Autocorrelation is the correlation of a return series with its own past values.
Autocorrelation is the correlation of a return series with its own past values. When returns are autocorrelated, one observation contains information about nearby observations, so the series is not independent across time.
That matters because many standard significance tests assume independent observations. Sonar’s strategy-validation research lists autocorrelation among the statistical properties that should be checked before relying on classical inference from a return stream. The same framework describes diagnostics such as autocorrelation-function analysis and Ljung Box testing as part of return-stream validation.【1】
The mechanism is straightforward. Standard tests such as naive t-statistics treat each return as if it were a separate draw. If returns cluster or persist through time, the effective amount of independent information is smaller than the raw sample length suggests. In that setting, standard errors based on independence assumptions can be misstated, confidence intervals can be too narrow, and hypothesis tests can give misleading significance conclusions.【1】
Sonar’s research on strategy validation explicitly warns that inference should account for dependence in the return stream rather than relying on a naive backtest summary. Its validation checklist includes testing for autocorrelation and then using significance measures that remain valid when return data depart from idealized assumptions.【1】
The same point appears in Sonar’s glossary entry on the Deflated Sharpe Ratio. That document explains that simple Sharpe-ratio inference can be distorted by non-normality and by time-series effects in backtest data, and it presents the Deflated Sharpe Ratio as a more conservative significance framework for evaluating an observed Sharpe ratio in the presence of backtest distortions and multiple testing concerns.【3】 While that glossary page is not a dedicated treatment of autocorrelation, it supports the broader principle that naive significance measures can overstate evidence when return data violate classical assumptions.【3】
Sonar’s Backtest Overfitting Audit tool also frames strategy evaluation as more than reading headline performance statistics. The tool is designed to stress-test whether apparent evidence survives stronger validation procedures rather than assuming that a single backtest significance readout is reliable on its own.【2】 In the context of autocorrelation, that means first detecting dependence and then applying methods that are robust to it, instead of treating the return stream as independent by default.【1】【2】
In practice, the key idea is not that every autocorrelated series is unusable. The key idea is that autocorrelation changes the meaning of standard inferential statistics. If a return stream shows meaningful dependence through autocorrelation diagnostics, inference should be adjusted with methods that are robust to those distortions before drawing conclusions from significance tests.【1】
Covered in depth in the Strategy validation & overfitting pillar hub.