Aveyx
Where Aveyx started

🏀 NBA Analytics

Before MLB, before soccer — this model was built on basketball first. Every player, every night, run through a real multi-factor prediction pipeline: recency-weighted form, matchup and pace context, schedule fatigue, live injury impact, and market calibration — not a single flat average dressed up as a prediction.

20+

Live Factors

11

Props Tracked

5

Signal Groups

82-26

Games/Season

How It Actually Works

Five Signal Groups, One Number

Every prediction starts from a season/career baseline, then moves through these five groups in order — each one a real, capped adjustment, not a black box.

Player Form

How good is this player right now, not just this season?

  • Season + career baseline, dynamically blended
  • Median-based outlier trimming (IQR) so one 40-point night doesn't skew the read
  • Last-3-vs-prior momentum detection (hot/cold streak)
  • Role-change detection when recent usage diverges sharply from history

Matchup Context

Who are they facing, and how fast will the game actually be played?

  • Composite opponent defense rating vs. league average
  • Shot-zone matchup analysis — their shot profile vs. what the defense allows there
  • Primary defender-specific adjustment
  • Real possessions-based pace factor (FGA − OREB + TOV + 0.44×FTA)

Schedule & Fatigue

Tired teams and cold benches produce different numbers.

  • Back-to-back discount (0 days rest)
  • 3-games-in-4-nights fatigue, distinct from the back-to-back check
  • Blowout + coach-tendency risk, from real point-differential history
  • Home/away performance split

Injury Impact

Missing teammates and missing defenders both move the number.

  • Live ESPN injury report, both teams
  • Historical with/without-teammate performance, not a flat guess
  • Opponent's missing defenders factored the same way
  • Return-from-injury minutes regression for the player themselves

Market Calibration

The sportsbook line already prices in things we don't see — respect that.

  • Stacked ML ensemble blended with the statistical model
  • Tiered regression toward the line when our number diverges sharply
  • Team possession ceiling — can't exceed a player's realistic share of the scoring pie
  • Confidence capped 5–80%, widened for real fat-tail/outlier risk

What confidence actually means

A 65% confidence pick is expected to hit roughly 65% of the time — not 95%, not a lock. Real basketball has real variance; the model widens its own uncertainty band when the signal is genuinely volatile instead of pretending every pick is a sure thing.

Sample Factor Breakdown

Illustrative example only — a hypothetical points prop, not a live prediction or a real player.

Baseline → Final

24.625.1

Props Covered

PointsReboundsAssistsPRA3-Pointers MadeStealsBlocksQ1 PointsQ1 ReboundsQ1 AssistsQ1 PRA

Run It On Tonight's Slate

Pick a player and see every one of these factors broken out line by line.

Open ML Predictions