โ† Back to PropTensor

System Architecture

How PropTensor Thinks

PropTensor isn't a stat tracker or a trends page. It's a machine learning engine trained on three NBA seasons and four WNBA seasons of player data. For each matchup it weighs 24 inputs (NBA) or 21 inputs (WNBA) to project 17 stat categories.

01Data Quality

Clean Data In, Better Projections Out

If a player rolled an ankle in the first quarter and only played 5 minutes, those stats don't reflect what they normally do. Averaging them in like a regular game makes a player look like they're slumping when they aren't.

PropTensor filters those games out. When a player's minutes fall far below their normal workload โ€” an early exit, foul trouble, or sitting out a blowout โ€” the game is left out of the learning. The engine focuses on games where the player had a normal role.

No peeking at the answer. When the models learn from past games, every input comes only from information that existed before tip-off: recent form, opponent defense up to that date, and so on. That way the models practice under the same conditions they face when you run a real projection.

02Multi-Variable Analysis

24 Inputs (NBA) ยท 21 Inputs (WNBA)

For every projection, the engine looks at the matchup from four angles:

๐Ÿ“Š
Player MomentumLast-10 baseline vs. last-3 trend for the stat, minutes, and shot volume, plus how the player does at home vs. on the road.
๐ŸŽฒ
Game ContextDays of rest, rest advantage, opponent back-to-backs, home court, and rivalry games. For the NBA, it also uses the spread and game total you enter from your sportsbook.
๐Ÿ›ก๏ธ
Opponent MatchupThe opponent's defensive rating, pace, and how they defend this player's position โ€” based on their last 30 days so it reflects trades and injuries.
๐Ÿš‘
Injury ImpactWhen you mark players as out, the engine estimates how much of their usual shot volume shifts to others โ€” for both teams.

Why so many inputs? A single stat like a season average misses a lot: who's hurt, who's tired, and who's guarding whom. Each input covers a different part of the game, so no single number decides the projection.

03The Prediction Engine

A Dedicated Model For Every Stat

PropTensor doesn't use one generic model for everything. Each league has a separate model for each of the 17 stat categories. The points model learns scoring patterns. The assists model learns playmaking patterns.

Each model is tested on games it never saw during training, and that test is where the accuracy numbers below come from.

PTS
REB
AST
3PM
STL
BLK
TOV
P+R
P+A
R+A
S+B
FGA
FGM
FTM
3PA
FAN
DD

04Precision Transparency

We Show You How Accurate The Models Are

Every projection comes with a Mean Absolute Error (MAE): how far off the model was, on average, when tested on games it didn't train on.

If the model projects 24.5 points with a ยฑ4.81 MAE, a typical result lands somewhere around 19.7 to 29.3. Keep in mind this is an average miss, not a guarantee. Real results land outside that range fairly often.

For how the picks actually perform against real lines, check the Track Record, which only counts picks from the current models (NBA-V8 ยท WNBA-V3).

ยฑ4.81

NBA Points MAE

NBA-V8

ยฑ4.49

WNBA Points MAE

WNBA-V3

17

Stat Categories

Per League

05AI Analyst Deep-Dive

The Numbers, Explained In Plain English

The projection is step one. Step two is the Analyst Breakdown: a language model receives the engine's output and writes a 5-step breakdown of the matchup in plain English.

It's instructed to use only the engine's numbers โ€” the projection, the range, the defensive matchup, and the injury inputs โ€” and to explain why the model leans over or under.

The Safe Bet and Lotto Play lines are calculated directly from the model's range by the site itself, not written by the AI, so they always point in the right direction.

06Built-In Safety Nets

The System Warns You When Something Doesn't Add Up

PropTensor doesn't just hand you a number. It checks the projection against recent results and your line, and flags situations worth a second look:

Recent Games Disagree

If the model says UNDER but the player has been clearing the line recently (or the reverse), the system flags it so you know you're betting against a streak.

Typical Range

Every projection shows a typical range based on the model's average miss, so you see more than a single number. If the line sits inside that range, the verdict tells you to be careful.

Too Close To Call

When the projection and the line are within 5% of each other, the system says so instead of forcing an over or under.

Low Data Warning

Rookies and new signings with only a few games on file get a clear warning, because the projection is less reliable until more games are played.

See The Numbers

Run A Projection And Judge The Line Yourself

Get Started

Projections are estimates, not guarantees. Please bet responsibly. If gambling stops being fun, call 1-800-GAMBLER.

PropTensor โ€“ NBA & WNBA Player Stat Projections