System Architecture
PropTensor isn't a stat tracker or a trends page. It's a proprietary machine learning engine trained on two full seasons of NBA data, analyzing every matchup through 15+ quantitative variables to generate institutional-grade projections across 18 stat categories.
01Data Quality
Garbage In, Garbage Out — So We Eliminate The Garbage
Most models treat every game equally. If a player rolled their ankle in the first quarter and only logged 5 minutes, those terrible stats get averaged in like a normal game. The model thinks the player is declining. It's wrong.
PropTensor doesn't make that mistake. Our proprietary data cleaning pipeline automatically identifies and removes contaminated games — early exits, ejections, illness, and low-rotation noise. The engine only learns from games where the player actually competed.
The result: cleaner training data produces sharper projections. Every fraction of a point matters when you're looking for an edge.
02Multi-Variable Analysis
15+ Variables. Every Angle Covered.
For every prediction, the engine evaluates the matchup across four proprietary intelligence layers:
Why 15+ variables? Simple models use 3 or 4 inputs and get simple results. Our engine captures the full context of a matchup — the same way a professional analyst would break down a game, but with mathematical precision across thousands of data points.
03The Prediction Engine
18 Specialized Models. One For Every Stat.
PropTensor doesn't use one generic model for everything. We train 18 dedicated prediction models — one for each statistical category. The points model understands scoring patterns. The assists model understands playmaking patterns. Each brain is a specialist.
This means when you ask for a rebounds projection, you're getting a model that was specifically built, tuned, and validated for rebounding — not a one-size-fits-all guess.
04Precision Transparency
We Tell You Exactly How Accurate We Are
Every projection comes with a Mean Absolute Error (MAE) — the model's average miss distance on historical data. This isn't a vague confidence rating. It's a hard number.
If the model projects 24.5 points with a ±4.69 MAE, you know the realistic range is roughly 19.8 to 29.2. That floor and ceiling lets you evaluate the line with real math, not gut feeling.
±0.57
Blocks MAE
Tightest Model
±4.69
Points MAE
Core Model
18
Stat Categories
Full Coverage
05AI Analyst Deep-Dive
Machine Learning Meets Natural Language Reasoning
The projection is step one. Step two is the Analyst Deep-Dive — a large language model that receives the engine's full output and writes a 5-step breakdown of the matchup in plain English.
It's grounded exclusively in our engine's data. No hallucinated stats, no recycled narratives. The AI explains why the model likes the over or under, references the defensive matchup, flags injury impacts, and gives you alternative lines to consider.
Think of it as having a quant analyst sitting next to you, walking you through every angle of the bet — except it runs in seconds and never forgets a variable.
06Built-In Safety Nets
The System Warns You When Something Doesn't Add Up
PropTensor doesn't just hand you a number. It cross-checks its own projection against recent results and flags potential traps:
Trend Divergence Alert
If the model says UNDER but the player has been crushing the over recently, the system flags it as a potential "Hot Trap" — warning you that you're betting against a streak.
Floor / Ceiling Range
Every projection includes a mathematically validated floor and ceiling based on the model's historical precision — so you always know the realistic range, not just a single number.
Stop Guessing
Deploy The Engine & Find Your Edge