Devig Methods Compared: Multiplicative, Additive, Power and Shin
Devigging converts a book's prices into fair probabilities by removing the margin. Multiplicative scales every implied probability equally, additive subtracts an equal share, power raises each to a common exponent, and Shin models insider money. They agree on close two-way markets but diverge on longshots, so test which one is calibrated on your markets.
What devigging does
A sportsbook's prices imply probabilities that add up to more than 100 percent. The excess is the margin, also called overround or vig. To compare a price with a fair value, the margin has to be removed so the probabilities sum to one. Every devig method does this. They differ only in how they distribute the margin across outcomes, which is a modelling assumption about how the book priced the market.
Start with implied probability q for each outcome, which is 1 divided by decimal odds, and the sum S of those probabilities. S minus 1 is the overround. Each method below maps q to a fair probability p with the p values summing to one.
The four common methods
Multiplicative, also called basic normalisation or proportional: p equals q divided by S. Every outcome loses the same proportion of its implied probability. It is the simplest method and the default in most calculators.
Additive: p equals q minus (S minus 1) divided by n, where n is the number of outcomes. Every outcome loses the same absolute amount. In markets with many outcomes this can push a longshot below zero, which makes it unsuitable for outrights.
Power: p equals q raised to a power k, with k solved numerically so the probabilities sum to one. Because k is above one when there is margin, small probabilities shrink proportionally more than large ones. Joseph Buchdahl's writing on the wisdom of the crowd calls this the logarithmic method.
Shin: derived from Hyun Song Shin's papers in the Economic Journal in 1992 and 1993, which model a book setting prices when a share z of the money comes from insiders. The fair probability is p equals the square root of (z squared plus 4 times (1 minus z) times q squared divided by S), minus z, all divided by 2 times (1 minus z), with z solved so the p values sum to one. The insider model produces a favourite longshot bias, so Shin also takes more margin out of longshots.
Worked example: a close two-way market
Take an illustrative moneyline of -150 and +130, decimal 1.667 and 2.30. Implied probabilities are 60.0 and 43.5 percent, an overround of 3.48 percent. The fair favourite probability comes out at 57.98 percent multiplicative, 58.26 percent additive, 58.26 percent Shin, and 58.40 percent power.
The spread between methods is under half a percentage point. On prices like these, method choice rarely changes a decision. Note that Shin and additive match exactly. With only two outcomes the two methods are equivalent, a point the documentation for the R package implied also makes.
Worked example: where they disagree
Now take an illustrative -400 and +300, decimal 1.25 and 4.00, with a 5.0 percent overround. The fair underdog probability is 23.81 percent multiplicative, a fair price of about +320. It is 22.50 percent additive and Shin, about +344. It is 21.76 percent power, about +360.
Suppose another book offers the underdog at +340, decimal 4.40. Expected value per unit is fair probability times 4.40 minus one. Multiplicative says +4.8 percent. Additive and Shin say -1.0 percent. Power says -4.2 percent. One method flags a value bet; the other two say the price is fair or poor. The disagreement is concentrated exactly where bettors most often look for value: longshot prices from a market with a fat margin.
Three way markets
With three or more outcomes Shin separates from additive. Take an illustrative football match at 1.40, 4.80 and 8.50, an overround of 4.03 percent. The fair probability of the 8.50 outsider is 11.31 percent multiplicative, 10.42 percent additive, 10.55 percent power and 10.66 percent Shin, with an estimated insider share z of about 2 percent. The favourite moves the other way: 68.66 percent multiplicative against 69.72 to 70.22 percent for the rest. Multiplicative again treats the longshot most generously.
Why longshots drive the difference
Betting markets have long shown a favourite longshot bias: longshots tend to win less often than their prices imply, so books load more margin onto them. Multiplicative assumes the margin is spread proportionally and therefore leaves longshots too likely if the bias is present. Additive, power and Shin all take more out of the longshot, by different amounts. If a market has no such bias, multiplicative can be the better fit. The method is a hypothesis about the market, not a neutral calculation.
What comparative studies report
Erik Štrumbelj's 2014 paper "On determining probability forecasts from betting odds" in the International Journal of Forecasting found that probabilities from Shin's model beat basic normalisation, and regression-based approaches, for every bookmaker and sport pair it tested, scored on the ranked probability score. Stephen Clarke, Stephanie Kovalchik and Martin Ingram compared additive, multiplicative, power and Shin in "Adjusting bookmaker's odds to allow for overround" in the American Journal of Sports Science in 2017, across three large bookmaker datasets from three different sports, and reported that the power method universally outperformed multiplicative and outperformed or matched Shin. The documentation for the implied package says basic normalisation tends to be the least accurate of the methods it offers. So the two head-to-head studies point in different directions, Štrumbelj's toward Shin and Clarke's toward power, and note that each used a different scoring rule from the log loss or Brier score recommended below. None of this settles the question for a specific market.
How to choose a method for a dataset
Treat the choice as a calibration test. Collect historical closing prices for the markets in question, preferably from a sharp book. Devig them with each method. Score the fair probabilities against outcomes with log loss or Brier score, and check calibration by probability bucket, paying attention to the longshot buckets. Use the method that scores best out of sample, and re check when the source book or market type changes.
Open source implementations exist: the implied package for R, and the shin and penaltyblog packages for Python. Reimplementing is straightforward, but compare outputs against one of them to catch errors in the root finding.
Frequently asked questions
- Which devig method is most accurate?
- There is no universal winner. Štrumbelj's 2014 paper reports Shin's method beating basic normalisation across every bookmaker and sport it tested, Clarke, Kovalchik and Ingram's 2017 comparison reports power outperforming multiplicative universally and matching or beating Shin, and the implied package notes basic normalisation tends to be least accurate. But accuracy depends on the sport, the book and the market. Score each method on historical closing prices for the markets in use and pick the best calibrated one.
- Why does the multiplicative method overrate longshots?
- It removes the same proportion of margin from every outcome. If the book actually loaded extra margin onto longshots, which the favourite longshot bias suggests is common, proportional removal leaves those longshots looking more likely than they are. Additive, power and Shin all remove relatively more from small probabilities, which usually lowers the longshot's fair probability.
- Is Shin's method the same as additive?
- Only in two-way markets. With two outcomes the Shin solution matches the additive one exactly, as the illustrative -150 and +130 example shows. With three or more outcomes they diverge, because Shin's insider model shapes how margin is spread across outcomes of different sizes, while additive removes an identical amount from each one.
- Can devig methods produce negative probabilities?
- The additive method can. It subtracts the same absolute amount from every outcome, so in a market with many outcomes and a large margin, a longshot's implied probability can be smaller than its share of the overround. That makes additive unsuitable for outright and futures markets. Multiplicative, power and Shin always return probabilities between zero and one.