Core Path · L6.1.03

Time series: stationarity and autocorrelation

14 minARMAACF and PACFUnit rootsDickey-FullerHalf-life

Know these first

By the end you will be able to

  1. 1Tell AR(1), MA(1) and ARMA(1,1) apart by their ACF and PACF, fit them, and judge the fit with Ljung-Box on the residuals.
  2. 2Explain why risk models work with changes rather than levels, and what stationarity buys.
  3. 3Run and read the augmented Dickey-Fuller test, and explain why its critical values are not the t-test's and why it has little power near φ = 1.
  4. 4Estimate a mean-reversion half-life with an interval, recognise when an AR(1) is too simple for it, and convert φ into Vasicek's κ.

Why it matters

The US 10-year yield began 2006 at 4.33% and ended 2025 at 4.20%, after a low of 0.52% and a high of 5.20%. Does it come back? The Vasicek model of L5.3.02: The Vasicek model says yes, at speed κ. Twenty years of daily data cannot tell: the standard test for a unit root gives p = 0.39. Whether a series has a mean to return to decides whether a risk model may use its level, how fast its shocks fade, and whether a correlation between two such series means anything at all.