4.2 Summary of nonparametric spectral estimation
In summary, the different steps to undertake are, in order,
- Differencing (or detrending) should be performed prior to any analysis. The series should be centered around zero by subtracting the mean from the data to remove the frequencies near \(\omega=0\), which would otherwise dominate.
- For tapering, multiply centered observations by a data window. The cosine bell taper, also known as Hann taper, is the default choice in the R function
spectrum
. - Zero-pad to increase the resolution and so that the length of your time series is a highly composite number, typically \(2^l>n\). If you want to convert your spectrum to an autocovariance, zero-padding helps avoid “circular autocorrelation”. It also reduces the computation burden for the FFT calculation.
- Smooth the periodogram. It amounts to convoluting the periodogram with a kernel. The amount of averaging should increase as \(n \to \infty\). This reduces the variance, but introduces leakage. The amount of smoothing is controlled by the bandwidth parameter.