The problem it solves
In many games you cannot see the thing you most want to know: the cards in an opponent’s hand, the type of player across the table, or whether a stranger on your team is cautious or reckless. What you can see is evidence, such as a bet, a pause, or a move. Bayes rule is the standard method for turning that evidence into a revised estimate.
It answers a question people often get backwards. You usually know how likely a behaviour is if a certain hidden fact is true. What you want is how likely the hidden fact is given the behaviour you saw. Those are different numbers, and mixing them up is one of the most common reasoning errors.
The three ingredients
Bayes rule combines three quantities:
- Prior: how likely the hidden fact was before you saw anything.
- Likelihood: how often the evidence appears when the fact is true, and how often it appears when the fact is false.
- Posterior: the updated probability after the evidence.
In symbols, P(A given B) = P(B given A) × P(A) ÷ P(B). The denominator P(B) is the total chance of seeing the evidence under every possibility, which you get by adding up each case. You do not need to memorise the formula to use it. Counting an imaginary crowd works just as well.
A worked example with natural frequencies
Suppose you believe that about 20 percent of the opponents you meet in a certain card game are aggressive types, and the rest are cautious. You also believe that aggressive players make a large raise on 60 percent of their turns in a given situation, while cautious players do so only 10 percent of the time. An opponent makes a large raise. How likely is it that they are aggressive?
Imagine 1,000 such opponents.
- 200 are aggressive. Of these, 60 percent raise: 120 players.
- 800 are cautious. Of these, 10 percent raise: 80 players.
In total, 120 + 80 = 200 players raise. Of those 200, 120 are aggressive. The posterior is 120 ÷ 200 = 0.6, or 60 percent.
Two things are worth noticing. First, the raise moved your belief from 20 percent to 60 percent, a large shift but not certainty. Second, 40 percent of the raisers were cautious players who simply raised once, because there are so many more of them. The prior mattered.
The inputs here are assumptions made for illustration. In a real game you would be estimating them from experience, and they might be quite uncertain. The method is only as good as those estimates.
The base rate trap
Ignoring the prior is called base rate neglect. Imagine a rare type of opponent who almost always makes a particular move. If that type is only 1 percent of the field and some common players also make the move occasionally, then seeing the move may still point to the common type. Evidence that is very typical of a rare thing does not automatically make that thing probable.
This links to wider thinking about cognitive biases, where intuitive judgement tends to focus on the vivid evidence and skip the background frequency.
Using Bayes rule with hidden cards
The same machinery works for cards. Say you are fairly sure from earlier play that an opponent holds either a strong hand or a weak one, and that the chance of a strong hand dealt to a player is, for the sake of illustration, 25 percent. If strong hands bet 80 percent of the time and weak hands bet 20 percent of the time, then in 100 deals you would see:
- 25 strong hands, 80 percent betting: 20 bets.
- 75 weak hands, 20 percent betting: 15 bets.
A bet appears 35 times, 20 of them from strong hands. The posterior probability of a strong hand is 20 ÷ 35, about 0.571, or roughly 57 percent.
Check the arithmetic: 25 × 0.8 = 20, and 75 × 0.2 = 15. Their sum is 35, and 20 ÷ 35 = 4/7 ≈ 0.571.
Notice how this connects to bluffing signals. If a player bluffs at the right frequency, the betting action tells you less, because strong and weak hands both produce it. Part of the point of mixed strategies is to keep the likelihoods close together so that your posterior stays near the prior.
Updating more than once
Bayes rule can be applied repeatedly. The posterior from one observation becomes the prior for the next. After a large raise leaves you at 60 percent, a second piece of evidence starts from 60 percent, not from the original 20. This is how opponent reading builds up over a session: each action nudges the estimate.
There is a trap in chaining updates, though. The observations should be reasonably independent given the hidden fact. If the same underlying habit produces two actions, counting both as separate evidence overstates what you know.
Practical limits
- Inputs are guesses. Likelihoods in real games come from limited observation. Treat the output as a rough guide, not a precise figure.
- Opponents adapt. A player who knows you are reading them may change their behaviour, which changes the likelihoods.
- Small samples mislead. A few observations can suggest a pattern that is not really there.
- Posteriors are beliefs, not outcomes. A 60 percent estimate is wrong 40 percent of the time even when you calculate it correctly.
Used carefully, Bayes rule is less about producing exact numbers and more about keeping you honest. It reminds you to ask three questions whenever you see evidence: how likely was this to begin with, how often does this evidence appear if I am right, and how often does it appear if I am wrong?