What does the probability of backtest overfitting (PBO) measure?

Updated 21 September 2026 · Versión en español

PBO is the probability that the configuration which looked best in-sample ranks below the median of everything you tried, once you look out of sample. It judges the search, not the strategy: a PBO near 0.5 means picking the in-sample winner was about as good as picking at random — the optimizer found noise.

How it is estimated: CSCV

Bailey, Borwein, López de Prado and Zhu (2017) estimate it with combinatorially symmetric cross-validation:

Take the return series of every configuration you tried — N of them, over the same history.
Cut that history into S equal blocks (S even; ten is common).
For every way of choosing half the blocks as in-sample and the other half as out-of-sample — with ten blocks that is 252 splits — find the configuration with the best in-sample Sharpe.
Look at where that same configuration ranks out of sample. Its relative rank is ω = rank ÷ (N + 1), and its logit is λ = ln(ω ÷ (1 − ω)).
PBO is the share of splits in which λ ≤ 0 — that is, in which the in-sample winner landed at or below the out-of-sample median.

Because every split is used and in-sample and out-of-sample swap roles symmetrically, no single lucky cut of the history decides the answer.

How to read the number

PBO	What it says about the search
near 0	The in-sample ranking tends to survive out of sample.
around 0.5	Choosing the in-sample best was a coin flip.
above 0.5	The in-sample winner tends to do worse than a typical configuration out of sample — the search is selecting for luck.

An illustration, not a result: you try 20 configurations over a history cut into 10 blocks, giving 252 splits. If the in-sample winner lands below the out-of-sample median in 139 of them, PBO is 139 ÷ 252 ≈ 0.55. Whatever that winner's backtest looked like, you have no evidence it is better than any other setting you tried.

What it cannot tell you

It says nothing about future returns. A low PBO means the ranking held up on this history; markets can still change.
It needs a real search to judge: with two or three configurations the estimate is coarse, and with one it does not exist.
It inherits the data. Costs left out, survivorship bias or look-ahead in the backtests themselves pass straight through it.

Its companion: the Deflated Sharpe ratio

Bailey and López de Prado (2014) ask the same question from the other side: given how many configurations were tried, how non-normal the returns are and how short the sample is, what is the probability that the best Sharpe ratio is genuinely above zero? The more you search, the higher the bar the winner has to clear.

Where TradingZona shows it — and where it does not

Portfolio Lab ▸ Optimize computes PBO by CSCV and the Deflated Sharpe and prints both beside the optimized result, with the number of splits and trials behind each. The backtest results add block-bootstrap confidence intervals, because one Sharpe over one history is a point estimate.

Strategy optimization reports a different and simpler number: an overfitting score — how much of the in-sample Sharpe is lost out of sample — plus a robustness report built from the walk-forward windows. That report is what the deploy gate reads before it lets a strategy go live. It is not PBO, and we do not call it PBO.

Sources

Bailey, D. H., Borwein, J., López de Prado, M., Zhu, Q. J. (2017). The Probability of Backtest Overfitting. Journal of Computational Finance 20(4), 39–69.
Bailey, D. H., López de Prado, M. (2014). The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality. Journal of Portfolio Management 40(5), 94–107.

TradingZona is research and signal software. It places no orders, it is not a broker and it is not investment advice. A backtest is hypothetical: past or simulated performance does not guarantee future results. Trading involves substantial risk of loss.