Bayesian Updating for Trading Expectations

Frequentist statistics asks whether an edge is real; Bayesian statistics asks for your best current estimate given the evidence. For live trading, the Bayesian question is the useful one.

The Framework

You hold a prior belief about a system's win rate, then update it with each trade. The result is a posterior distribution reflecting everything you know, weighted by evidence.

A win rate is a probability, so use a Beta prior with parameters α and β, interpretable as "prior wins" and "prior losses":

• Prior: Beta(α0, β0)
• After observing w wins and l losses: Beta(α0 + w, β0 + l)
• Posterior mean win rate: (α0 + w) / (α0 + β0 + w + l)

Choosing the Prior

The prior encodes your belief before seeing data:

• Skeptical Beta(5, 5): equivalent to 10 prior trades at 50%, strong shrinkage toward 50%.
• Weak Beta(1, 1): uniform, lets data dominate immediately.
• Industry-informed Beta(7, 6): encodes that trend systems typically win 40-55% of trades.

Use the skeptical prior for new systems; it prevents overreacting to the first 20 lucky trades.

How Updating Works in Practice

Start with Beta(5, 5), mean 0.50. After 30 live trades with 18 wins, 12 losses:

Posterior = Beta(5+18, 5+12) = Beta(23, 17) Posterior mean = 23/40 = 0.575

After 200 live trades with 110 wins:

Posterior = Beta(5+110, 5+90) = Beta(115, 95) Posterior mean = 115/210 = 0.548

Notice the shrinkage: the early 60% regressed toward 54.8% as evidence accumulated; the posterior self-corrects.

Decisions From the Posterior

The posterior is a distribution, not a point. Use it for decisions:

• Credible interval: if the 90% interval sits entirely above 50%, you have a real edge; if it straddles 50%, evidence is inconclusive.
• Position sizing: scale size with posterior mean minus 50% (the no-edge baseline); tighter variance permits larger size.
• Kill rule: if the posterior 5th percentile drops below 0.50, the system likely has no edge; reduce size or stop.

Updating Other Quantities

The same logic extends beyond win rate: use a normal-inverse-gamma prior on mean R-multiple (expectancy) and update with each trade's R. Bayesian Sharpe estimation produces a distribution accounting for sample size, more honest than a point estimate.

Avoiding Common Errors

• Too strong a prior: with Beta(50, 50), 30 trades barely shift the posterior. Match prior strength to the evidence you expect to need.
• Updating on design trades double-counts. Start the Bayesian clock on fresh forward trades, not backtest trades.
• Treating the posterior mean as the true win rate ignores uncertainty. Use the credible interval, not the point, for sizing and kill decisions.

The Honest Benefit

Bayesian updating formalizes the discipline of revising beliefs with evidence. Instead of declaring a system "good" after 30 trades or "dead" after a 10-trade drawdown, you maintain a living probability distribution of the edge, and decisions follow from it rather than from emotional reactions.