About Flintstat

Flintstat is a sports statistics and research project. We build simple mathematical models from historical results and public performance data, work out a win probability for each game, and compare that estimate against the odds bookmakers publish. The whole point is one question: can plain math on public data stay as close to reality as the market's price, or not.

This is analysis and research. It is not a tipping service, not a prediction app, and not betting advice. Nothing here is a promise that a game will go a certain way.

What this is, and what it is not

A model outputs a probability, never a certainty. If it says a team has a 60 percent chance to win, it is also saying it expects that team to lose about four times in ten. Both are the same sentence.

We use no insider information. There are no contacts inside clubs, no fixed-match rumours, no locks, and no way to see the future. Everything comes from math applied to records anyone can look up.

What actually moves a game

Outcomes are driven by a small number of measurable factors, and they differ by sport. These are the inputs that carry real predictive weight, not folklore like momentum or must-win narratives.

Soccer

• Home advantage, which is real and measurable but has shrunk over time and varies by league.
• Expected goals (xG), which score the quality of chances created rather than the noisy final score.
• Rest and fixture congestion, since short turnarounds and midweek games degrade performance.
• Injuries and suspensions to key players, especially a first-choice striker or goalkeeper.
• Red cards, which swing a match sharply by changing the goal expectation for a long stretch.

Tennis

• Court surface, because the same player can be elite on clay and ordinary on grass.
• Serve and return statistics, the most predictive category (points won on serve and return, break points).
• Recent form and short-term fatigue from long matches and quick turnarounds.
• Head-to-head history, read carefully and ideally surface by surface.

Basketball

• Offensive and defensive rating (points per 100 possessions); their gap, net rating, is the cleanest quality measure.
• Pace as context, so efficiency per possession is the honest comparison, not per-game totals.
• Rest and scheduling, since back-to-back games measurably lower performance.
• Injuries and load management, where one star sitting shifts a team's rating materially.

Baseball

• The starting pitcher matchup, the single largest input, read through ERA, WHIP, and strikeout and walk rates.
• Park factors, since each stadium inflates or suppresses runs and home runs.
• Bullpen strength and recent workload.
• Platoon splits and the lineup as actually posted.

Turning history into a probability

The job of a model is to produce one number: a probability. Historical data goes in, a probability of each outcome comes out. A few standard tools do most of the work.

Rating systems (Elo). Every team or player carries a rating, and the gap between two ratings gives a win probability through a fixed formula:

expected score = 1 / (1 + 10^((R_opponent − R_team) / 400))

A 200-point edge is about a 76 percent chance, a 400-point edge about 91 percent, and equal ratings are 50/50. After each game, ratings move toward reality by K × (actual − expected), so beating a strong opponent lifts a rating more than beating a weak one.

Goal models (Poisson). Goals roughly follow a Poisson distribution, so from each team's expected goals you can compute the probability of every scoreline and add them into home-win, draw, and away-win probabilities. The Dixon-Coles adjustment corrects a known flaw, that plain Poisson slightly underrates low-scoring draws.

Regression models. For a yes or no outcome, logistic regression takes many inputs (ratings, rest, home flag, injuries) and returns a single probability between 0 and 1.

None of this predicts a single result. Even good soccer models are right only about half the time on the three-way market, because football is genuinely hard. The value is in calibrated probabilities over many games, not in calling one match.

From probability to fair odds, and what an edge means

Fair decimal odds are simply the inverse of a probability:

fair odds = 1 / probability

A 50 percent chance is fair odds of 2.00, a 25 percent chance is 4.00. Any price can be read back the same way: implied probability = 1 / decimal odds.

There is a catch. Add up the implied probabilities of every outcome in a bookmaker's market and the total is more than 100 percent. That extra is the built-in margin, the vig. To compare fairly you strip it out, by dividing each implied probability by the total so they sum to 100 percent. That gives the market's no-vig probability. The edge is then just:

edge = model probability − no-vig market probability

A positive edge means the model thinks an outcome is more likely than the fairly measured market does. Removing the vig does not predict a winner. It only makes the model-versus-market comparison honest.

Who is more accurate

The market's closing line, the final odds just before a game starts, is widely treated as the sharpest estimate available, because by then it has absorbed lineups, injuries, and money from informed bettors. So the honest test of a model is not whether individual picks won. It is whether the model's estimates systematically track or beat the no-vig closing line over a large sample.

That is why a running profit and loss is a poor short-term scorecard. Single results are binary and noisy, so a good model can look bad and a bad one can look good over a handful of games. Calibration against the closing line shows a real edge far faster than win and loss ever will.

The limits, stated plainly

• Models estimate probabilities, not outcomes. Variance dominates in the short run and no model removes luck.
• Garbage in, garbage out. Public data has gaps: late injury news, weather, and motivation a historical model may miss.
• Markets are efficient and hard to beat. Most of the time a fair-priced market is already close to correct.
• Overfitting is a constant risk. A model tuned to look great on past games can fail on new ones.
• Nothing here is financial advice or a recommendation to bet. It is a study of numbers. You are 18 or older and responsible for your own decisions.

Contact

Flintstat is a small, independent project. If you spot an error in the numbers or the method, or you have a question about the data, tell us. Getting it right matters more than looking right.