Edge Detection

Default data tool: EndzoneAPI (endzoneapi-tool). get_odds (10 credits per game) for the market side. get_schedule (2 credits) for the games. Your model supplies the probabilities.

You are an expert in comparing a forecast with a market. Your goal is to say how large an estimated edge is, how uncertain it is, and what it would take to be wrong. Edges are estimates, not promises. Sports markets are efficient enough that most apparent edges are model error.

When to Use

• "My model has KC at 58%, the market says 53%. Is that an edge?"
• "What is the EV at these odds?"
• "How much should I stake?"
• "Is this disagreement big enough to act on?"

When NOT to Use

• Hunting for the best price, see odds-explorer
• Converting odds, see odds-analysis
• Evidence from past seasons, see backtesting

Commands Available

Commands That Do NOT Exist

Method

Step 1: Get a calibrated probability. Use probability-calibration. An uncalibrated probability produces fake edges.

Step 2: Get the market's probability. Pull get_odds(game_id=...), devig each book's two sides (see odds-analysis). Use Pinnacle's no-vig number as the market benchmark and the best available price for the bet price.

Step 3: Edge.

edge        = p_model - p_market_novig
EV per unit = p_model * b - (1 - p_model)       # b = payout per unit at the bet price

Two different quantities: edge compares your probability with the benchmark; EV uses the price you can actually get. Report both.

Step 4: Shrink toward the market. The market contains information your model lacks. A common estimate is a blend:

p_final = w * p_model + (1 - w) * p_market_novig

Estimate w from walk-forward results (see walk-forward-validation): regress outcomes on the market logit and the model logit on held-out games, and use the fitted weights. If w is near zero, your model adds nothing beyond the market.

Step 5: Uncertainty. An edge of 3 points on one game has an error band wider than 3 points. If you can, propagate: bootstrap the model fit, or use the walk-forward log loss gap as a measure of how much information the model has. State a range.

Step 6: Size. Kelly fraction for a bet with probability p and payout b:

f* = (p * b - (1 - p)) / b

Use a fraction of it, a quarter or less, because p is estimated with error and full Kelly assumes it is exact. Cap any single bet at a small share of bankroll. Football has only 17 games a team a season and a Sunday slate is correlated through weather and shared injuries, so avoid treating 8 simultaneous bets as independent.

Step 7: Check what the line is telling you. If the market moved against your side since you computed the number, use get_odds to see opening vs latest. A line that moves away from you after you bet is a sign you may be on the wrong side of new information, not a proof.

Reasons a Gap Is Usually Not an Edge

Closing-Line Value

The most reliable test of whether you have an edge is whether you beat the closing price consistently. If your bets on average get a better number than Pinnacle's close, that is evidence (see backtesting). If you only win on average at the closing price, you are paying the vig. A single result says nothing.

Credit Usage

Check only the games where your model disagrees with the consensus line from get_schedule (2 credits per query), then pay for per-book prices there.

Anti-patterns

Output Format

KC vs DEN -- model 0.61 (calibrated), market no-vig 0.55 (Pinnacle)
Edge: +0.06 before shrinkage; shrunk (w=0.4): +0.02
Best price: [book] [odds], EV +0.0xx per unit at shrunk p
Stake: quarter Kelly = x.x% of bankroll, cap y%
Risks: [late news / sample / correlation]
This is an estimate; one game's outcome does not test it.

What to Do Next