Reading ARIMA Output Without Mixing Up the Pieces
An ARIMA summary and its forecast table answer related but different questions. The model order describes how the series was differenced and how the resulting series is represented; coefficient output describes estimated terms; residual tests assess whether remaining serial dependence is evident; and forecast intervals quantify uncertainty at specified horizons. A sound interpretation keeps these pieces separate.
The most consequential traps are treating differencing as proof that the original series is stationary, interpreting every constant as a fixed increase in the level, and using the one-step forecast standard error at every horizon. Each mistake changes the conclusion or the forecast uncertainty.
For example, an ARIMA(\(1,1,1\)) model describes an ARMA(1,1) structure for \(\Delta Y_t\), not an ARMA(1,1) structure for \(Y_t\). If that differenced series is stationary, the levels can still be nonstationary. Differencing is a modeling operation; it is not a test result about the original levels.
Drift Depends on the Differencing Order
A constant in the equation for the differenced series is often called drift when \(d=1\). In an ARIMA(\(0,1,0\)) with drift \(c\), the model is \(\Delta Y_t=c+\varepsilon_t\), where the innovations have mean zero. The expected change in the level each period is \(c\). Thus the level forecast has a linear trend.
When \(d=2\), a constant instead appears in the equation for second differences. It changes the expected first difference each period, producing a quadratic rather than a linear pattern in the level forecasts. Output labels vary by software: inspect the model equation and the differencing order rather than relying only on a label such as “constant” or “mean.”
Coefficient Output and Residual Tests
A coefficient estimate divided by its reported standard error gives a Wald \(z\)-statistic (or, in some settings, a \(t\)-statistic). The associated p-value tests a null hypothesis about that coefficient under the fitted model’s assumptions. It does not test whether the entire model is adequate. In particular, a small p-value for drift does not show that residuals are uncorrelated.
A Ljung–Box test assesses the joint null that residual autocorrelations through a specified lag are zero. Its test statistic is compared with a chi-square reference distribution; a common degrees-of-freedom adjustment subtracts the number of fitted AR and MA coefficients from the number of tested lags. Software may use a different adjustment, so interpret the degrees of freedom shown with the reported result.
A large p-value means the test does not provide sufficient evidence against its null at the chosen significance level. It does not prove that residuals are independent, normally distributed, or otherwise ideal. The test concerns autocorrelation, not every possible form of model misspecification.
Worked Example: Separating Coefficient Evidence From Residual Evidence
A hypothetical insurer fits an ARIMA(1,1,1) model to quarterly claim-cost indices. The reported estimates are drift \(1.20\) with standard error \(0.50\), AR(1) coefficient \(0.35\) with standard error \(0.14\), and MA(1) coefficient \(-0.20\) with standard error \(0.11\). The residual Ljung–Box output reports \(Q(12)=14.8\), \(10\) degrees of freedom, and \(p=0.14\). At the 5% level, interpret the output. Assume the displayed p-values use two-sided large-sample coefficient tests and the stated chi-square reference for the residual test.
Method. First interpret \(d=1\), then calculate each estimate-to-standard-error ratio, and finally assess the residual test separately.
Because \(d=1\), the model is for first differences. The drift estimate means the fitted differenced process includes a constant term; it is not automatically a statement that the original levels are stationary.
For drift, \(z=1.20/0.50=2.40\), giving a two-sided p-value of approximately \(0.016\). For AR(1), \(z=0.35/0.14=2.50\), with p-value approximately \(0.012\). For MA(1), \(z=-0.20/0.11=-1.818\), with p-value approximately \(0.069\). At 5%, the first two coefficient tests reject their zero-coefficient nulls; the MA(1) test does not.
For the residual test, the stated \(p=0.14\) exceeds \(0.05\). We therefore fail to reject the null of no residual autocorrelation through lag 12 under the test’s reference distribution. This is not proof that the model is correct, nor does it make the MA coefficient significant.
Check. The three coefficient ratios are \(2.40\), \(2.50\), and approximately \(-1.82\); the separate residual-test conclusion is based on \(0.14\), not on those ratios. Common mistake avoided: using a coefficient p-value as if it were the residual-adequacy test.
Forecast Intervals Must Match the Horizon
A forecast table usually gives a point forecast and an interval for each lead time \(h\). A prediction interval concerns a future observation, not merely uncertainty about the conditional mean. Its width depends on the accumulated forecast error at that horizon. For the random walk with drift, assume independent innovations with mean zero and variance \(\sigma^2\), and condition on the observed value \(Y_T\). The \(h\)-step forecast error is the sum of the next \(h\) innovations.
The variance identity follows because the innovations are independent: the variance of their sum is the sum of their variances, \(h\) terms each with variance \(\sigma^2\). Under normally distributed innovations, a 95% prediction interval is the point forecast plus or minus \(1.96\) forecast standard errors. For other ARIMA models, forecast errors are weighted sums of innovations; the weights depend on the model, so the random-walk formula does not generally apply unchanged.
Worked Example: Recalculating Two Forecast Horizons
A hypothetical portfolio’s quarterly aggregate claim cost is measured in thousands of dollars. At time \(T\), the observed value is \(Y_T=100\). A fitted ARIMA(0,1,0) with drift has \(c=1.5\) thousand dollars per quarter and innovation variance \(\sigma^2=16\) in squared thousands of dollars. Assume independent normal innovations and known parameters. Calculate the two-step forecast and its 95% prediction interval, and compare it with the one-step interval.
Method. For this random walk with drift, use \(\widehat{Y}_{T+h|T}=Y_T+hc\) and forecast-error variance \(h\sigma^2\).
At horizon 1, the point forecast is \(100+1(1.5)=101.5\). Its standard error is \(\sqrt{1(16)}=4\). The margin is \(1.96(4)=7.84\), so the interval is \(101.5\pm7.84\), or \([93.66,109.34]\) thousand dollars.
At horizon 2, the point forecast is \(100+2(1.5)=103\). The forecast-error variance is \(2(16)=32\), and the standard error is \(\sqrt{32}=5.656854249\ldots\). The margin is \(1.96\sqrt{32}=11.087434\ldots\). The endpoints are \(103-11.087434\ldots=91.912566\ldots\) and \(103+11.087434\ldots=114.087434\ldots\). Rounded to two decimals, the interval is \([91.91,114.09]\) thousand dollars.
The two-step interval is wider because its forecast error includes two innovations rather than one. Check. The two-step point forecast is \(103\), while the standard error is \(\sqrt{32}\), not \(4\). Common mistake avoided: reusing the one-step uncertainty at horizon 2, or multiplying \(1.96\sqrt{32}\) inaccurately.
Identify \(p,d,q\), the series to which the ARMA terms apply, and whether the output includes a constant or drift.
Use estimate divided by standard error for a coefficient test, while keeping its null hypothesis distinct from model diagnostics.
Check the test statistic, lag count, degrees of freedom, and p-value; state only what the residual autocorrelation test supports.
Read or calculate the forecast error variance for that lead time before constructing an interval.
Worked Example: Interpreting a Constant With Second Differences
A hypothetical quarterly exposure index is modeled by \(\Delta^2Y_t=0.8+\varepsilon_t\), an ARIMA(0,2,0) with a constant in the second-difference equation. At time \(T\), \(Y_T=250\) and \(\Delta Y_T=3\). Assume future innovations have conditional mean zero. Find the forecasts at horizons 1 and 2.
Method. Take conditional expectations in the second-difference equation, then reconstruct the first differences and levels in sequence.
At horizon 1, \(E[\Delta^2Y_{T+1}\mid T]=0.8\). Since \(\Delta^2Y_{T+1}=\Delta Y_{T+1}-\Delta Y_T\), the expected first difference is \(3+0.8=3.8\). Therefore \(\widehat{Y}_{T+1|T}=250+3.8=253.8\).
At horizon 2, the expected second difference is again \(0.8\), so the expected first difference becomes \(3.8+0.8=4.6\). Therefore \(\widehat{Y}_{T+2|T}=253.8+4.6=258.4\).
The level increases by \(3.8\) and then \(4.6\); it does not increase by \(0.8\) each quarter. Check. The successive first differences rise by \(0.8\), as required by the model equation. Common mistake avoided: interpreting a constant in a second-difference equation as a constant increment in the original level.
Common Exam Traps
- Calling the level stationary because \(d=1\). Differencing may produce a stationary series, but the differenced model does not by itself establish stationarity of the original levels.
- Reading drift as a level intercept. Translate the constant through the differencing operation before describing its effect on forecasts.
- Claiming a large residual-test p-value proves adequacy. It means only that this test did not detect sufficient autocorrelation evidence at its selected lags and significance level.
- Confusing coefficient tests with residual tests. A significant AR or drift coefficient is not evidence that all remaining residual dependence has been removed.
- Using a single standard error for all leads. Forecast uncertainty generally changes with horizon; consult the corresponding row or derive the model-specific forecast-error variance.
- Calling a prediction interval a confidence interval for the mean. A prediction interval accounts for the future observation’s innovation as well as forecast uncertainty.
Check Your Understanding
For each problem, identify the relevant model feature before calculating or interpreting the result.
- An ARIMA(0,1,0) with drift has \(Y_T=52\), drift \(0.7\), and innovation variance \(4\). Under independent normal innovations, find the three-step point forecast and 95% prediction interval. Round endpoints to two decimals. (Answer: Point forecast \(54.1\); interval \([47.31,60.89]\).)
- A fitted ARIMA(2,1,0) has residuals that pass a Ljung–Box test at 10 lags with p-value \(0.22\). What does this say about the level series and the residuals? (Answer: The model uses first differences; the test does not reject no residual autocorrelation through the tested lags, but it proves neither level stationarity nor overall model adequacy.)
- A residual Ljung–Box test has \(Q=17.1\), 6 degrees of freedom, and a 5% critical value of \(12.592\). State the test decision. (Answer: Reject the null of no autocorrelation through the tested lags because \(17.1>12.592\).)
- For \(\Delta^2Y_t=0.5+\varepsilon_t\), \(Y_T=40\), and \(\Delta Y_T=2\), with zero-mean future innovations, calculate the forecasts at horizons 1 and 2. (Answer: \(\widehat{Y}_{T+1|T}=42.5\) and \(\widehat{Y}_{T+2|T}=45.5\).)
- A random walk with drift has innovation variance \(9\). Compare the forecast standard errors at horizons 1 and 4. (Answer: \(3\) and \(6\), respectively; the four-step variance is \(4(9)=36\).)