Two kinds of question
Probability answers two different kinds of question, and they behave very differently.
The first asks about a single event: will this attack crit, will this match be won, will this chest contain the rare item? For these, probability gives odds but no certainty. Even a 95 percent chance fails one time in twenty.
The second asks about many events together: what share of attacks will crit across a season, what fraction of matches will a team win across a hundred games? For these, probability gives something much closer to certainty. This is the core idea behind the shift from chance to certainty: the more trials you combine, the more predictable the overall result becomes, while each individual trial remains as uncertain as ever.
The law of large numbers, in plain terms
The law of large numbers says that as the number of independent trials grows, the observed average tends to move toward the true underlying average. Flip a fair coin and the share of heads will wander early on, then settle near one half.
Two common misunderstandings are worth clearing up.
First, the law does not say the counts even out. After 1,000 flips, you might be ten heads above 500. After a million flips, you might be a few hundred above half. The raw gap can actually grow, but it grows more slowly than the number of flips, so the proportion gets closer to one half.
Second, the law does not say that the coin compensates for earlier results. Each flip is independent, and nothing is “due”. The proportion settles because new results are added to a growing pile, not because later results correct earlier ones.
How fast does it settle?
For a yes-or-no event with probability p, the typical spread of the observed proportion after n trials is the square root of p × (1 − p) divided by the square root of n. For a fair coin, p = 0.5, so the numerator is 0.5.
| Number of flips | Typical spread in proportion of heads | Typical spread in number of heads |
|---|---|---|
| 10 | about 0.158 (15.8 points) | about 1.6 heads |
| 100 | 0.05 (5 points) | 5 heads |
| 1,000 | about 0.016 (1.6 points) | about 16 heads |
These figures are standard deviations, so results often fall a bit outside them without anything being wrong. Even so, the pattern is clear. Multiplying the number of trials by 100 only shrinks the spread in the proportion by a factor of 10. Certainty arrives slowly, and each additional digit of precision is expensive.
The practical rule is that to halve the noise in an estimate you need roughly four times as many trials. This is why a short record tells you very little, and why the question of sample size comes up whenever someone claims a change made a difference.
A game example: win rates
Suppose a player’s true chance of winning any game is 55 percent. After 20 games, the standard deviation of their win count is the square root of 20 × 0.55 × 0.45, which is about 2.2 wins. The expected count is 11, so a record anywhere from roughly 9 to 13 wins would be unremarkable. That is a range from 45 to 65 percent, covering players from clearly below average to clearly strong.
After 400 games, the standard deviation of the win rate is about the square root of 0.55 × 0.45 divided by 20, which is about 0.025, or 2.5 points. A 55 percent player will usually show a win rate in the range of roughly 50 to 60 percent. The signal has started to separate from the noise, though only moderately.
This is also why a skill rating needs many games before it settles, and why rating volatility is large when a rating is based on few results.
What stays uncertain
Large numbers make averages stable, but they do not make individual outcomes predictable. Knowing that a 20 percent crit will land on about one in five attacks in the long run tells you nothing certain about the next attack. Several practical limits follow:
- The next result is unaffected. Past results do not shift the odds for independent events.
- Average and spread are separate. A long-run average can be very stable while single results remain highly variable. The earlier example of a 20 percent crit with double damage has an average of 120 per attack and a standard deviation of 40, no matter how many attacks occur. What shrinks with more attacks is the uncertainty of the average, not the variability of each attack.
- Rare events can still happen. With millions of players making attempts, even one-in-a-million outcomes occur regularly. Aggregate certainty and individual surprise exist together.
- Independence matters. The law applies to independent trials. If results are linked, because a hot streak changes confidence or because a game adjusts odds after failures, the averaging works differently.
Why designers and analysts rely on it
Game designers use this behaviour to set target averages and let variance create drama. Published drop rates work this way: a drop that appears at a stated rate over millions of attempts will land very near that rate, even though individual players see huge differences. The concept of expected value is useful here, because it summarises what a repeated event delivers on average.
Analysts use the same logic in reverse. When they see an observed rate that differs from the intended one by more than the noise predicts for that number of trials, it suggests the real odds differ from the stated ones, or that something else is influencing the result. Judging how large a gap counts as “more than noise” is the hard part, and it depends on knowing the square-root rule above.
Keeping the two views apart
A useful habit is to ask which question you are answering. If the question is about one event, treat the result as uncertain whatever the odds say. If it is about many events, use the average, and ask how many trials stand behind it. Being clear about which question applies prevents two common errors: expecting certainty from a single roll, and expecting wild swings from a very large sample.