EDGECOURT

How it works

EdgeCourt is NBA player-prop analytics for fans: clear enough for someone seeing a prop line for the first time, deep enough for someone who prices them. It reads the books' lines, gives its own number, shows where the two disagree, and grades itself in public.

It is analytics, never betting advice. It never tells anyone to wager, sizes no stakes, and links to no sportsbook.

New to props? The idea in one minute

A player prop is a line on one player's stat tonight: say, Jalen Brunson over 25.5 points. A sportsbook puts a price on each side, like +130 or −110. Every price carries a chance inside it: the price is only fair if the thing happens that often. EdgeCourt works out its own chance from the player's games, sets it beside the price's, and shows you where the two disagree and by how much. Then, after the game, it checks itself against the box score, in public.

If you only read one thing on a card, read its plain sentence, just above the price. It says our chance, the price, and the chance the price needs, in plain words. Everything below it is the working.

Every pick is logged before the game, and graded after it

This is the part that matters most. Every pick on the nightly board is written to a database when the board is published, before the game is played. Once the game goes final, each pick is graded automatically against the real box score: over or under the line, or a push when the stat lands exactly on it. A player who does not play is never counted as a loss: once the box score shows he did not play, the pick is marked void and left out of the record, like a push.

You cannot edit that record, and neither can we in the ordinary sense: no account can change a pick, and results are written only by the grading job. There is no step at which a bad week can be tidied up. Every pick counts, and the record always shows its denominator.

What the grader records

For each graded pick: the game it was matched to, which stat it graded, the number it read from the box score, and which data source answered. A verdict you cannot check is an assertion, so the evidence is stored alongside the result.

One board a night

EdgeCourt publishes one board each night, the same for everyone, about four hours before the first tip-off. It covers that night's games and three stats: points, assists and threes. Rebounds are not on the board for now.

The prices come from four sportsbooks licensed in Ontario (DraftKings, FanDuel, BetMGM and BetRivers), read about an hour before the board is made. They are those books' US prices; the price at your own book can differ, so every card offers a calculator to enter yours.

How a pick is made

Our estimate. A statistical model projects each player from his own last 82 games at most: his per-minute rates, pulled toward league norms when there is little history, so a small sample is not trusted as though it were a large one, times the minutes we expect him to play tonight. Those minutes start from his recent games, weighted toward the last few, and move for a second game on consecutive nights, a return from a missed game, and regular teammates listed out. That gives a range of outcomes, not a single guess, which is why each card draws the line through a shape. The model's number is then corrected stat by stat, by how its past estimates actually turned out over a held-out season.
The market's number. Each book's price has a margin built in. Taking it out ("no-vig") and averaging across the books gives the market's own chance.
The shortlist. A line makes the shortlist only when our estimate is at least 3 points above the market's, the best price is better than fair on our number, the player is not listed out or doubtful, and the gap is no more than 15 points. A bigger gap usually means the model is missing something, not that the books are.
The review. Claude, Anthropic's AI model, reads the shortlist with the night's availability news. For each line it can keep it, drop it, or move our number by up to 5 points, and it must say why. It never sets the number from scratch and never chooses the grade.
The grade. Strong, Value or Lean is arithmetic, not opinion: how far the final number sits above the price's break-even, measured in the model's own typical error. That error is large (about 13 points on a steady starter's points, more for bench players), so most picks grade Lean, and they are labelled that way rather than left off. The record shows whether Lean picks hold up.

When a team's top scorer is listed out, his teammates' numbers change in a way the ordinary model misses. On those nights the board takes no unders on his teammates, and prices their points with a version of the model fitted on nights like that. The card says so.

Tonight's longshots

Separately from the board, EdgeCourt shows up to three longshots a night: threes lines priced between +500 and +2000 where our corrected estimate sits furthest above what the price needs, never two from the same player or the same game. Longshots miss most of the time by design; that is what the price says. They are not reviewed by the AI, and they are counted on their own line of the record, never mixed into the board's.

Reading a card

A card names the player, the stat, the side and the line (Over 25.5), then draws the range of outcomes our model expects, with the line through it. Under that is one sentence that needs no vocabulary, for example:

Our model gives this a 48.2% chance. At +130, it needs 43.5% to break even, and clears it by 4.7 points.

Then the price row: the best price we found, the book that has it, and the EV at that price. Anything that changes how to read the number, such as a player new to his team or a reviewer's adjustment, is named at the top of the card. Tap Details for the rest:

Three numbers side by side: Model est (our chance), Break-even (the chance the price needs) and Market (the books' own chance). Our case is the gap between the first two; the third shows whether the books agree with us.
The working: the grade, how often a number like this misses, the range we expect for the player, and rows that appear only when they apply, such as a reviewer's adjustment, a top scorer out, the spread of prices across books, a book away from the others, how a price moved, and how picks like this one have done.
The last ten games, every book's price, the context and the reasoning, and the buttons to check your own book's price, add a note or share.

Blue means better than the price; coral means worse. Any label with a dotted underline explains itself in one sentence when you tap it. Those same sentences are below.

Every term, in one sentence

Model est: Our model's chance this happens, worked out from the player's own games and checked against how its past estimates actually turned out.
Model est: On board edges published since Oct 31: the model's estimate pulled part of the way toward the market's number, because the model's own has run high this season. The card says how far.
Break-even: How often something at this price has to happen for the price to be fair: the chance the price itself implies.
Market: The books' own chance once their built-in margin is taken out ("no-vig"), averaged across the books that price this line.
EV: Expected value: how much better or worse this price is than fair, as a long-run percentage, if our estimate is right. It is an average over many games, never a promise about one.
Edge: How many percentage points our estimate sits above the break-even chance.
Grade: How far our estimate sits above break-even, measured in the model's own typical error: Strong is two of those or more, Value one or more, Lean less than one.
Misses: How often a pick with this chance does not happen.
Likely: The range our model sees as likely for this player tonight, and how often ranges like it held last season.
Model's own: The model's own number for this line, before it is pulled toward the market's number.
Adjusted: Our AI reviewer can move our number by up to 5 points, and only with a stated reason.
Priced for: When a team's top scorer is listed out, his teammates are priced with a version of the model fitted on nights like that.
Price range: The best and worst price on this exact line across the books we track; a better price needs to happen less often to be fair.
Off-market: A book whose line or price sits well away from the other books'. Prices like that are often corrected quickly.
Moved: How this book's price changed between the times we checked it.
New team: The estimate comes from his games for another team, so his new role and teammates are not in it. Treat it as less certain than the grade says.
Win chance: Our model's chance each team wins, from the players expected to play, their per-minute production and home court. It is our estimate, not the books' line.
Picks like this: How earlier board picks of the same grade, stat and side have done, shown once at least 10 have been graded.

Where the numbers come from

Lines and prices: The Odds API, timestamped, from the four books above. These are facts, and the app labels them as such.
Box scores and player game logs: ESPN's public endpoints. The same source grades the picks and feeds the model, deliberately: two parsers over one payload eventually disagree, and then grading and the model disagree about what a player actually scored.
Availability and rosters: ESPN's injury report and today's team rosters, shown with their source and time.
Estimates: our statistical model over each player's own game history, reviewed as described above.

What it will not do

No stake sizing, bankroll advice, or bet-size recommendations of any kind.
No claims of certainty. "Edge" describes a computed difference between two chances, never an outcome.
No unverifiable context presented as fact. Anything the model cannot verify from the data it was given is labelled unverified, or is not shown.
No cherry-picking. The record includes every pick, and always shows its denominator.

Honest about what is not proven yet

The model was checked stat by stat against games it had never seen before it was trusted with a number. A backtest is not a record, though, and some of the board's thresholds (the 15-point ceiling, what counts as a book away from the others) are chosen, not yet measured. The live record is how they get measured.

The graded record opens with the regular season on 20 October 2026, at 0–0. Small samples swing; read it with its denominator.

21+ US · 19+ ON

EdgeCourt is for adults: 21 or older in the United States, 19 or older in Ontario and elsewhere. It is for research and entertainment, and it is not financial advice. If play stops being fun, help is available at ConnexOntario, ResponsibleGambling.org and GamblingTherapy.org.

Open EdgeCourt →