ARMA Correlogram Lab: Reading AR and MA Order from the ACF and PACF

Two rows of bars on a plinth: sample ACF and PACF of a simulated ARMA series, blue beads for the exact theory, glass slabs for the white-noise band. Drag to orbit.

Switch to 2D
Controls

Presets

Parameters

Quality

Camera

Model checks

The science

Model. A zero-mean ARMA(2,1) process xt = φ1xt-1 + φ2xt-2 + et + θ et-1 with Gaussian noise. The sliders are the lag-1 and lag-2 partial autocorrelations a1, a2 (so φ2 = a2, φ1 = a1(1-a2)), which keeps the process stationary for every slider setting. Theory: with the psi-weights ψ0 = 1, ψ1 = φ1+θ, ψj = φ1ψj-1+φ2ψj-2, the autocovariance is γk = σ² ∑j ψjψj+k.

Reading the plots. A pure AR(p) has a PACF that cuts off after lag p while its ACF tails off; a pure MA(q) has an ACF that cuts off after lag q while its PACF tails off; a mixed ARMA tails off in both (Box and Jenkins' identification rule). The dotted band ±1.96/√m is the 95 % limit for white noise only; for a correlated series the true standard error is larger (Bartlett's formula, used in the checks).

Simplifications. The rule-of-thumb "reading" counts bars outside the band and is a heuristic, not a model-selection criterion (AIC/BIC would be used in practice). Gaussian noise, known zero mean, no seasonality, no differencing. Checks: a 100,000-point simulation against theory in Bartlett standard errors, psi-weight ACF against the Yule-Walker recursion, Durbin-Levinson PACF against direct Toeplitz solving, closed-form AR(1)/MA(1) special cases, and the process variance.

Loading model...
Loading Blender model...
-
-
Made with 🟠 Blender · Three.js