Hands-On Machine Learning for Algorithmic Trading

Serial correlation

Serial correlation means that consecutive residuals produced by linear regression are correlated, which violates the fourth GMT assumption. Positive serial correlation implies that the standard errors are underestimated and the t-statistics will be inflated, leading to false discoveries if ignored. However, there are procedures to correct for serial correlation when calculating standard errors.

The Durbin-Watson statistic diagnoses serial correlation. It tests the hypothesis that the OLS residuals are not autocorrelated against the alternative that they follow an autoregressive process (that we will explore in the next chapter). The test statistic ranges from 0 to 4, and values near 2 indicate non-autocorrelation, lower ...