crtdrops
NOTE · ENGLISH ·

Luck makes winners too

On the left, a thousand simulated players with no edge after 100,000 hands, with seed 1: bars for how many landed in each 1 bb/100 band, from −15 to 15, brighter from zero up, and a vertical line at the 5 bb/100 of the player with a real edge; a red dot marks the best, at 12.5. On the right: end up ahead, 516 of 1,000; “prove” an edge, 54; the best, 12.5 bb/100; hands to prove 5 bb/100, 247,303.

I have played a fair amount of poker: at home with friends and in an online room with play chips, and once with 10 real dollars someone gave me. I always tried to use the probabilities to decide when to bet, when to fold and when to keep going. But I never kept a record of my results, and even if I had, I would not know what to conclude from it. How many hands does it take before winning says something about how you play?

Everything that follows is a simulation of a model, not the results of any player, mine included. The model is a Texas Hold’em cash table where each hand is independent and its result follows a normal distribution. The figures in the examples are illustrative: the 5 bb/100 edge, the 100 bb/100 variability and the 30 hands an hour. At a real table big pots are more common than in a normal distribution and the variability has to be estimated, so the normal can fall short. This is the math of the game, not an invitation to gamble, and it stays out of the legal debate over whether poker is a game of skill or of chance: it only asks how many observations it takes to tell a small edge from luck.

The unit: big blinds per 100 hands

At a cash table, what you win or lose is counted in big blinds, and the rate, per 100 hands: bb/100. The example in this note is a player who wins 5 bb/100 on average, with a variability of 100 bb/100. That is, per hand they win 0.05 blinds on average, but the result of a single hand swings about 10 blinds up or down. In a single hand, the edge is two hundred times smaller than the noise.

A thousand players with no edge at all

I simulated a thousand players with zero edge, with seed 1. None plays better than any other. After 1,000 hands, 526 were winning. After 10,000, 498. After 100,000, 516.

Half the players with no edge end up winning, and that does not change with more hands. What changes is how much they win: at 1,000 hands the best was at 115.2 bb/100, at 10,000 at 30.4, at 100,000 at 12.5. With more hands the results bunch up around zero, but half stay above it.

Winning is not proving

Ending up ahead proves nothing. To call it an edge, the result has to be so high that luck would rarely produce it. At 95 % confidence, after 100,000 hands that threshold is 5.20 bb/100: a player with no edge gets past it only 5 times in 100.

And that is what happened: 54 of the thousand players with no edge passed the threshold at 100,000 hands, 56 at 1,000 and 48 at 10,000. Over 200 seeds, the median was 49 to 51 at all three cuts. The test does what it promises: it raises a false alarm for 5 in 100 players with no edge. But with a thousand players, 5 in 100 is fifty “proven edges” that do not exist.

The player who does have an edge

Now the 5 bb/100, with the same variability. After 1,000 hands, the probability of being behind is 43.7 %. After 10,000, still 30.9 %. After 100,000, 5.7 %. A player with a real edge is behind after a thousand hands almost half the time, and at 10,000 is still behind almost one time in three.

Proving it is worse: with 100,000 hands, the 95 % test finds the edge only 47.5 % of the time. In the simulation of a thousand players at 5 bb/100, 488 proved it. The 95 % margin around the observed rate is ± 6.20 bb/100 after 100,000 hands: more than the edge itself.

How many hands it takes

For the test to find the edge 80 times in 100, at 95 % confidence, it takes 247,303 hands. At about 30 hands an hour at a live table, also an illustrative figure, that is more than 8,000 hours. At 99 % confidence and 90 % power, 520,678 hands.

The count grows with the square: twice the edge, 10 bb/100, needs a quarter of the hands, 61,826. An edge of 2 bb/100 needs 1,545,640.

Hands No edge: ahead No edge: “prove” an edge 5 bb/100: behind 5 bb/100: prove it
1,000 526 of 1,000 56 43.7 % 6.9 %
10,000 498 of 1,000 48 30.9 % 12.6 %
100,000 516 of 1,000 54 5.7 % 47.5 %

The two no-edge columns come from the simulation with seed 1; the ones for the player with an edge are the model’s exact figures.

Looking only at the winners

If after 100,000 hands you look only at the best of the thousand players with no edge, with seed 1 they are at 12.5 bb/100. That seed came out high: over 200 seeds, the best of the thousand landed between 8.3 and 13.5 bb/100, with a median of 10.0. The typical case is twice the edge of the player who really has one, and in all 200 it stayed above their 5 bb/100.

It is the problem of gathering many and keeping the one who came out best. Watch enough players and you will find extraordinary results even when none has an edge, and the most extraordinary one will not be the best player, just the luckiest.

The same with an automated strategy

The same question comes up when evaluating an automated investment strategy. I have one, TradingAgent, in a 91-day paper test that ends on January 11, 2027. The criteria for the decision were set before it started and are not changed midway or based on the results so far; that is why I give no partial results here.

The reason is the one in this note. A strategy’s positive result can be a favorable run of luck, just like a player with no edge who is ahead after 100,000 hands. If many variants are tried and the one that did best is picked, the thousand-player problem repeats: the best may just be the luckiest.

And looking several times is misleading too. If the same thousand players with no edge are checked at 1,000, 10,000 and 100,000 hands, 134 pass the threshold in at least one of the three checks, not 50. That is the median over 200 seeds, which ran from 98 to 163. Each extra look is another chance for luck to pass for an edge; that is why the criteria are set beforehand and not changed midway.

To try it with your own edge and variability, there is the skill or luck calculator: it runs in the browser, tells you how many hands it takes to prove an edge, and simulates a thousand players without it and a thousand with it, from a seed you can repeat.

In poker, winning does not prove we play better than our opponents, and losing does not prove our decisions were bad. The difference between skill and luck is not in one hand or one night: it is in an edge that holds up and can be proven over enough observations. The same goes for an automated strategy: a positive result can be promising, but it is not, on its own, evidence of a real edge. Luck makes winners too; statistics helps find out which ones have an edge worth taking seriously.

Israel Negrete Lepe is an electronics engineer. Twenty-five years making different systems work together under real conditions: from the sensor, the vehicle or the transaction to the backend, the data, security and operations. Today, increasingly, edge AI. I take on projects by the hour or by deliverable: israel0negrete@gmail.com.