The algorithm finds the edge.Our analysts decide if it survives.
Every market on the board enters the engine. It gets scored, stress-tested, priced against the book, and then handed to a person. Most of it never makes it out.
- 010Markets ingestedEvery line on the slate
- 020Survive the modelsSignal clears the bar
- 030Reach an analystPriced, risk-gated
- 040ReleasedApproved and posted
Illustrative volumes for a typical slate. Graded results live on the public record.
Rather than describe the process, here is one market going through it — start to finish.
The machine
Everything on the board goes in. Most of it is thrown out before a model ever runs.
Intake, integrity, and the model bank
Before anything is scored, the engine has to agree that the inputs are real.
Everything that moves a line, on the board at once.
The engine pulls the full picture for every game — not the handful of stats that fit a narrative.
- Team + player performance
- Recent form
- Matchup profiles
- Injuries + projected lineups
- Rest, travel, situational spots
- Sportsbook prices + line movement
- Historical comparables
Bad inputs get killed before they can become a pick.
A model is only as honest as what it was fed. Anything unreliable is dropped at the door.
- Missing or partial data
- Stale prices
- Unconfirmed lineups
- Low sample size
- Conflicting signals
- Unreliable market conditions
Six independent reads. No single model gets a vote of confidence.
Each layer evaluates the market on its own terms. Agreement between them is the signal — a lone outlier is not.
- Performance model
- Matchup model
- Situational model
- Market model
- Historical model
- Volatility model
The read
Where the edge is, what it is worth, and whether it survives being attacked.
We simulate the game thousands of times and map where the edge lives.
Model outputs are resolved into a probability surface. Peaks are where the models converge and the market disagrees. Only what breaks the release threshold moves forward.
- Simulated Outcome Field
- Every modelled result, mapped
- Signal Convergence
- Where the layers agree
- Projected Edge
- Model probability vs market price
- Volatility Zone
- Unstable — discarded
- Release Threshold
- The bar a candidate must clear
Conceptual visualisation of the simulation output. Not a live feed.
The edge is the gap between what the book charges and what the outcome is worth.
The engine converts its own probability into a fair price, then measures it against the number actually on the screen. No gap, no pick.
- True probability
- Simulated outcome frequency
- Fair odds
- What the price should be
- Market price
- What the book is offering
- Edge delta
- The mispricing we get paid for
Edge below our release floor is logged and dropped, not downgraded into a play.
A candidate has to survive being attacked.
Before a human ever sees it, the engine tries to break its own conclusion.
- Price movement toleranceDoes the edge survive the line moving against us?Pass
- Signal stabilityDoes the read hold across model runs?Pass
- Assumption dependenceIs this leaning on one fragile input?Pass
- Volatility ceilingIs the outcome range too wide to price?Pass
- Market correctionHas the book already started fixing it?Pass
The judgement
A person makes the call. Then the result comes back and changes the engine.
AI identifies the opportunity. Analysts validate the decision.
A model cannot read a beat reporter, judge a coach’s language, or tell that a number has already gone stale. Our analysts can. Every candidate is reviewed by a person before release.
- NBA-4471Grizzlies @ Kings
- NFL-2280Steelers @ Ravens
- MLB-9134Padres @ Dodgers
- Context + news interpretation
- Lineup confirmation
- Is the market still playable
- Is the edge actually real
- Is the price still available
- Anything the model can’t see
Rotation confirmed, price held at open. Cleared for release.
ApprovedWe don’t release everything the model finds. We release what survives.
One record, posted to the Discord, logged to the public ledger the moment it goes out — win or lose.
- Posted with the reasoning attached
- Graded in public, including the losses
- Timestamped before the game starts
- Price
- -108
- Book
- Consensus
- Edge
- +4.6%
Every result is fed back into the thing that produced it.
The engine is scored on its own output. So is the desk. What we learn on Sunday changes how Monday gets modelled.
- 01OutcomeWon, lost, pushed
- 02Closing-line valueDid we beat the close
- 03CalibrationDid 60% hit 60% of the time
- 04Model performanceWhich layer carried the read
- 05Signal usefulnessWhat actually predicted
- 06Analyst accuracyWere the overrides right
Machine-scale analysis. Human-level verification.
Every market enters the engine. Only a few make it through. That filter is the product.
Common questions about the process
- How are ProPickz picks actually selected?
- Every market on the board is ingested, checked for data integrity, and scored by six independent model layers. Surviving candidates are simulated into a probability surface, priced against the sportsbook number, and stress-tested at a risk gate. Each one that clears is then reviewed by a human analyst before anything is released. Most candidates never make it out.
- Are picks generated by AI or by people?
- Both, in that order. The models find and price the edge; an analyst has the final say on whether it is released, and can hold or reject one the model likes. Neither side publishes alone.
- Why are so few picks released?
- Because the bar is a price, not a schedule. A pick goes out when the number is wrong by enough to be worth betting, so a quiet slate produces very few. Manufacturing volume to fill a day is how a record gets ruined.
- What does an analyst actually check?
- Context and news the model cannot read: confirmed lineups, whether the market is still playable, whether the price is still available, and whether the edge is real rather than an artefact of one fragile assumption.
- What happens after a pick is graded?
- The result is fed back into the engine. Outcome, closing-line value, calibration, which model layer carried the read, and analyst accuracy all get scored, and that changes how the next slate is modelled. Nothing is deleted, so the calibration data includes every loss.