Model Guide

Scorecards and Predictions

How Triathlon Lab scores athletes across distances, then turns those scorecards into race reads, start-list analysis, and model odds.

Core idea

Historical strength is separate from race prediction

Scorecards answer how strong an athlete has been recently for a specific distance and discipline.

Predictions ask what happens when that athlete enters a specific start list with a specific course, field, weather profile, and tactical race shape.

Distance

Full, middle, short-course, and broad long-course views stay separate.

Discipline

Overall, swim, bike, and run scorecards are scored independently.

Context

Start-list predictions adjust for the actual field and race shape.

Scorecard families

One athlete can have several true scores

A 70.3 score, a full-distance score, a WTCS score, and a split score can tell different stories. The model chooses the right lens for the race instead of forcing every athlete into one universal ranking.

tl_overall_ironman

Full-Distance Specialists

Full IRONMAN and Challenge full-distance races.

Primary evidence is full-distance racing. Middle-distance evidence can support thin profiles, but it is discounted and capped so a 70.3 specialist is not treated as a proven full-distance favorite.

tl_overall_middle_distance

Middle-Distance Specialists

IRONMAN 70.3, T100, 100K, half, and middle-distance races.

Primary evidence is middle-distance racing. Full-distance results can support the profile when needed, but full-distance durability and middle-distance speed are not treated as identical skills.

tl_overall_long_course

Long-Course PowerIndex

A broad long-course power view across full, 70.3, T100, and Challenge races.

Useful for general long-course strength, while race-specific tools prefer the distance-specific full or middle profiles.

tl_ironman_pro_series

IRONMAN Pro Series

Official-style IRONMAN Pro Series ranking context.

Winners receive event maximum points and every second behind the gender winner subtracts one point. Best five results count, with no more than three full IRONMAN races.

tl_wtcs_race

WTCS Race Rankings

WTCS race evidence in the main PowerIndex model.

A WTCS-specific lens for World Triathlon Championship Series racing.

tl_short_course_wt

Short-Course Specialists

WTCS and Olympic/championship-final-level short-course racing.

Athletes need current elite short-course evidence to appear. World Cup and Continental Cup results can support eligible athletes, while junior, U23, and development races are excluded for now.

Start lists

Race distance chooses the active profile

WT / WTCS races use tl_short_course_wt.
Full-distance races use tl_overall_ironman.
Middle-distance races use tl_overall_middle_distance.
Nearby fallback profiles can fill gaps, but fallback scores are heavily reduced.

Prediction overview

Predictions are structured scouting reports

The output should explain who is strongest, who controls the race shape, who needs a T2 gap, who can run people down, and where the uncertainty lives.

1Distance-specific scorecards show recent athlete strength.
2Swim, bike, and run scorecards identify discipline weapons.
3Timing signatures show how athletes usually move through a race.
4Race context profiles describe whether similar races were swim-bike pressure races, runner races, bike bridge races, or balanced races.
5Weather and course context can change how costly a chase or bike-heavy race becomes.
6Start-list snapshots compare athletes against the actual field entered in that race.

Detailed reference

Formulas and model notes

These are the current high-level mechanics behind the scorecards and prediction layers. The database keeps evidence rows so model output can be traced back to the races carrying the score.

PowerIndex overall race score

powerindex_score =
  position_score * 0.35
  + sof_score * 0.35
  + time_score * 0.30

Position score rewards placing, with steeper or softer drop-off based on race tier.
SOF is strength of field, either stored or conservatively seeded from tier.
Time score compares the athlete's finish time to the gender winner's finish time.
The current PowerIndex build uses finished results from the past 52 weeks.

Full-distance aggregation

2+ primary rows: average(best 2 primary) + small depth bonus, capped at 100
1 primary + fallback: average(primary + best fallback), capped at 85
1 primary only: primary * 0.82, capped at 78
fallback only: fallback average, capped at 72

Primary full-distance evidence is full IRONMAN plus Challenge full.
Fallback middle-distance evidence is discounted and kept in dark-horse territory.

Split evidence score

split_score =
  split_rank_score * 0.35
  + split_time_score * 0.20
  + split_sof_score * 0.35
  + split_dominance_score * 0.10
  + split_premium_bonus

Evidence score is capped by split tier ceiling, then multiplied by profile discount.
Run split dominance is field-strength gated so soft-field blowouts do not overwhelm premium-field evidence.

Race forecast v2

race context
-> result movement samples
-> athlete race-shape profiles
-> start-list forecast tables

The current builder writes race contexts, result samples, checkpoints, and athlete race-shape profiles from recent pro results.
The schema also includes forecast summary, athlete role, phase, and scenario tables.

Building race view...

Predictions are structured scouting reports

The output should explain who is strongest, who controls the race shape, who needs a T2 gap, who can run people down, and where the uncertainty lives.

1Distance-specific scorecards show recent athlete strength.

2Swim, bike, and run scorecards identify discipline weapons.

3Timing signatures show how athletes usually move through a race.

4Race context profiles describe whether similar races were swim-bike pressure races, runner races, bike bridge races, or balanced races.

5Weather and course context can change how costly a chase or bike-heavy race becomes.

6Start-list snapshots compare athletes against the actual field entered in that race.

2+ primary rows: average(best 2 primary) + small depth bonus, capped at 100 1 primary + fallback: average(primary + best fallback), capped at 85 1 primary only: primary * 0.82, capped at 78 fallback only: fallback average, capped at 72