Deflated Sharpe Ratio (DSR)

The foundational statistical problem that the DSR solves. When a researcher tests thousands of trading rules on the same dataset, the probability of finding a false positive (a strategy that looks great purely by luck) increases dramatically.

• Bonferroni Correction: A simple but overly conservative adjustment that divides the significance threshold by the number of trials.
• Family-Wise Error Rate (FWER): The probability of making one or more false discoveries. The DSR controls for this in the context of Sharpe ratios.
• False Discovery Rate (FDR): An alternative framework that controls the expected proportion of false positives among all rejected hypotheses.

Also known as p-hacking or data dredging, this is the core pathology the DSR corrects. It occurs when a dataset is used more than once for inference or model selection, leading to over-optimistic performance estimates.

• Mechanism: A researcher iterates on a strategy until it backtests well, implicitly fitting the noise in the historical sample.
• Consequence: The in-sample Sharpe ratio becomes a severely biased estimator of the true out-of-sample Sharpe ratio.
• DSR's Role: The DSR explicitly models the distribution of the maximum Sharpe ratio across all trials, effectively asking: 'Given that I tried N strategies, how impressive is my best result?'

A practical heuristic closely related to the DSR, introduced by Harvey and Liu (2015). It provides a simple formula to deflate an observed Sharpe ratio based on the number of correlated strategies tested.

• Formula: The haircut is a function of the number of trials, the average correlation among strategy returns, and the observed Sharpe ratio.
• Application: A Sharpe ratio of 1.0 might be impressive if it was the only strategy tested, but could be statistically insignificant if it was the best of 500 correlated trials.
• Comparison to DSR: The Haircut Sharpe Ratio is a point estimate, while the DSR provides a full probability of significance.

The gold-standard empirical complement to the DSR's theoretical correction. Instead of testing a strategy on a single historical window, walk-forward analysis simulates real-time deployment by sequentially optimizing on an in-sample window and validating on a subsequent out-of-sample window.

• Anchored vs. Rolling: The in-sample start date can be fixed (anchored) or advanced (rolling) to create multiple validation periods.
• Purging: Removing overlapping data points between training and testing sets to prevent information leakage.
• Synergy with DSR: A high DSR probability combined with a robust walk-forward Sharpe ratio provides strong evidence of genuine predictive power.

A theoretical result formalizing the intuition behind the DSR. It states that with a sufficiently large number of trials, the expected maximum Sharpe ratio of a purely random strategy is strictly greater than zero.

• Mathematical Basis: The expected maximum of a set of random variables drawn from a distribution with unbounded support increases with the number of draws.
• Practical Implication: A Sharpe ratio of 0.5 or even 1.0 is not inherently impressive without knowing the context of the search effort.
• DSR Connection: The DSR operationalizes this theorem by computing the probability that the observed Sharpe ratio exceeds this expected maximum for a given number of trials.

A distinct but compounding bias that inflates backtested Sharpe ratios. It occurs when a dataset includes only entities that have survived to the present day, excluding those that have delisted, gone bankrupt, or been acquired.

• Impact: The average return of a survivorship-biased index is artificially high because the poor performers have been removed.
• Interaction with DSR: Survivorship bias inflates the nominal Sharpe ratio, making it even more likely to be a false positive. The DSR can help identify if an inflated ratio is statistically meaningful after accounting for this and other biases.
• Mitigation: Using point-in-time databases that preserve the historical composition of an index.