Learn
How to find value bets, and how to tell a real edge from a measurement error
Finding value is two jobs stacked on top of each other: produce a probability the market disagrees with, and then work out whether the disagreement is real or just noise in your own estimate. The second job is the one that gets skipped.
Educational only. quant+ does not take wagers and cannot guarantee a return.
A price is a value bet when your probability times the decimal price is greater than one. Everything else on this page is about where the probability comes from and how much of it to believe.
You believe the side wins 55% of the time. p = 0.55
The best price available is -105. d = 1.9524
Expected value per unit is positive. 0.55 x 1.9524 - 1 = +7.38%
The fair price for 55% is: 1 / 0.55 = 1.8182, about -122
At -130 the same belief is a losing bet. 0.55 x 1.7692 - 1 = -2.69%
The same probability is a value bet at one price and a bad bet at another. Value is a property of the pair, not of the side.
Get a probability from one of two places
There are exactly two honest routes to the number on the left-hand side of that test. They have different costs, different failure modes, and different evidence requirements.
Route A: your own estimate
You form a probability from ratings, a model, or a structured judgement about a specific situation the market has not priced. This is the only route that can find value in a market where the price is genuinely wrong, and it is also the only route whose error you cannot see.
The failure mode is not being wrong. It is being wrong in a way that feels like insight, because a considered opinion about a team is much easier to form than it is to validate.
Cost: high. Evidence needed: a long graded record, or a demonstrated calibration.
Route B: borrow a sharp number
Take a two-way market from a venue whose prices are known to be sharp, remove the margin, and treat the result as an estimate of the true probability. Then look for a softer price elsewhere that beats it. The estimate is not yours, which is the point.
The failure mode is different: a sharp benchmark is a good estimate on average and is still capable of being wrong on any single game, so a disagreement with it is a lead rather than a verdict.
Cost: low. Evidence needed: a benchmark that actually is sharp in that market.
What the de-vig actually produces
Removing the margin from a two-way market means dividing each implied probability by their sum. The benchmark venue does not have to be named; what matters is that the two prices come from the same market at the same moment.
| Sharp two-way price | Overround | Fair probability | What you must beat |
|---|---|---|---|
| -115 / -105 | 4.71% | 51.08% / 48.92% | -115 or better |
| -120 / +100 | 4.55% | 52.17% / 47.83% | -120 or better |
| -110 / -110 | 4.76% | 50.00% / 50.00% | -110 or better |
| -125 / +105 | 4.34% | 53.25% / 46.75% | -125 or better |
The last column is the whole method. Once the margin is out, the fair probability is a target, and any venue offering a longer price than that target on the same side of the same market is offering value against the benchmark.
The working order, from picking a market to grading the bet
This is the sequence that produces a defensible bet rather than a hopeful one. The order matters: several of these steps exist only to stop a later step from fooling you.
Pick one market and learn its prices before you look for edges
The same edge in a market you understand and a market you do not are different bets. Choose a league and a market type and stay in it long enough to know what a normal price looks like, because an anomaly is only visible against a baseline.
Form the probability before you see the price
This is the step that prevents the most common failure, which is anchoring. Once you have seen a price it is very hard to un-see it, and an estimate formed after the fact drifts toward the number you were shown. Write the number down first, with a range.
Put a range on it, not a point
A probability of 55% is not a fact about the world, it is a summary of an estimate with error around it. State the error. A belief of 55% plus or minus 3 points is a different input from a belief of exactly 55%, and the next step depends on which you have.
Convert the market price to the same units
Turn the price into a required probability so the two numbers are comparable. At -110 the market is asking for 52.381%; at +120 it is asking for 45.455%. Comparing a probability with a price without converting is comparing two different things.
Discard marginal disagreements
If the offered price sits inside your range, you have no bet, not a small bet. A price that the market and your own uncertainty both endorse is a coin flip with a commission attached. This step removes most candidates and it is supposed to.
Shop the price before you commit
The same side at -105 instead of -110 is worth about 2.27 percentage points of return, and it costs nothing. Price is the only part of this process you control with certainty.
Log the bet with the price and the close
Recording the price you took and the price it closed at turns each bet into evidence about your process rather than a data point about your luck, and it produces a reading after dozens of bets instead of thousands.
Steps 03 and 05 are the ones people drop, and they are the two that carry the entire difference between a value bet and a number that looks like one.
Most measured edges are smaller than they look
When you only bet on the games where your estimate disagrees with the price, you are selecting the games where your estimate is most likely to be wrong. This is the winner's curse, and it has a size.
Suppose your probability estimates carry a typical error of about three percentage points, and you bet whenever your measured edge exceeds two points. The games that pass the filter are not a random sample. They are disproportionately the games where your error was positive, which means the average edge you measure on the bets you place is larger than the average edge that exists.
| Your true edge | Edge you measure on selected bets | How much of it is real |
|---|---|---|
| 0.0 points | 3.80 points | none of it |
| 1.0 point | 4.06 points | 25% |
| 2.0 points | 4.39 points | 46% |
| 3.0 points | 4.80 points | 63% |
Read the first row carefully. A bettor with no edge at all, filtering on a two-point measured edge with three points of estimation error, will see an average measured edge of 3.80 points across the bets they place. The number is real, the arithmetic is correct, and the edge does not exist.
Tighten the estimate
The size of the curse is set by your error. Halving the error roughly halves how much inflation you see on selected bets, which is the argument for better inputs rather than for a better filter.
Raise the bar, and shrink what passes
Demanding a bigger measured edge removes the worst candidates and does not remove the effect, because the games that clear a high bar are still the games where your error ran in your favour. The honest correction is to shrink the edge you act on, not to trust it more.
Prefer the borrowed number
A de-vigged sharp price does not have this problem in the same way, because the error being selected on is the market’s rather than yours, and the market’s error is much smaller. This is the strongest practical argument for Route B.
When you scan hundreds of games, most flags will be noise
A screen does not find value. It finds candidates, and the mix of candidates it returns is governed by the base rate of real opportunities and the false positive rate of your filter.
The arithmetic of a scan
Take a market where roughly one price in forty is genuinely soft, and a filter that flags five percent of games as opportunities when nothing is there. Scan 200 games.
real opportunities 200 / 40 = 5
false flags 200 x 0.05 = 10
total flags 15
share that are real 5 / 15 = 33%
Two out of three flags are artefacts of the filter. This is not a criticism of the filter; it is what a base rate of two and a half percent does to a five percent error rate. The only way to change the ratio is to lower the false positive rate or raise the base rate by working in a market you know is soft.
Three ways a flag is not an opportunity
- A stale price. A number that has not moved because nobody has updated it is not a soft number, it is an old one, and it will be repriced before or during your bet.
- A different market. Two prices that look comparable can settle under different rules, and the gap between them is the rule difference rather than an edge.
- A limit you cannot get. A price that only stands for a small stake is a real price and not a usable one, and a screen that ignores stake size is reporting an opportunity you cannot take.
What a price has to be to clear a threshold
Turning a belief into a minimum acceptable price is one division. If you want a margin of m on top of a probability p, the decimal price must be at least (1 + m) / p.
| Your probability | Fair price | With 2 points of margin | With 3 points of margin |
|---|---|---|---|
| 55% | 1.8182 (-122) | 1.8545 (-117) | 1.8727 (-115) |
| 45% | 2.2222 (+122) | 2.2667 (+127) | 2.2889 (+129) |
A useful habit is to decide this threshold before you look at what is available. Deciding it afterwards, when you can see the price, is the anchoring failure from step 02 in a different costume.
Record the six fields that make a bet auditable
A bet log is not a diary. It is the dataset that will tell you, later, whether your probability estimates are any good, and it only works if the fields are recorded before the result is known.
Your probability, and the date you formed it
Recorded before the price is seen where possible, and always before the result. This is the field that lets you measure whether your estimates are calibrated.
The price you actually took
Not the best price you saw. The price you got, at the size you got it, because that is the only number that describes the bet you made.
The closing price, and the timestamp
The close is what your price gets compared against, and the timestamp is what makes the comparison valid. A price captured hours before the start is not a close.
The reason, in one line
Writing down why a bet qualified forces the reasoning to be explicit, and it makes patterns visible later: a cluster of bets that all failed for the same reason is a finding, and it is invisible in a win-loss column.
The stake, and the rule that set it
Staking recorded separately from selection is what lets you tell a bad selection process from a bad sizing rule. The two produce the same bankroll curve and need different fixes.
The market and the venue type
Recording where the price came from, without necessarily naming it, separates the performance of different sources. It is common to find that one source of candidates has been contributing most of the losses.
What following this process does not give you
It does not guarantee a profit. The process raises the quality of the decisions and it does not change the fact that a genuine edge loses on plenty of individual bets. Value is a statement about the price, not a prediction about tonight.
It does not make the estimate correct. Every step after the first depends on the probability being roughly right, and a wrong probability produces a disciplined, well logged, correctly priced bet on nothing.
It does not survive a change in your inputs. Small changes in assumptions move a marginal edge across zero, which is why the discard step matters more than the selection step. The bets you decline are the process working.
And it does not remove the operational constraints. Limits, the time between spotting a price and taking it, and the possibility that a venue responds to being beaten on price all sit between finding value and keeping it. A process that only models the maths and not the execution is modelling half the problem.
Turn the filter into a process
A scanner that reports a candidate, its price, its stake limit and the benchmark it beat gives you something to act on. A scanner that reports a percentage does not.
Questions people ask
What is the formula for finding a value bet?
A price is a value bet when your probability times the decimal price is greater than one. Believing 55% at a price of -105 gives 0.55 x 1.9524 - 1, which is +7.38% per unit. The same 55% belief at -130 gives 0.55 x 1.7692 - 1, which is -2.69%. Value belongs to the pair of a probability and a price, not to a side on its own.
Do value bets work?
They work in the sense that a bet placed at a price longer than the true probability is profitable on average, and that is a mathematical statement rather than a promise. Whether they work for a particular person depends entirely on whether their probability estimates are better than the market's, which is the hard part and the part that no method can supply.
How do I find value bets without building a model?
Borrow a number instead of making one. Take a two-way market from a venue whose prices are known to be sharp, divide each implied probability by their sum to remove the margin, and treat the result as your estimate of the true probability. Then look for a longer price on the same side of the same market somewhere else. This avoids the winner's curse of selecting on your own estimation error, though a sharp benchmark can still be wrong on any single game.
Why do my value bets keep losing?
Two common reasons. The first is that the measured edge was largely your own estimation error rather than a real disagreement with the market: with three points of error and a two-point filter, a bettor with no edge at all still measures an average edge of 3.80 points on the bets they place. The second is that the price was not available at the size taken, or had already moved. Both are addressed by shrinking the edge you act on and by recording the price you actually got.
How many games should I scan to find value?
Scanning more games returns more candidates and does not improve the ratio of real ones. If roughly one price in forty is genuinely soft and your filter has a five percent false positive rate, scanning 200 games yields about five real opportunities and ten false flags, so only a third of what you look at is real. Improving the inputs and lowering the false positive rate changes the ratio; scanning harder does not.
Sources
- How to Find Value Bets, betmetricslab.com
- How to Find Value Bets Without Chasing False Edges, sportnewsing.net
