Oddsprey

Forgot your password?

Don't have an account yet? Create one

Guide

How Oddsprey builds its predictions, what value means, and how we size stakes. Read this once and the rest of the app makes sense.

What Oddsprey does

Odds of 2.20 are not good or bad. They are just a price. The only question worth asking is whether that price is too high.

At 2.20 the bookmaker is saying this happens about 45 times in 100. If our model says 52, the price is too generous and we back it. If our model says 40, that same 2.20 is a bad bet and we leave it. Same odds, opposite call.

That is the whole product. We rate every fixture with a statistical model, put our number next to the bookmaker's, and only bet when ours is higher. We are not trying to pick winners. We are hunting mispriced bets.

The bookmaker's odds become an implied chance, our model estimates its own chance, and the difference between the two is the edge book odds 2.20 the price offered implied chance 45% what 2.20 means our estimate 52% from the model edge +7 points, so we bet
Turn the price into a percentage, put our own next to it, and bet only when ours is higher. A negative number here is a bet we skip.

Each match we cover comes with:

Reading odds

Odds are the bookmaker's price on an outcome. They also tell you the payout.

Decimal odds

The standard across Europe. Odds of 2.50 return €2.50 for every €1 staked, your stake included. Return = stake × odds.

A one euro stake at odds of 2.50 returns 2.50: the original stake plus 1.50 profit your stake €1.00 profit €1.50 returns €2.50
Odds of 2.50 on €1. The return always includes your stake back, so the profit is €1.50.

Implied probability

Flip the odds to get a probability. 2.50 works out to 40% (1 ÷ 2.50). That figure carries the bookmaker's margin, so a book's implied chances always add up to more than 100%.

The three implied probabilities in a match add up to 105 percent; the extra five points are the bookmaker's margin 100% home 2.22 45% draw 3.70 27% away 3.03 33% margin
Three prices on one match, flipped to probabilities. They total 105%. The shaded 5% past the line is the bookmaker's margin, which is why implied odds are never a fair estimate of the real chance.

Edge and value

Edge is the gap between our probability and the odds-implied probability. When our number is higher, the bet has value.

Say a team has a 50% chance in our model, but the odds are 2.10 (47.6% implied). That is roughly a 2.4% edge. Back enough spots like that and the maths works in your favour over time.

Our 50 percent probability against the 47.6 percent implied by odds of 2.10; the 2.4 point gap is the edge our number 50% book implied 47.6% edge 2.4%
That thin band is the entire edge. It looks like nothing, and it should. Real edges are small. What turns them into profit is repetition, across hundreds of bets.

We flag a bet when three things line up:

How we size stakes

Every stake comes from the Kelly criterion. Kelly ties the bet to the edge: more edge, more stake, but never so much that a bad run wipes you out.

The formula

Kelly % = (p × odds − 1) / (odds − 1). Here p is the model's chance for the selection and the odds are decimal. The top half is the edge, what a euro staked returns on average, and dividing by (odds − 1) turns that into a share of the bankroll.

A 50% call priced at 2.10 gives (0.50 × 2.10 − 1) / 1.10, so 4.5% of the bankroll at full Kelly. The same call at 2.00 gives zero. No edge, no stake.

How we apply it

We stake a fraction of full Kelly, usually 10% to 50%, scaled by how much we trust the number and how big the edge is. On a €1,000 bankroll, a 2% Kelly call is a €20 bet. Fractional Kelly grows the account on winners and cushions the losing weeks.

Stake rises with edge. Full Kelly is the steep upper line; we stake in the shaded band between 10 and 50 percent of it full Kelly we stake in here 0 4% 8% our edge on the bet 0 5% 10% stake, % of bankroll
More edge earns a bigger stake, but never the full Kelly amount. Full Kelly maximises long-run growth and is brutal on the way there, so we deliberately stake a fraction of it.

How the numbers are built

We are not going to publish the method in full. How the inputs are weighted, and where we let the model disagree with the market, took the longest to get right and it stays in-house. Here is the shape of it.

We start from data and finish with a review.

The core model is Dixon-Coles, a standard method for football scorelines. We calibrate it against past results, then our analyst writes up the standout matches and flags the risk and context the model alone would miss.

What the confidence label means

Every match carries High, Medium or Low confidence. It is not how likely the bet is to win. That number is the model chance, printed next to it. Confidence answers a different question: how much history the model had to work with on this fixture, and whether it could use its strongest method.

Two things decide it, and the rule is fixed rather than a judgement call:

LabelWhat it means
HighPriced by Dixon-Coles, with at least five recent matches for both sides.
MediumEnough recent form to price the match, but either on a weaker method or with a shorter run of matches behind it.
LowFewer than three recent matches for one of the sides. The prediction still stands, but it rests on a thin record.

So a Low confidence match is not a bad bet and a High confidence one is not a safe bet. The label tells you how much the number behind it is leaning on, which matters most when the two disagree: a big edge on a Low confidence fixture is the one to treat carefully.

Then comes the part we care most about. Every call is scored against what actually happened, and so is every bet we passed on. Losses tell us where the model is too confident. Missed winners tell us where it is too cautious. Both get fed back, and the weights move. The goal is a system that keeps sharpening itself, so the version pricing next month's matches is better than the one pricing this month's.

Four data inputs feed the Dixon-Coles model, which is calibrated on history and reviewed by an analyst; every result feeds back into the model match history & xG team form market odds injuries & context Dixon-Coles model Calibrate on history Analyst review results retrain the model, automatically
The loop is the product. Results feed straight back into the weights, so the model that prices next month's matches has been trained by this month's, with no one deciding what to learn from.

What a subscription gets you

Check our homework before you pay a cent. Every settled result stays free forever, wins and losses alike, with nothing quietly deleted. Most tipsters show you their best week. We show you all of them.

A subscription opens what is coming next:

Everything already settled is free forever; a subscription opens the matches still to be played today Settled results free forever Upcoming matches subscription
The split runs on one line: the past is open so you can audit us, the future is the paid product.

And the results keep working after the whistle. Every graded match goes back into the model, so the wins compound and the losses get paid forward as a sharper number next week. You are not buying a list of tips. You are backing a system that gets better the longer it runs.

New accounts get 30 days with everything open. No card up front.

Common questions

How accurate are the predictions?
See for yourself on the Results page. We track win rate, ROI, and profit or loss for every bet we have ever posted. It varies by market, league, and stretch of time. Over a long enough run, a real edge shows up in the numbers.
What do I get for free?
Every settled result, one sample match in full, and the model's chances for all fixtures. New accounts also get 30 days with every match open. After that you can subscribe or stay on the free tier.
Should I follow every bet?
No. Each one stands on its own. Only bet when you agree with the read and the price you can get still holds the edge. We give you the analysis; the call is yours.
How do I use the suggested stakes?
They are already sized with fractional Kelly for a €1,000 bankroll, so scale them to yours. Want less risk? Take a smaller fraction. Never stake more than the suggested amount.
Can I lose money?
Yes. Betting carries risk, always. Even a real edge loses over short stretches, because variance is real. Our numbers improve the long-run return; they do not guarantee any single bet. Only stake what you can afford to lose.

Getting started

Create a free account to open every match for 30 days. Nothing to pay up front.

Want the plans first? They are on the pricing page.