Betting edge is the difference between your fair win probability and the probability implied by the market price. The subtraction is easy. The hard part is earning the fair number.
A calculator can convert odds into implied probability in seconds. It cannot tell you whether your model is calibrated, whether the input data was available before the game, whether the price was actually offered, or whether a hot streak is being mistaken for skill. Most bad edge claims fail before the formula begins.
The edge formula has two named inputs
Write the calculation in plain language: edge = fair win probability - market implied probability. A positive result means your estimate assigns a better chance than the price does. A negative result means the price asks you to pay for a chance your model does not support.
Both inputs need provenance. The market side needs the book, market, price, and capture time. The fair side needs a model version, feature window, training cutoff, and output produced without postgame information. If either side is missing, the displayed edge is not auditable.
The vig and hold guide explains why a raw book probability includes margin. When comparing several sides of the same market, remove or account for that margin before treating the price as a fair market belief. Otherwise the book’s fee is being compared with a model number as though both were built on the same scale.
A fair probability is a forecast, not a fact
The phrase “fair probability” can sound more certain than it is. It is an estimate generated under assumptions. Those assumptions should be visible: what data was used, which outcomes were predicted, how missing values were handled, and whether the model was calibrated on games outside its training sample.
Volume helps only when it matches the question. One cited row says Christian McCaffrey was fantasy’s top scorer after matching a career high with 413 touches . That is a substantial workload observation. It does not, by itself, provide a fair probability for his next game, a team spread, or a player prop.
Another cited row says Jack Campbell finished second in the league with 176 tackles and added five sacks . Those counted events can inform a player or defense model. They still need opponent, role, health, pace, and price context before they become part of an edge estimate.
The point is not that large samples solve everything. It is that each input must be connected to the target. A season workload is better evidence for future opportunity than a stray quote, but it remains only one feature inside a forecast.
Expertise can improve a model without becoming a record
A cited biography note gives Aaron Schatz more than 20 years of NFL analytics experience . That experience can guide feature design, identify data problems, and sharpen assumptions. It is not an ATS win-loss record.
The distinction protects both the analyst and the reader. Biography explains why someone may be worth listening to. A graded record shows what happened to predeclared calls. The two can complement each other, but one cannot substitute for the other.
The other source rows illustrate the problem. They contain directional references to Dallas , the Giants , and Philadelphia . Without picks, prices, timestamps, and outcomes, those mentions cannot be scored. They are commentary, not a denominator.
Convert the price before comparing it
Use the conversion rule that matches the odds format and preserve the original offered price alongside the result. Do not round so aggressively that a small apparent edge is created by formatting. Do not compare a model’s probability for one market with a book price from another.
For a spread, the forecast should refer to covering the posted number at the posted price. For a moneyline, it should refer to winning the game. For a total, it should refer to the stated over or under. The market definition is part of the probability.
The NFL spreads guide covers that contract. The edge formula cannot rescue a mismatched target; clean arithmetic on different events is still wrong.
Calibration decides whether the fair number deserves trust
A model that labels many events with the same probability should see roughly that share occur over a sufficiently broad, out-of-sample set. Calibration checks whether stated confidence matches observed frequency. Discrimination checks whether the model ranks stronger opportunities above weaker ones. Both matter.
Backtests must respect time. Features should be limited to information available before the price being evaluated, and the test games should sit outside the training window. A model that sees closing information while pretending to make an earlier forecast will report an edge it could never have traded.
The calibration-curve guide explains the visual check. The Week One model guide explains why distribution shift can make a historically calibrated number less reliable at the start of a new season.
Grade the claim in the form it was made
If the claim is that a side beats the spread, report the record as ATS wins and losses, the ATS percentage, the evaluation window, and the sample size. If pushes are possible, state how they are treated. Do not replace that record with a bankroll-return summary when the article is making a hit-rate claim.
This source set contains no graded betting calls. It therefore supports no record for the named pundits, teams, or players. The correct output is educational: here is the formula, here are the required fields, and here is why the available mention rows cannot prove an edge.
The track record is where predeclared picks and prices can be graded. A result published after the game is not evidence of a forecast, and a record without a window or sample is not a record readers can evaluate.
A publishable edge statement is reproducible
A strong edge claim names the event and market, the book price and capture time, the model version and fair probability, the vig treatment, and the resulting difference. It also states the conditions that would invalidate the estimate, such as an injury update or a move beyond the trigger price.
That format is less theatrical than “the model loves it,” but it can be checked. Another analyst can reconstruct the comparison, challenge the assumptions, and grade the result against the close.
The formula is the smallest part of the job. Subtract the market probability from the fair probability only after both sides have earned their names. Until then, the number on the screen is not edge. It is an unsupported opinion with a percent sign waiting to happen.
Breakeven win rate at recorded American prices
Breakeven probability is calculated only from American prices that were actually captured in the odds-history table.
