Bayesian Versus Frequentist Interpretation

Two analysts backtest the same strategy and both report "there's a 95% chance the true Sharpe ratio is above 1.0." One of them is using that sentence loosely and technically shouldn't say it; the other means it exactly as written. The difference isn't sloppiness — it reflects two genuinely different definitions of what "probability" even means, and knowing which one you're using changes what you're allowed to claim from the same data.

An analogy: a loaded die you can't re-roll

Imagine asking "what's the probability this specific die, right now, showing a 4, is actually loaded?" A frequentist recoils from that question — probability, to a frequentist, is a long-run frequency: the fraction of times an event happens under infinite repetition. The die is either loaded or it isn't, so "probability it's loaded" isn't well-defined in that framework. Instead a frequentist asks a repeatable question: "if I ran this test many times on a fair die, how often would it get flagged as loaded?" A Bayesian answers the original question directly — probability, to a Bayesian, is a degree of belief, and beliefs can attach to one-off facts. The die's status is fixed, but your knowledge of it is uncertain, and that uncertainty is what a Bayesian probability quantifies.

The maths, one symbol at a time

Let $\theta$ denote an unknown fixed quantity you care about — a strategy's true Sharpe ratio, a coin's true bias, a model's true coefficient. Let $D$ denote observed data.

The frequentist treats $\theta$ as a fixed, non-random constant. Estimation procedures are evaluated by how they behave over hypothetical repeated data: a 95% confidence interval is a procedure such that, across many repeated experiments, 95% of the intervals it produces contain the true $\theta$. In any single instance, $\theta$ is either in the interval or it isn't — there's no probability statement about this interval.

The Bayesian instead treats $\theta$ itself as uncertain and assigns it a probability distribution, updated via Bayes' theorem:

$$p(\theta \mid D) = \frac{p(D \mid \theta)\, p(\theta)}{p(D)}$$

In words: your updated belief about $\theta$ after seeing the data ($p(\theta \mid D)$, the posterior) is proportional to how well $\theta$ explains the data ($p(D \mid \theta)$, the likelihood) times what you believed before seeing any data ($p(\theta)$, the prior). This lets a Bayesian make a direct probability statement about $\theta$ itself, like "there is a 95% probability the true Sharpe ratio exceeds 1.0" — a credible interval, not a confidence interval (see credible intervals).

Worked example 1: interpreting the same interval two ways

A backtest of 200 independent daily returns gives a 95% frequentist confidence interval for the strategy's true Sharpe ratio of $[0.3, 1.9]$. The correct frequentist reading: "if I reran this backtest on fresh, equally-typical data many times and built an interval the same way each time, 95% of those intervals would contain the true Sharpe." It says nothing about the probability the true Sharpe is in this particular $[0.3, 1.9]$. A Bayesian, running the same data through a prior and likelihood, might report a 95% credible interval of $[0.35, 1.85]$ and mean it literally: "there is a 95% probability the true Sharpe lies in $[0.35, 1.85]$."

Worked example 2: a coin with a strong prior belief

A coin is flipped 10 times and lands heads 8 times. A frequentist's maximum-likelihood estimate of the bias is simply $8/10 = 0.8$. A Bayesian who starts from a prior belief that most coins are close to fair — concentrated tightly around $0.5$ — combines that prior with the same data and might land on a posterior mean closer to $0.6$: the data pulls the estimate up from $0.5$, but the strong prior keeps it well short of the raw $0.8$. Ten flips is not much evidence, and the Bayesian approach lets that fact — via the prior — visibly temper the conclusion.

Try shifting the amount of data in the explorer above and watch the posterior move away from the prior as evidence accumulates — this is the mechanism that makes example 2 work.

What this means in practice

Frequentist tools (p-values, confidence intervals, hypothesis tests) dominate classical backtesting and academic factor research because they don't require specifying a prior. Bayesian tools show up wherever prior domain knowledge is valuable and cheap to encode — sizing bets under parameter uncertainty, updating a model as new data trickles in, or combining a small sample with a strong structural belief (see Bayesian inference). Neither is "more correct" in general; mixing up which one you're using is the single most common statistics mistake in quant writing.