Bayesian Thinking
Trading Intelligence
9 min read
Update your beliefs as new evidence arrives using Bayes theorem -- a framework for evolving your system without abandoning it.
Loading
9 min read
Update your beliefs as new evidence arrives using Bayes theorem -- a framework for evolving your system without abandoning it.
Markets change. Strategies weaken. But abandoning your edge too early can be just as dangerous as holding it too long. Here's how to think like a probabilistic strategist.
Prerequisites: Variance & Standard Deviation — you need to know what "within expected variance" means before you can update on it.
Bayes theorem — P(H|E) = P(E|H) · P(H) / P(E) — is the math of changing your mind correctly. In trading, H is "my system still has edge", E is "the last 30 trades." This lesson shows how to compute the update instead of guessing it.
Most traders operate in extremes:
But trading isn’t binary. It’s probabilistic.
Bayesian thinking gives you a framework to update your belief in a system over time — without emotional overreaction.
With it, you can:
Bayesian vs frequentist in one line: frequentist asks "how often would I see this evidence if the null were true?" Bayesian asks "given this evidence, how should I revise my belief?" Frequentist tests reject or fail-to-reject. Bayesian updates.
P(H|E) = P(E|H) * P(H) / P(E)
H = "my system still has edge" P(H) = prior — what you believed before this evidence P(E|H) = likelihood — probability of evidence if system still works P(E) = normalizer across all hypotheses P(H|E) = posterior — your updated belief
Without the likelihood term, you are not updating Bayesianly; you are guessing slowly.
"Start with a belief. As new evidence arrives, update that belief accordingly."
In trading, this means:
You don’t jump ship after 5 losses — and you don’t scale recklessly after 3 wins.
You adjust with probability, not emotion.
What Bayesian thinking does not do: it does not make your data better, it does not detect regime change you do not feed it, and it cannot rescue a wrong prior. It makes your uncertainty explicit and auditable. That is the entire value.
Foundational reading: Kahneman, Thinking, Fast and Slow (Ch. 16, base rates); Nate Silver, The Signal and the Noise (Ch. 8, Bayesian forecasting); López de Prado, Advances in Financial Machine Learning (Ch. 13, on sample-size adequacy before updating).
| Situation | Non-Bayesian Reaction | Bayesian Response |
|---|---|---|
| 5-trade losing streak | "System’s broken, quit." | "This is within expected variance. Reassess at 30–50 trades." |
| Sudden 4R winner | "Edge is insane, scale up fast!" | "Good outlier. Let’s see how it affects long-term EV." |
| Market volatility drops | "This strategy sucks now." | "System may underperform in this regime. Monitor and adapt slowly." |
| New filter seems great | "Add it now!" | "Test it separately over a sample. Compare Bayesian posterior EV." |
Pitfall — the prosecutor's fallacy: "I had 5 losses, my system is broken" confuses P(5L | edge) with P(edge | 5L). With a 55% win-rate system, P(5L | edge) = 0.018 — rare in isolation, but routine over 200 trades. Without the base rate of edge decay, the streak alone tells you almost nothing.
| Question asked | Output | Treats parameters as | Needs prior? | Best for traders when |
|---|---|---|---|---|
| Bayesian: "Given this evidence, how should I update my belief?" | Posterior probability distribution over edge | Random variables with distributions | Yes | You have a meaningful prior (backtest, history) and want a continuously-updated conviction |
| Frequentist: "How often would I see this if there were no edge?" | p-value, confidence interval | Fixed but unknown constants | No | You want a yes/no decision against a null hypothesis with a clean stopping rule |
“My strategy has an EV of +0.5R based on 100 trades.”
Last 30 trades show +0.1R, with higher drawdown and slippage.
With prior P(edge) = 0.70 and observing 30 trades at +0.1R when expected was +0.5R (likelihood ratio under "edge intact" vs "edge gone" ≈ 0.4), the posterior P(edge | evidence) ≈ 0.48. Edge is now a coin flip, not a conviction — that is what justifies a 30% size reduction, not gut feel.
It’s not about guessing the future. It’s about adjusting probability models with each trade.
| Quantity | Value | Notes |
|---|---|---|
| Prior P(true 55% WR) | 0.70 | From 100 backtested trades |
| Prior P(true 45% WR) | 0.30 | The "edge has decayed" hypothesis |
| Observed | 14W / 16L over last 30 | Live trades |
| Likelihood under 55% | ≈ 0.06 | Binomial(30, 0.55) at k=14 |
| Likelihood under 45% | ≈ 0.14 | Binomial(30, 0.45) at k=14 |
| Posterior P(true 55% | 14/30) | ≈ 0.50 | (0.06·0.7) / (0.06·0.7 + 0.14·0.3) |
| Action | Halve size, keep collecting | Edge is uncertain, not dead |
This is what "updating" actually looks like. You wrote down a prior. You computed a likelihood under each hypothesis. The math told you the posterior. No vibes.
Two failure modes dominate:
Bayes is bookkeeping, not safety. It exposes your assumptions; it does not validate them.
From your backtest or first 100 trades:
Win rate
EV
Variance
Max drawdown
Treat this as your “prior belief”
Compare to your prior:
"Are we still inside the confidence interval?" "Is this likely noise — or something has changed?"
Never test ideas live without journaling their results independently.
A checklist proxy — not a true Bayesian update, but cheap to maintain weekly.
Create a 0–100 confidence score for your strategy:
| Criteria | Points |
|---|---|
| EV positive over last 50 trades | +20 |
| Win rate within historical range | +20 |
| Drawdown within expected bounds | +20 |
| Aligned with current market regime | +20 |
| Executed with discipline | +20 |
| Deviating from strategy | –15 |
| New risk factors emerging | –15 |
Update weekly.
| Score Drop | Interpretation | Action |
|---|---|---|
| 90 to 70 | Normal variance | Continue, monitor weekly |
| 90 to 40 | Structural concern | Pause, investigate, or adapt |
No. With a 55% win-rate system, a 5-loss streak has P(5L | edge) ≈ 0.018 — rare per occurrence but routine across 200 trades. Reassess at 30–50 trades, not 5. A streak alone is evidence about the streak, not about the edge.
Posterior = (Likelihood × Prior) / Evidence. Concretely: write down your prior P(edge) from your backtest, compute the likelihood of the observed trade sequence under both "edge intact" and "edge decayed" hypotheses, then apply P(H|E) = P(E|H)·P(H)/P(E). The lesson's worked example shows the full numeric path.
No. A points-based +20/−15 checklist is a heuristic dashboard, not a posterior probability. It is useful as a quick weekly sanity check but it does not weight evidence by likelihood and has no proper prior. Do not confuse it with computing P(edge | data).
System hopping after small losing streaks, emotional sabotage from recent outliers, and unwarranted overconfidence after a hot run. It forces you to weight new evidence against your sample-size-backed prior instead of overreacting to the most recent trades.
Confusing P(losing streak | edge) with P(edge | losing streak). The first is a likelihood; the second is what you actually want to know. Without the base rate of edge decay, the streak alone tells you almost nothing about whether your system is broken.
Trading is not about finding a system that never fails. It’s about continually adapting your trust in your system — based on data.
Trading is not about finding a system that never fails. It is about updating P(edge | what just happened) honestly — and most traders cannot, because they never wrote down P(edge) in the first place.
Bayesian thinking protects you from:
Start with structure. Collect evidence. Update your belief. Grow steadily — without overreacting.
Recommended next: Law of Large Numbers & Confidence Intervals — when your posterior actually converges to truth — and the Kelly Criterion, which formalizes how to size positions to your posterior, not your prior.