Bayes' Theorem for Traders: Updating Beliefs in Real Time

Every trading decision is a belief-updating problem. You have a prior (what you believed before seeing new information), you receive new information, and you update to a posterior (what you believe now).

This is Bayes' theorem. And prop firm interviewers use it to screen for the one skill that separates good quants from great ones: the ability to update beliefs correctly when new information arrives.

The Formula

P(A|B) = P(B|A) × P(A) / P(B)

In words: the probability of A given B equals the probability of B given A, times the prior probability of A, divided by the total probability of B.

The denominator expands as: $$P(B) = P(B|A) \cdot P(A) + P(B|\neg A) \cdot P(\neg A)$$

The pieces:

• P(A): your prior — what you believed before seeing B
• P(B|A): the likelihood — how probable is the evidence if A is true?
• P(A|B): the posterior — what you believe after seeing B

Why Traders Need This

In markets, you're always updating:

• A stock gaps up on earnings. How much should you update your view of the company's fundamentals?
• An economic number comes in hotter than expected. How much does that change your rate expectations?
• Your algorithm shows unusual performance. Is it genuinely alpha, or is it noise?

The natural human reaction is to update too much (overreact to recent information) or too little (anchor to prior beliefs). Bayesian thinking gives you the mathematically correct update.

The Base Rate Problem: Why Intuition Fails

Classic Bayes problem: A disease affects 1% of the population. A test for the disease is 95% accurate (true positive rate = 95%, false positive rate = 5%). You test positive. What's the probability you have the disease?

Most people say "95%." The correct answer is much lower.

Step by step:

P(disease) = 0.01 (prior) P(positive | disease) = 0.95 (true positive rate) P(positive | no disease) = 0.05 (false positive rate)

P(positive) = 0.95 × 0.01 + 0.05 × 0.99 = 0.0095 + 0.0495 = 0.059

P(disease | positive) = (0.95 × 0.01) / 0.059 = 0.0095 / 0.059 ≈ 16%

Even with a 95% accurate test, a positive result only means a 16% chance of actually having the disease. The base rate (1%) dominates.

Why this matters for trading: This is the exact math behind why most signals are false positives. If the true alpha rate in the universe of strategies is 1%, and your backtest has a 95% chance of finding it if it's real (power = 0.95) and a 5% false positive rate — most "significant" strategies you find are noise.

12 Bayes Problems from Prop Firm Interviews

Group 1: Medical and Base Rate Problems

1. Rare disease: 0.1% prevalence. Test is 99% sensitive (true positive rate) and 99% specific (true negative rate = 1% false positive). You test positive. P(disease)?

P(disease) = 0.001 P(+|disease) = 0.99 P(+|no disease) = 0.01

P(+) = 0.99×0.001 + 0.01×0.999 = 0.00099 + 0.00999 = 0.01098

P(disease|+) = 0.00099 / 0.01098 ≈ 9%

Even a 99% accurate test on a 0.1% prevalent disease: only 9% chance you're positive. Base rates dominate for rare conditions.

2. Two tests, independent results. Same disease (1% prevalence), same test (95% accurate). You test positive twice, independently. P(disease)?

After first positive: P(disease) ≈ 16% (from earlier).

Now use 16% as the new prior and apply the test again:

P(+ second | disease) = 0.95, P(+ second | no disease) = 0.05

P(disease | 2nd positive) = (0.95 × 0.16) / (0.95×0.16 + 0.05×0.84) = 0.152 / (0.152 + 0.042) = 0.152 / 0.194 ≈ 78%

Two independent positive tests brings you to 78% confidence. Sequential Bayesian updating.

Group 2: Card and Ball Problems

3. A deck has 2 red cards and 1 black card. You draw without looking. I peek at your card and tell you truthfully "it is not the black card." P(red)?

Without information: P(red) = 2/3. "Not black" is consistent with you having a red card (certain) or... wait, you have either a red or black card. If "not black," you definitely have red. P(red | told not black) = 1 (100%).

This seems trivial, but note that the informative content of the message depends on its probability. "Not black" was almost certain (2/3 base rate) before the message, so it updates you a lot less than "not red" would.

4. Three cards: one gold/gold, one silver/silver, one gold/silver. You pick a card and see gold on one face. P(other face is gold)?

Let G/G = gold-gold card, S/S = silver-silver, G/S = mixed.

You see gold. The ways to see gold: you could be looking at: face 1 of G/G (prob 1/3), face 2 of G/G (prob 1/3), or the gold face of G/S (prob 1/3).

Given you see gold, P(G/G card) = (1/3 + 1/3) / (1/3 + 1/3 + 1/3) = 2/3.

If you have G/G, the other face is gold. P(other face gold | see gold) = 2/3.

Most people say 1/2. This is the same logic as the Monty Hall problem.

5. Bag A has 3 red, 7 blue balls. Bag B has 6 red, 4 blue. You pick a bag at random and draw a red ball. P(Bag A)?

P(A) = P(B) = 1/2 (equal prior) P(red|A) = 3/10, P(red|B) = 6/10

P(red) = (3/10)(1/2) + (6/10)(1/2) = 3/20 + 6/20 = 9/20

P(A|red) = (3/10)(1/2) / (9/20) = (3/20) / (9/20) = 1/3

Drawing a red ball makes Bag B more likely (it has more red balls).

Group 3: Conditional Independence

6. Alice tells you a stock will go up with 70% confidence. Bob independently tells you the same with 60% confidence. If you believe they're independent signals and your prior is 50% up, what's the posterior?

Treat each statement as evidence. P(Alice says up | stock goes up) = 0.7, P(Alice says up | stock goes down) = 0.3.

After Alice: P(up | Alice says up) = (0.7 × 0.5) / (0.7×0.5 + 0.3×0.5) = 0.35/0.50 = 0.7

After Bob (using 0.7 as new prior): P(up | Bob says up) = (0.6 × 0.7) / (0.6×0.7 + 0.4×0.3) = 0.42/(0.42 + 0.12) = 0.42/0.54 ≈ 77.8%

This assumes independence, which in practice requires the signals to be genuinely uncorrelated. If both analysts read the same report, they're NOT independent.

7. You have two fair coins and one two-headed coin in a bag. You pick a coin at random and flip it — heads. P(two-headed coin)?

P(2H) = 1/3 (prior) P(H | 2H) = 1, P(H | fair) = 1/2

P(H) = 1×(1/3) + (1/2)×(2/3) = 1/3 + 1/3 = 2/3

P(2H | H) = 1×(1/3) / (2/3) = 1/2

Now you flip again — heads again. P(two-headed)?

New prior: 1/2. P(H|2H) = 1, P(H|fair) = 1/2.

P(H) = 1×(1/2) + (1/2)×(1/2) = 1/2 + 1/4 = 3/4

P(2H | 2nd H) = (1 × 1/2) / (3/4) = 2/3

Each additional head makes two-headed more likely.

Group 4: Market Applications

8. You're a trader with a prior of 40% that a stock will beat earnings. An analyst you trust (who's right 70% of the time when the stock beats and 60% right when it misses) says "it will beat." Update your estimate.

P(beat) = 0.4, P(miss) = 0.6 P(says beat | beat) = 0.7, P(says beat | miss) = 1 - 0.6 = 0.4

P(says beat) = 0.7×0.4 + 0.4×0.6 = 0.28 + 0.24 = 0.52

P(beat | analyst says beat) = (0.7 × 0.4) / 0.52 = 0.28/0.52 ≈ 53.8%

The analyst moved you from 40% to 54%. A good analyst with a moderate track record provides modest but real information.

9. A stock historically gaps up 60% of days after a specific pattern. You see the pattern on Monday. By Thursday, 3 of 4 days have been positive. What's P(pattern works)?

This requires updating on data. Let's interpret: prior P(pattern works) = 60%, and when it works, each day is up with P = 0.7; when it doesn't, each day is up with P = 0.5.

P(3/4 up | works) = C(4,3) × 0.7³ × 0.3 = 4 × 0.343 × 0.3 = 0.4116 P(3/4 up | doesn't) = C(4,3) × 0.5³ × 0.5 = 4 × 0.125 × 0.5 = 0.25

P(works | 3/4 up) = (0.4116 × 0.6) / (0.4116×0.6 + 0.25×0.4) = 0.247 / (0.247 + 0.1) = 71%

Observing 3/4 up days moves the belief from 60% to 71%.

10. A trading signal fires with P = 5% on any given day. When it fires on genuine alpha days (which are 10% of trading days), it's 90% likely to fire. When it fires on non-alpha days, it fires randomly (P = 5%). Given the signal fires today, P(alpha day)?

P(alpha) = 0.10 P(signal | alpha) = 0.90 P(signal | not alpha) = 0.05

P(signal) = 0.90×0.10 + 0.05×0.90 = 0.09 + 0.045 = 0.135

P(alpha | signal) = 0.09 / 0.135 = 66.7%

A strong signal on a rare event: if the signal is specific to the event, P(alpha | signal) >> P(alpha).

Group 5: Sequential Updating

11. You're sequentially flipping a coin of unknown bias p. After 10 flips: 7 heads, 3 tails. Using a uniform prior over p ∈ [0,1], what's the posterior probability that p > 0.5?

With a uniform prior and binomial likelihood, the posterior for p is Beta(8, 4) (Beta(h+1, t+1) where h=7, t=3).

P(p > 0.5) = ∫₀.₅¹ Beta(8,4) dp

This is the regularized incomplete beta function I(0.5; 8, 4). Numerically ≈ 89.5%.

Intuitively: 7 heads out of 10 is substantial evidence the coin is biased toward heads. The posterior is concentrated above 0.5.

12. You're running a strategy. After 100 trades, your win rate is 55% (prior: 50/50 fair). Assuming independent trades and no alpha: P(win rate ≥ 55% by chance)?

Under H₀ (no edge, p=0.5): X ~ Binomial(100, 0.5). P(X ≥ 55) = ?

Using normal approximation: Z = (55 - 50) / √(100×0.5×0.5) = 5/5 = 1.0.

P(Z ≥ 1.0) ≈ 15.9%

A 55% win rate over 100 trades is NOT statistically significant (p ≈ 16%). You'd need about 400 trades at 55% to reach p < 5%. This is why professional funds require long track records before concluding alpha is real.

The Bayesian Trader Mindset

Bayesian reasoning implies certain habits that separate rigorous traders from intuitive ones:

1. State your prior explicitly. Before looking at new data, write down your prior belief and why you hold it. "I think this stock will beat earnings with 45% probability based on sector trends and recent guidance."

2. Distinguish P(B|A) from P(A|B). "If I'm right about the macro, rates will stay elevated" is P(macro right | rates stay up). But you want P(rates stay up | macro view), which requires Bayes to compute. They're not the same.

3. Base rates dominate rare events. For anything rare (a stock being a multi-bagger, a strategy having genuine alpha), the base rate pulls the posterior toward skepticism. Evidence has to be very strong to overcome a low prior.

4. Sequential updating beats one-shot analysis. Instead of waiting for a definitive answer, update continuously as information arrives. This is more computationally intensive but produces better-calibrated beliefs.

5. Independence is an assumption, not a fact. When combining signals (two analysts, two indicators, two test results), they must be independent for simple multiplication to work. Market signals rarely are. Correlated evidence updates less than independent evidence.

Bayes in Quantitative Research

For Two Sigma and HRT-style quantitative research roles, Bayesian thinking appears in:

Signal combination: If you have two signals that each predict price movement, combining them requires accounting for their correlation. Bayesian frameworks do this explicitly.

Model selection: Should you use Model A or Model B? Bayesian model comparison gives the posterior odds ratio: how much more likely is the data under A than under B?

Parameter estimation: Instead of fitting a point estimate for a model parameter, Bayesian estimation gives you a distribution over parameters. This naturally captures uncertainty.

Overfitting prevention: Bayesian models with appropriate priors automatically regularize — pulling extreme estimates toward the prior when data is limited.

---

Daily estimation practice at Fermiq develops the base-rate intuition that Bayesian reasoning requires. Practice the probability category specifically — many questions are disguised Bayes problems.