstrates particularly strong performance in middle-distance and distance events — precisely where traditional estimation methods struggle most due to low participation rates. In events such as the 400m, 800m, and 1500m, where athletes frequently lack recent competition marks, cross-event transfer models achieve median holdout errors well under 1.0%.

6. Natural Repeatability and Prediction Boundaries

An important finding from our research is that competitive swimming performance is not perfectly repeatable. Even when an athlete contests the same event repeatedly within a single competitive season, measurable natural variation exists between performances.

This natural variability provides valuable context when interpreting prediction accuracy:

• Realistic Accuracy Context: When evaluating model performance, prediction error should be interpreted relative to the natural variability of competitive swimming rather than assuming human performance is perfectly deterministic.
• Calibrated Confidence Bands: Because race execution inherently fluctuates, predictions are most practically represented as high-confidence target ranges rather than rigid single points.

Performance Transfer Across Known and Unfamiliar Strokes

Prediction confidence varies according to the completeness and nature of an athlete's available competition history:

• Extrapolating New Distances Within a Known Stroke: Models have strong foundations when an athlete has previously demonstrated proficiency in the target stroke. In these scenarios, the model primarily estimates distance endurance and pacing curves, delivering its highest consistency.
• Estimating Unfamiliar Strokes: Estimating performance in strokes where an athlete has no recorded competition history introduces additional physiological uncertainty. Core hydrodynamic mechanics — such as breaststroke kick timing, hip alignment, and ankle flexibility — cannot always be inferred from performances in other events.
• Confidence-Calibrated Intelligence: The system automatically communicates prediction confidence based on the depth of stroke-specific and multi-event data available for each individual.

7. Integration with the MyAthlete Ecosystem

Project M11 forms part of MyAthlete's broader performance intelligence ecosystem. Its predictions support downstream athlete analysis and training calibration, allowing incomplete competition histories to contribute to more comprehensive individual performance profiles:

• Automated Anchor Resolution: Incomplete race histories are supplemented automatically, enabling individualised training analytics for athletes who rarely contest distance events.
• Longitudinal Athlete Intelligence: As swimmers compete in new events, predictions continuously update to reflect emerging specialisations and technical development.
• Cross-Product Synergy: Performance transfer insights directly enrich Sesh 3.0 training calculations, bridging competition results and pool-deck training demands.