Experimental AI-assisted analytics — predictions and ratings are estimates, not official results.
APEX · Alliance Performance Expectation

How the APEX model works

A team's APEX rating estimates how many points they are expected to contribute above an average team, adjusted for opponent strength, alliance partners, event difficulty, and recent performance.

Experimental, AI-assisted analytics

predictFRC uses AI-assisted experimental analytics. Predictions, rankings, ratings, and insights are not guaranteed to be accurate and should not be treated as official results or professional advice. Data may be incomplete, delayed, miscategorized, or incorrect. predictFRC is a new project under active development, and accuracy will improve over time.

The core idea, in one sentence

Take everything a team has done this season, figure out how hard it was given who they played with and against, weight recent and high-stakes matches more heavily, and express the result as a single number: expected points contributed above average.

What goes in

The model blends many signals. None on its own tells the whole story — the rating comes from how they fit together.

Match scoresAlliance partnersOpponents facedEvent strengthScore breakdownsAuto pointsTeleop pointsEndgame pointsFouls & penaltiesRanking pointsRecent-match weightingPlayoff performance

Component ratings

Offensive rating

Points a team adds through scoring, above an average robot.

Defensive rating

Points a team prevents opponents from scoring.

Auto rating

Expected contribution during the autonomous period.

Endgame rating

Expected contribution from endgame and climb tasks.

Every metric, explained

What it means · how it's calculated · what feeds it · when to trust it.

The adjustments that matter

  • Partner & opponent adjustment

    Your numbers are corrected for the strength of who you played with and against, so easy schedules don't inflate ratings.

  • Event difficulty

    A strong performance at a tough event counts for more than the same result against a weak field.

  • Recency weighting

    Recent matches carry more weight, so the rating reflects current form, not the early-season version of a team.

  • Playoff weighting

    Elimination matches are weighted up because they test teams under real pressure.

Handling missing & messy data

  • Missing breakdowns: if phase-level data is absent, the metric falls back to a league-average prior and is tagged Data incomplete.
  • Tiny samples: ratings are pulled toward the mean until enough matches exist, preventing wild week-1 swings.
  • Surrogates & backups: excluded or down-weighted so stats aren't misattributed.
  • Outliers: a single lucky or unlucky match is damped, never allowed to dominate a rating.
  • Delayed data: stale values show an older "last updated" timestamp and reduced confidence.

How confidence scores work

Confidence (0–100) combines four factors: sample size (more matches → higher), data completeness (fewer missing fields → higher), recency (fresher data → higher), and matchup closeness (coin-flip games are inherently less certain). Low-confidence outputs are labeled, not hidden.

How predictions are tested

Every completed match feeds an audit comparing predicted vs actual winner and score. We track winner accuracy, average score error, and confidence calibration — whether a stated 70% really wins ~70% of the time. See the full Accuracy Report →

Avoiding overreaction

The model uses recency-weighted averaging with regression to the mean and outlier damping. One blowout win or a broken-down match nudges a rating slightly rather than rewriting it, so APEX tracks genuine improvement without chasing noise.

2026 REBUILT game intelligence

A game-specific config layer the model reasons over.

Estimate 2026 REBUILT scoring values are placeholders pending official manual ingestion.

Autonomous

First seconds: robots run pre-programmed routines with no driver input.

  • Auto Leave3 pts
  • Auto Score (Low)4 pts
  • Auto Score (High)6 pts

Teleop

Drivers control robots to acquire and score game pieces.

  • Teleop Score (Low)2 pts
  • Teleop Score (High)5 pts
  • Fast Cycle Bonus1 pts

Endgame

Final seconds: high-value parking / climbing objectives.

  • Endgame Park4 pts
  • Endgame Climb12 pts

Ranking point objectives (estimated)

  • 2 ranking pointsMatch WinAlliance wins the qualification match.
  • 1 ranking pointsMatch TieQualification match ends in a tie.
  • 1 ranking pointsScoring Threshold RPAlliance exceeds an estimated objective scoring threshold.
  • 1 ranking pointsEndgame RPAlliance meets an estimated endgame objective.
  • 1 ranking pointsCoopertition BonusBoth alliances cooperate to meet a shared objective (lowers a separate threshold).AI-assisted

REBUILT scouting metrics

Real vs estimated vs AI-inferred.

Game Piece ThroughputAverage game pieces scored per match.Estimate
Auto ReliabilityHow often the autonomous routine executes as intended.AI-assisted
Cycle TimeAverage time for one acquire-and-score cycle.Estimate
Scoring AccuracySuccessful scores divided by attempts.AI-assisted
Preferred Scoring LocationWhere the robot scores most often.AI-assisted
Driver EfficiencyPath efficiency and decision quality under pressure.AI-assisted
Defense ResistanceAbility to keep scoring while defended.Estimate
Endgame ReliabilityHow often the endgame objective is completed.Estimate
Ranking Point ContributionEstimated RP this team helps its alliance earn.AI-assisted
Penalty RiskLikelihood of incurring fouls (lower is better).Estimate
Alliance Role FitBest complementary role on an alliance.AI-assisted

The data pipeline

How real FRC data will flow once connected.

  1. 1

    Fetch

    Pull raw events, teams, matches and score breakdowns from public APIs.

  2. 2

    Validate

    Reject malformed payloads, flag missing fields and duplicate keys.

  3. 3

    Normalize

    Map every source into a single internal schema for teams, events and matches.

  4. 4

    Categorize

    Split score breakdowns into game-specific phases (auto / teleop / endgame).

  5. 5

    Compute Metrics

    Derive APEX contributions, consistency, momentum, schedule difficulty, etc.

  6. 6

    Assign Confidence

    Score each output by sample size, data completeness and recency.

  7. 7

    Store

    Cache computed outputs so pages render instantly without recomputation.

  8. 8

    Display

    Render with provenance labels, freshness timestamps and uncertainty.

FRC knowledge the model assumes

Qualification matches

Randomized round-robin matches that set the ranking order.

Ranking points

Bonus points for wins, ties, and meeting game objectives — separate from match score.

Alliance selection

Top-seeded teams draft partners to form playoff alliances.

Playoff brackets

Double-elimination bracket that decides the event winner.

District vs regional

Districts use a cumulative points system across events; regionals advance winners directly.

District points

Points earned at district events that determine district championship qualification.

Championship divisions

Large championships split teams into divisions that feed a final playoff.

Strength of schedule

Some teams simply draw harder partners and opponents — APEX corrects for it.

Coopertition

Game-specific incentives that reward both alliances for cooperating.

Auto / teleop / endgame

The three scoring phases, each modeled separately.

Fouls & penalties

Rule violations award points to the opponent and add noise to raw scores.

Surrogate matches

Extra matches that don't count toward a team's ranking.

Backup robots

Substitute robots called in during playoffs — flagged so stats aren't misattributed.

Event advancement

How teams progress from districts to district champs to worlds.

Season improvement

Teams iterate their robots across weeks; recency weighting captures it.

Raw score vs true contribution

A high score can come from strong partners — APEX isolates a team's own contribution.

A note on honesty

APEX is an original model built for predictFRC. It is not affiliated with, nor a copy of, any other analytics provider. Ratings are estimates — they describe expected performance, not certainties, and this build uses simulated data.