Solving ARMA(1,1) Parameter and ACF Problems

Solving ARMA(1,1) Problems

An ARMA(1,1) process combines dependence on the previous observation with dependence on the current and previous innovations. Problems often ask for more than substitution into a formula: a candidate may need to recover parameters from autocorrelations, determine whether the model is stationary and invertible, or identify which parameter solution is admissible. The order of those steps matters. In particular, do not use a stationary ACF formula before checking the stationarity condition.

Throughout this tutorial, use the following sign convention. Let \(X_t\) be a process with mean \(\mu\), and let \(Z_t\) be white noise with mean zero and variance \(\sigma_Z^2\). The model is

White noise here means that the \(Z_t\) values have constant variance and are uncorrelated at distinct times. They need not be normally distributed for the second-moment results below. The sign convention is important: some sources write the moving-average term with a minus sign and therefore report a different signed MA parameter.

Conditions Before Calculating the ACF

The autoregressive lag polynomial is \(1-\phi B\), where \(B X_t=X_{t-1}\). Its root is \(1/\phi\) when \(\phi\ne 0\). Causality requires that this root lie outside the unit circle, which gives \(|\phi|<1\). The moving-average polynomial is \(1+\theta B\), with root \(-1/\theta\) when \(\theta\ne 0\). Invertibility requires that root to lie outside the unit circle, giving \(|\theta|<1\). When a parameter is zero, its polynomial has no finite root to test and its corresponding condition is satisfied.

For the usual stationary ARMA(1,1) autocovariance calculation, first establish \(|\phi|<1\). Invertibility is not needed merely to calculate autocovariances, but it is important when identifying a unique, standard MA representation from the model. In parameter-recovery problems, use it to reject a noninvertible alternative when the question specifies an invertible model.

Deriving the ARMA(1,1) Autocovariances

When \(|\phi|<1\), the model has the causal representation

Thus the coefficient of \(Z_t\) is \(\psi_0=1\), and for \(j\geq1\), the coefficient of \(Z_{t-j}\) is \(\psi_j=(\phi+\theta)\phi^{j-1}\). Since distinct white-noise terms are uncorrelated, the variance is \(\sigma_Z^2\sum_{j=0}^{\infty}\psi_j^2\). For lag \(k\geq1\), the covariance is \(\sigma_Z^2\sum_{j=0}^{\infty}\psi_j\psi_{j+k}\). These sums converge when \(|\phi|<1\).

For completeness, the positive-lag simplification follows by summing the covariance terms:

In the last equality, \(1-\phi^2+\phi(\phi+\theta)=1+\phi\theta\). The variance simplifies because \(1-\phi^2+(\phi+\theta)^2=1+\theta^2+2\phi\theta\). Therefore,

and

In particular, \(\rho_k=\phi\rho_{k-1}\) for \(k\geq2\). The ACF therefore tails off geometrically after lag one; it does not generally cut off after lag one. The denominator in \(\rho_1\) is positive under stationarity because \(1+\theta^2+2\phi\theta=(\theta+\phi)^2+1-\phi^2>0\).

Recovering Parameters From Several ACF Values

If \(\rho_1\ne0\), the lag relation gives a direct estimate of the autoregressive parameter from theoretical autocorrelations:

After finding \(\phi\), substitute it and \(\rho_1\) into the lag-one formula to solve for \(\theta\). The resulting equation is quadratic:

This equation can have two solutions. Check the original model conditions for each candidate; in an invertible model retain only a solution with \(|\theta|<1\). If \(\rho_1=0\), dividing \(\rho_2\) by \(\rho_1\) is invalid. Under stationarity, \(|\phi\theta|<1\), so \(1+\phi\theta\ne0\); the lag-one formula then implies \(\phi+\theta=0\). This is a cancellation case, not a reason to use the ratio formula.

Common Exam Traps

• Using the wrong MA sign convention. The formulas here assume \(+\,\theta Z_{t-1}\). If the equation instead has \(-\,\theta Z_{t-1}\), replace \(\theta\) by \(-\theta\) before applying these formulas.
• Using \(\rho_1/\rho_2\) for \(\phi\). The correct relation is \(\rho_2=\phi\rho_1\), so when \(\rho_1\ne0\), \(\phi=\rho_2/\rho_1\).
• Confusing stationarity with invertibility. Check \(|\phi|<1\) for causality and \(|\theta|<1\) for invertibility. One condition does not imply the other.
• Ignoring multiple parameter roots. The equation for \(\theta\) can produce more than one algebraic solution. Apply the stated invertibility condition and verify the selected root in the original formula.
• Assuming the ACF cuts off at lag one. That cutoff is characteristic of an MA(1), not a general ARMA(1,1). For this model, \(\rho_k=\phi^{k-1}\rho_1\) for positive lags.

Parameter recovery from exact theoretical autocorrelations is an algebra problem. In applied work, estimated sample autocorrelations are subject to sampling variation and may not obey the theoretical relations exactly. The formulas remain useful for understanding the model, checking candidate parameters, and recognizing whether a proposed solution is consistent with stationarity and invertibility. These techniques are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.