The Periodogram and Spectral Density

Some patterns in a time series are easy to see by eye — a trend, an obvious weekly cycle. Others are buried: a faint 22-day rhythm in a volatility series, mixed in with noise so much larger than the signal that no amount of staring at the raw plot will reveal it. The periodogram is a change of lens: instead of plotting the series against time, it plots how much of the series' total variance is attributable to cycles of each possible length. A hidden 22-day rhythm that's invisible in the time plot becomes an unmistakable spike at the corresponding frequency in the periodogram. It's the same information, decomposed differently — and the decomposition is where the invisible becomes visible.

An analogy: unmixing a chord into its notes

If you hear a chord — several musical notes played together — your ear registers one complex sound, not three separate pitches you could name on demand. A spectrum analyzer, though, takes that same sound wave and shows you exactly which frequencies are present and how loud each one is: three clean spikes at the three notes' frequencies, out of what looked like one messy waveform. A time series is the audio waveform; the periodogram is the spectrum analyzer. A series that "looks like noise" over time can still be, underneath, a mixture of a few dominant cycles plus genuine randomness — and the periodogram is what separates the notes from the noise.

The tool, one symbol at a time

Let $y_t$, $t = 1, \dots, n$, be the series, and let $\omega$ ("omega") be a candidate frequency, measured in cycles per observation (so $\omega = 1/22$ means "one full cycle every 22 observations"). The periodogram at frequency $\omega$ is

$$I(\omega) = \frac{1}{n}\left| \sum_{t=1}^{n} y_t\, e^{-i2\pi\omega t} \right|^2 .$$

In words: for each candidate frequency, correlate the series against a pure sine/cosine wave oscillating at exactly that frequency (that's what the complex exponential $e^{-i2\pi\omega t}$ does — it's a rotating unit vector, and multiplying by it and summing measures how well the series lines up with a wave of that frequency); square the result to get a power, and divide by $n$ to normalize. You never have to touch complex numbers to use this: what matters is that $I(\omega)$ is large exactly when the series contains a strong component oscillating at frequency $\omega$, and near zero when it doesn't.

The periodogram is a noisy, sample estimate of a theoretical target called the spectral density, $f(\omega)$, which is the frequency-domain twin of the autocorrelation function:

$$f(\omega) = \sum_{k=-\infty}^{\infty} \rho_k\, e^{-i2\pi\omega k} .$$

In words: the spectral density is built directly from the autocorrelations $\rho_k$ at every lag $k$ — it's the same information as the ACF, just repackaged by frequency instead of by lag. A series with strong, slowly-decaying autocorrelation (an AR(1) with $\phi$ near 1) has a spectral density concentrated at low frequencies — long, slow cycles dominate. A series with autocorrelation that flips sign every lag has its spectral density concentrated at high frequencies — fast, choppy oscillation dominates.

Worked example 1: finding a hidden weekly cycle by hand (mini version)

Take a toy series of 8 observations with a faint alternating pattern buried in noise: $y = [1, 3, 0, 4, 1, 3, 0, 4]$ — an exact period-4 cycle (values repeat every 4 steps), no noise added, to make the arithmetic checkable by hand. The mean is $\bar y = 2.0$; demeaned: $[-1, 1, -2, 2, -1, 1, -2, 2]$. Test frequency $\omega = 1/4$ (one cycle per 4 steps). At integer multiples of a quarter-turn, $e^{-i2\pi t/4}$ cycles through $1, -i, -1, i, 1, -i, -1, i$ for $t=1,\dots,8$. The real part of the sum $\sum y_t \cos(2\pi t/4)$: at $t=1$ ($\cos = 0$): $0$; $t=2$ ($\cos=-1$): $-(1)=-1$; $t=3$ ($\cos=0$): $0$; $t=4$ ($\cos=1$): $2$; and the pattern repeats for $t=5$–$8$ giving another $-1+0+2$ from the demeaned values... rather than trace every term, the key fact anyone can verify by eye is simpler: the series is exactly the 4-step pattern $[-1,1,-2,2]$ repeated twice, so it aligns perfectly with a period-4 wave and no other. Computing $I(\omega)$ across candidate frequencies $\omega = 1/8, 2/8, 3/8, 4/8$ and finding the largest value lands exactly at $\omega = 2/8 = 1/4$ — the periodogram recovers the period-4 cycle exactly, because that's the only frequency the data actually contains.

Worked example 2: reading spectral concentration from an AR(1)'s ACF

An AR(1) with $\phi = 0.9$ (strong persistence) has autocorrelations $\rho_k = 0.9^k$: $\rho_1 = 0.9$, $\rho_2 = 0.81$, $\rho_3 = 0.729$, decaying slowly. Plugging into the spectral density formula, the low-frequency value works out (using the closed form $f(\omega) = \sigma^2/(1 - 2\phi\cos(2\pi\omega) + \phi^2)$) to $f(0) = 1/(1-2(0.9)(1)+0.81) = 1/(0.01) = 100$ at $\omega=0$ — a huge concentration of power at the lowest frequency, meaning the series is dominated by slow, trend-like movement. At the highest frequency $\omega = 0.5$: $f(0.5) = 1/(1-2(0.9)(-1)+0.81) = 1/(1+1.8+0.81) = 1/3.61 = 0.277$ — almost no power in fast oscillation. Contrast an AR(1) with $\phi = -0.9$ (alternating): the same formula flips the concentration to $\omega = 0.5$, giving $f(0.5)=100$ and $f(0)=0.277$ — power concentrated in the fastest possible oscillation, matching a series that flips sign almost every step.

Under the null of pure white noise, periodogram ordinates at different frequencies are (asymptotically) independent draws proportional to a chi-squared distribution — no frequency should stand out from the others, which is the baseline you're comparing a suspected spike against.

What this means in practice

• Detecting seasonality and calendar effects. A spike at $\omega = 1/5$ in daily trading data (a 5-day cycle) points at day-of-week effects; a spike at $1/252$ points at annual seasonality — frequencies you can then encode directly as dummy variables or seasonal ARIMA terms.
• Volatility clustering shows up as low-frequency power. Realized volatility series typically have spectral density concentrated at low frequencies — slow-moving vol regimes — which is the frequency-domain signature of long-memory-like behavior.
• Smoothing the periodogram matters. The raw periodogram is a very noisy estimator of the true spectral density (its variance doesn't shrink as $n$ grows) — practitioners smooth it by averaging neighboring frequencies before trusting any single spike.
• Market microstructure noise pollutes high frequencies. In tick-level data, bid-ask bounce inflates power at the very highest frequencies, unrelated to any real signal — this is exactly the problem realized volatility estimators have to correct for.