What Odds Actually Represent, and How to Read Them as Probabilities

Odds look like several different systems and are one system with three notations.

Decimal. The total return per unit staked, including the stake. A price of 2.50 returns two and a half units for one.

Fractional. The profit relative to the stake. Three to one means three units of profit per unit staked, which is decimal 4.00 once the stake is included.

American. A positive number is the profit on a hundred units. A negative number is the stake required to profit a hundred. Positive 150 is decimal 2.50, and negative 200 is decimal 1.50.

Converting between them is arithmetic, and it is worth doing everything in decimal internally because the probability conversion is trivial from there.

That conversion is the whole point: implied probability equals one divided by the decimal price. A price of 2.50 implies 0.40, a price of 1.50 implies about 0.667, and a price of 4.00 implies 0.25. Every analysis downstream starts here, which is why the format question matters less than people expect and the conversion matters more.

Take a two-outcome market and convert both prices to implied probabilities. They will sum to more than one.

That surplus is the margin. It is variously called the overround, the vigorish, or the juice, and it exists because the operator prices the market to hold a share of the total staked regardless of the result.

This has a direct consequence for anyone modeling. A raw implied probability is systematically inflated relative to what the market actually believes, and the inflation is not evenly distributed across outcomes in most pricing schemes. So comparing a model output of 0.42 against a raw implied 0.44 is comparing a probability to a probability plus an unknown share of margin.

Removing it is called devigging, and the methods differ in what they assume.

Proportional. Divide each implied probability by the sum. Simple, and it assumes the margin is applied evenly in proportion to probability, which tends to understate the favorite's true price.

Additive. Subtract an equal share of the excess from each outcome. Assumes a flat margin per outcome, which tends to distort long shots.

Power and shin methods. Fit a parameter so the adjusted probabilities sum to one under a specific assumption about how margin scales, generally handling the favorite-longshot pattern better at the cost of being less transparent.

No method is correct in the abstract. Each encodes an assumption about how the operator distributed the margin, and the right choice depends on the market and should be tested rather than assumed.

Margin varies by market

It is tempting to treat a devigged implied probability as the market's estimate of truth. That is close enough to be useful and wrong in a specific way worth understanding.

A posted price reflects two things at once: what informed participants believe, and what the operator wants to hold. Those diverge when money arrives unevenly. A price can move because new information appeared, or because one side attracted volume and the operator adjusted to manage exposure, and from outside those look identical.

So line movement is information about flow as much as about belief, and reading every move as new knowledge overinterprets it.

What prices genuinely are good at is aggregating. Across liquid markets, prices incorporate injury news, lineup changes, weather, and public sentiment faster than any individual process, which is why a model that disagrees with a liquid market usually deserves scrutiny before the market does.

The honest framing for anyone building models: the market is a strong baseline, devigged prices are the correct thing to compare against, and disagreement is a hypothesis to test rather than an edge to assume.

A few operational points that matter more than the theory.

Store the price and the timestamp. A price without a time is unusable for analysis, because markets move and a comparison to a closing price is a different exercise from a comparison to an opening one.

Store the source. Different operators price differently, and a dataset mixing sources without labeling them makes every subsequent comparison ambiguous.

Keep raw and derived values separate. Keep the original quoted price, and store implied and devigged probabilities as derived fields with the method recorded. When you change devigging method later, and you will, you want to recompute rather than to have lost the original.

Normalize outcome identifiers. The same event and the same outcome are labeled differently across sources, and reconciling them is most of the work in any multi-source odds dataset.

Parlay Ledger works at that layer: ingesting odds across sports, normalizing them across sources, and supporting analytics and parlay modeling on top. The unglamorous part, making prices from different origins comparable with their timestamps and outcome identifiers intact, is the part that determines whether any downstream analysis means anything.