Statistics Glossary → Periodogram and spectral density
Periodogram and spectral densityconcept
A tool that shows how the total variance of a time series is distributed across frequencies (cycle lengths), revealing the dominant periods in the series.
Any time series can be written as a sum of sine and cosine waves at different frequencies. The periodogram uses the Fourier transform to compute how strong these waves are at each Fourier frequency (ωⱼ = 2πj/n): I(ωⱼ) = (1/n)·|Σ xₜ e^(−iωⱼt)|². A clear peak indicates a regular cycle at the corresponding period (period = 1/frequency); in monthly data, for example, a peak at frequency 1/12 is annual seasonality. The periodogram of white noise is roughly flat.
The raw periodogram is an inconsistent estimator of the true spectral density: the variance at each frequency does not shrink as the sample grows, which is why the plot looks very jagged. Windows such as Daniell, Bartlett, Tukey-Hanning, Tukey-Hamming or Parzen average neighbouring frequencies to give a smoothed spectrum, which reduces variance but spreads out sharp peaks. For two series, the cross-spectrum shows at which frequencies they move together (coherence) and the time shift between them (phase). A trend creates dominant power at low frequencies, so it is usually removed before the analysis.
An everyday example
An electricity company examines hourly consumption data. The periodogram shows two clear peaks, at periods of 24 hours and 168 hours (weekly). They indicate that consumption fluctuates regularly both within the day and between weekdays and weekends, so the forecasting model should include both cycles.
In MerQur
In time series analysis, the periodogram extra output gives the dominant periods from the raw periodogram. The 'Spectral window' option (None (raw periodogram), Daniell, Daniell (modified), Bartlett, Tukey-Hanning, Tukey-Hamming, Parzen) gives a smoothed spectrum, and 'Spectral extras: Cross-spectrum' gives the coherence and phase between two series.