A simulation large enough to be meaningful always converges on the same answer, and it is the one printed in the game rules. What the arithmetic shows, why short runs look like patterns, and the one dataset worth keeping.
March 30, 2026 • Reading time: 6 minutes • By Isabelle Grant, writer on bankroll and casino mathematics
Articles claiming to have run ten thousand bets and uncovered exploitable patterns are a fixed feature of this industry. They are worth examining, not for the patterns, but because the arithmetic that makes them impossible is simple enough to do in a paragraph — and doing it explains more than any simulation result could.
To be clear about what follows: no simulation was run for this article. None is needed, and that is precisely the point.
The arithmetic that needs no simulation
Every game with a fixed house edge has a straightforward expected cost: stake per bet, multiplied by the edge, multiplied by the number of bets. A game returning 96% carries a 4% edge, so €1 staked ten thousand times has an expected cost of €400 — regardless of the order the wins and losses arrive in.
This is what a large simulation converges on. Not a pattern, not a profitable window, not a time of day: the number already printed in the game information panel. A simulation that produced anything else would be evidence that the simulation was wrong.
| Return to player | Edge | Expected cost of 10,000 × €1 | Expected cost per hour at 600 spins |
|---|---|---|---|
| 98% | 2% | €200 | €12 |
| 96% | 4% | €400 | €24 |
| 94% | 6% | €600 | €36 |
| 88% | 12% | €1,200 | €72 |
Two things follow from that last column, and they matter more than any strategy. Spin speed multiplies cost as directly as stake size does — the same €1 bet at 600 spins an hour costs twice what it costs at 300. And the difference between a 96% game and an 88% game is a factor of three, available for free before a single bet is placed. Those are the decisions that actually change what play costs.
Why short runs look like signals
Expected value describes the long run. What a player experiences is a short one, and short runs of random outcomes look nothing like the average.
The spread around the expected result grows with the square root of the number of bets, while the expected loss grows in direct proportion. Over a few hundred spins the spread dominates and results scatter widely on both sides of the average, which is exactly why a winning session is unremarkable rather than evidence of anything. Over tens of thousands, the proportional term takes over and the outcome settles close to the arithmetic.
This is also why streaks are not anomalies. In any long series of independent outcomes, runs of six, eight or ten identical results are expected to appear — their absence would be the surprising finding. A player watching one table sees the stretch that looks like a signal and not the thousands of spins that make it ordinary, which is the raw material for most of the errors in the subject.
What backtests get wrong
Betting systems are usually presented with a history attached, and the history is usually the weakest part of the claim.
- Overfitting. A rule with enough conditions will fit any past dataset perfectly and predict nothing. The tell is specificity that has no reason behind it — a threshold of 2.35 rather than 2, a window of seventeen days.
- Survivorship. The systems that get published are the ones that happened to work in the sample. The identical systems that failed are not written up, so the visible population is selected on the outcome being demonstrated.
- Costs left out. Backtests are usually run on prices that were not actually available — the best line across every book, at the moment of the result rather than the moment of the bet. Margin, limits and closing movement remove most published edges on their own.
- Progression disguised as strategy. A doubling progression produces a long, smooth series of small wins in any backtest, followed by one loss that erases them. Cut the sample before the loss and the system looks flawless.
None of this makes modelling worthless. It makes the standard of proof higher, and the honest standard is narrow enough to state in one line — did the bet beat the closing price, and did it do so often enough to clear the margin.
The one dataset worth keeping
There is a set of numbers a player can collect that is genuinely informative, and it is not about the games. It is about the account.
- Total deposited and total withdrawn, per month. Two figures, and together they are the only honest measure of what this costs.
- Time played per session, recorded externally. Estimates made afterwards are consistently short.
- The stake size at the start of a session and at the end of it. A divergence between them is the single most useful early signal available.
- Which sessions ended because a limit was reached, and which ended for some other reason.
That last one is the closest thing to a finding this article has. Sessions that end on a pre-set limit and sessions that end when the balance runs out are two different activities, and the ratio between them describes a month far better than the profit figure does. The point at which time stops being noticed at all is its own subject.
The honest summary
A simulation tells you the house edge, which the game already told you. A backtest tells you what would have worked in the past, which is not the same question. Session data tells you what you actually did, which is the only one of the three you can act on.
Everything else marketed as a data-driven edge is either the edge itself restated, or a pattern found in noise. And for anyone tempted to search for gaps in the operator's systems rather than in the mathematics — the operator holds every record of every account, and the asymmetry is not in the player's favour. The broader statistical background is summarised on online gambling, and analytics methodology as applied to sport is discussed at the MIT Sloan Sports Analytics Conference.
Published March 30, 2026 • BetBeast Blog