Autocorrelation
In a general context, "autocorrelation" (also called "serial correlation") is the correlation between observations of a time series with lagged observations of the same series.
In a performance context, positive autocorrelation of period returns can result in a positive (negative) relatiev return in one period followed by another positive (negative) relative return the next period. Negative autocorrelation can result in a positive (negative) relative returns followed by negative (positive) relative returns. On the other hand, a sign change (or an absence of) from observation to observation is not a requirement for the presence of autocorrelation.
Positive autocorrelation has also been observed in real estate, respectively subprime and ABS and CDO containing subprime mortgage exposures and hedge fund return time series. One reason for the existence of autocorrelation is illiquidity, which results in observed prices not being market prices reflecting all relevant information.
The problem with positive autocorrelation is that it will have a lower volatility than the uncorrelated series. As less volatility is usually preferred to more, the attractiveness of returns is therefore distorted (overestimated) in the presence of positive autocorrelation.
Several approaches to dealing with autocorrelation exist. A popular one is the Blundell/Ward filter, which uses the filter below...
r*(t) = 1/(1-a1) * r(t) - a1/(1-a1) * r(t-1)
r*(t)... the "decorrelated" return time series
r(t)... the original return time series
r(t-1)... the lagged (by one period9 return time series
a1... a coefficient from the regression below...
r(t) = a0 + a1*r(t-1) + e(t)
a0... a constant
e(t)... the usual regression error term
This filter has the advantage that the mean return remains more or less unchanged. So calculating risk-adjusted performance (see chapter 'External Analysis') will result in much less biased results.
A spreadsheet illustrating the Blundell/Ward filter can be downloaded here.
Durbin-Watson test for autocorrelation ...
Box-Ljung Statistic: H0 = autocorrelation is zero
|
