What Is Expected Value in Sports Analytics?

Expected value is the single most useful — and most misquoted — idea in sports analytics. It answers a precise question: if a decision were repeated many times at the same probability and the same price, what would the average result be? This guide defines it, works through real prices, shows why it is only as good as the probability behind it, and explains how it shows up in a graded record as ROI.

One thing it does not do is predict a single result. A decision with positive expected value loses often; one with negative expected value wins often. Expected value is about the long run, and the long run is longer than most people think.

The definition

For a two-outcome decision, expected value per unit staked is the probability of winning multiplied by the net profit if it wins, minus the probability of losing multiplied by the stake lost. With decimal odds d and win probability p, that is EV = p × (d − 1) − (1 − p) × 1. When EV is zero the price is fair; when it is positive the price is better than the probability deserves; when it is negative the price is worse.

The whole calculation rests on p. The price is public; the probability is an estimate. Every claim of positive expected value is therefore a claim that your estimate of p is better than the price implies, and that is a claim that has to be tested against results.

A worked example at -150

American odds of -150 mean risking 150 to win 100, which is decimal 1.667. The break-even probability — the p at which EV is zero — is 150 ÷ 250 = 60.0%. Now suppose your estimate for the favorite is 62%. EV = 0.62 × 0.667 − 0.38 × 1 = 0.413 − 0.380 = +0.033, or about +3.3% of the stake per repetition. If instead your estimate is 58%, EV = 0.58 × 0.667 − 0.42 = 0.387 − 0.420 = −0.033, about −3.3%.

Notice how thin the margins are. A four-point difference in the probability estimate swings EV from clearly positive to clearly negative. This is why calibration — whether your 62% calls come true 62% of the time — matters far more than any single number.

From price to break-even probability

The first thing to compute for any price is the probability you would need for the decision to break even. For negative American odds the break-even is |odds| ÷ (|odds| + 100); for positive odds it is 100 ÷ (odds + 100).

The table makes a familiar point concrete: the standard -110 on a point spread needs 52.4% to break even, so a spread method must be right more often than it is wrong just to stand still.

The vig, and why the market’s own number is your baseline

Add the break-even probabilities of both sides of a market and you get more than 100%. At -150 / +130, the favorite needs 60.0% and the underdog 43.5%; the sum is 103.5%. The extra 3.5 points is the bookmaker’s margin — the vig. Remove it by scaling both sides so they sum to 100% and you get the fair, market-derived probabilities: 58.0% and 42.0%.

That fair probability is the baseline against which any expected-value claim should be measured. It is the number Sportslyx shows on every game page — the moneyline pair de-vigged, not a model’s guess — because a claim of value has to start from what the market believes with its margin stripped out. Against the fair 58.0%, the -150 price carries a negative expected value equal to the favorite’s share of the margin; you need to believe the true probability is above 60.0% before the price is attractive, and you need evidence for that belief.

Correlated legs and combos

Expected value becomes treacherous when outcomes are combined. The naive way to price a two-leg combo multiplies the two probabilities, which is only right if the legs are independent. In the same game they rarely are: a favorite covering the spread and the total going over are correlated when the favorite’s path to covering is scoring a lot. Multiplying independent probabilities then understates the joint chance, and the “value” you compute is an artifact of the wrong model.

The correct approach prices legs from one model of how the score is distributed, so that correlation is built in. That is how the Sportslyx Combo Builder works: same-game legs are priced from a single score distribution rather than multiplied together, and the expected value shown is the honest one for the ticket, not for its parts.

Variance and sample size

A positive-EV method still loses in streaks. Suppose a method is genuinely 55% at even money, an edge of +10% per unit. Over fifty results the standard deviation of the hit rate is about seven percentage points, so a 48% month is entirely ordinary. Over five hundred results it is about 2.2 points, and a 48% stretch would be surprising. The edge is real in both cases; only the second sample is large enough to see it.

This is the reason Sportslyx measures its record over a fixed calendar month per sport and prints the settled volume next to every figure. Expected value is a statement about many repetitions; a record that does not show how many repetitions it contains cannot be read.

Expected value in a graded record

Realized ROI is expected value observed after the fact. On the Sportslyx pick tracker, ROI is measured at a flat stake per read: total profit or loss divided by total amount staked, with each read counting one unit. Hit rate is wins divided by wins plus losses, with pushes excluded. The two can point in different directions — a method that takes long prices can have a low hit rate and a positive ROI, and a favorite-only method the reverse — which is why both are published.

Some grading rules change the arithmetic and are stated openly. A soccer draw on Sportslyx is a push, so a losing outcome is removed from the record while the 1X2 price is kept; that inflates the soccer ROI relative to a true draw-no-bet price. Voids — a cancelled fight, a player who did not play — are removed rather than counted as losses. Knowing the rules is part of knowing what the ROI means.

Limits of the idea

• EV assumes the probability is right; it cannot check itself
• It assumes the price is available at that number, and closing lines move
• It says nothing about a single result, and single results are what people remember
• Correlation breaks the simple formula; combined outcomes need a joint model
• Past ROI is a noisy estimate of EV, not a forecast of the next month

Used with those limits in mind, expected value is the clearest way to reason about any estimate in sport. The methodology page fixes the definitions Sportslyx uses; the sports analytics guides cover probability and calibration in more depth.