Why a Break Can Look Like Persistence
A time series that rises and then stays high may appear to have strong autoregressive dependence: recent values are informative about future values. But there are two different explanations. The series may have persistent stochastic behavior, in which shocks continue to affect later observations, or it may have undergone a structural break, such as a sudden change in its average level or trend. These explanations imply different forecasts, so confusing them can matter in an actuarial setting.
A structural break is a change in the process generating observations. This tutorial focuses on changes in level and trend. A level break shifts the series to a new baseline; a trend break changes its systematic rate of increase or decrease. Persistent dependence, by contrast, describes how shocks to a stable process carry forward over time. In practice, a break and persistent dynamics can coexist, and finite data rarely identify the explanation with certainty.
The key is to assess multiple kinds of evidence together: the time plot, changes in the series, residual behavior under a candidate model, and forecast errors on both sides of a suspected change. A high estimated AR coefficient or a slowly declining sample autocorrelation is not, by itself, proof of a persistent process.
Useful Patterns in Levels, Differences, and Residuals
Suppose a series has a stable baseline \(\mu\), plus a level shift of size \(\delta\) at time \(\tau\), and a disturbance \(e_t\):
Here \(I(t\geq\tau)\) equals 1 at and after the change and 0 before it. If the disturbance is otherwise stable, the level shift contributes one large change at \(\tau\), then no further changes from the step itself. This differs from a persistent shock, which generally has a sequence of effects after it occurs.
For example, in a stable AR(1) process centered at its mean, a shock of size \(a\) has expected effects \(a,\phi a,\phi^2a,\ldots\), where \(|\phi|<1\). These effects decay gradually. A level break instead contributes a lasting displacement to the mean. If an ordinary ARIMA model is fitted without representing that break, the estimated persistence may partly be an attempt to accommodate the lasting shift.
A trend break leaves a different signature. A piecewise-linear mean can be written as:
where \((t-\tau)_+=\max(t-\tau,0)\). Before \(\tau\), the slope is \(\beta_1\); after \(\tau\), it is \(\beta_1+\beta_2\). A trend change can therefore produce a sustained difference in the average increments, rather than a single isolated increment. A plot of levels and a plot of first differences can be useful together.
Worked Example: A Sudden Level Shift or Ongoing Persistence?
A hypothetical insurer records monthly counts of a particular type of roadside assistance request. The series is \(100,101,99,100,110,111,109,110,111,109,110\). Assess whether the pattern is more consistent with a level shift or with continuing persistence. For this diagnostic, suppose the within-regime disturbances fluctuate around zero; do not assume the observations are a random sample from a known distribution.
Step 1: Inspect the levels. The first four values are \(100,101,99,100\), centered at 100. The next seven are \(110,111,109,110,111,109,110\), centered at 110. The latter values stay near a different baseline rather than gradually returning to the earlier one.
Step 2: Calculate all first differences. Subtract each observation from the next:
The fourth difference is \(110-100=10\). The remaining differences are small and fluctuate around zero. This is the pattern expected from a level shift of about 10: the shift contributes one large increment, while the new level persists.
Step 3: Compare with gradual persistence. If the series were high because one shock was gradually dissipating, a sequence of elevated values tending back toward the original baseline would be more plausible. Here the values do not show that decline. That does not rule out persistent disturbances in addition to a break, but the visible change is better described by a level shift than by gradual decay alone.
Check. The mean of the first four values is \((100+101+99+100)/4=400/4=100\). The mean of the last seven is \((110+111+109+110+111+109+110)/7=770/7=110\). The 10-unit difference between these group averages agrees with the isolated 10-unit increment. The common mistake is to read the high later values as evidence of a highly persistent AR process without checking whether the series changed baseline.
Residuals and Forecast Errors as Evidence
Residual diagnostics help assess whether a proposed model has represented the pattern. If a stable ARIMA model is fitted to a series with an omitted level change, residuals may show a run of one sign, a large residual near the change, or a shift in their average. These patterns indicate that the model has left systematic structure unexplained. The exact pattern depends on the model and on the timing of the break.
Residual autocorrelation is also relevant, but it is not a break detector by itself. A residual ACF with notable correlations may indicate remaining dependence, an omitted level or trend change, or another model inadequacy. Check the residual time plot as well: autocorrelation summaries can conceal a clear change in location. Conversely, a model can absorb some effects of a break and leave residuals that appear less problematic, while still forecasting poorly after the change.
Forecast errors provide a decision-focused check. Compare errors from forecast origins before and after a suspected change, keeping forecast horizon and evaluation period in view. If a model performs adequately before the change but has a sustained bias afterward, that is evidence of a changed process or a specification that cannot adapt to it. A single large error is weaker evidence; it might be ordinary noise. Rolling-origin evaluation and horizon-specific summaries help show whether the pattern is systematic.
Worked Example: Persistent AR Response Versus an Omitted Step
Consider two hypothetical deviations from a baseline. In the first case, a centered AR(1) model has \(\phi=0.8\), and an initial shock is 10 with no later innovations. In the second case, a process has a level shift from 0 to 10 and remains at 10. For each case, calculate the first five post-event deviations and the corresponding AR(1) one-step residuals \(r_t=Y_t-0.8Y_{t-1}\), using the preceding value 0 at the first post-event time.
Case A: Persistent AR(1) response. Start with a deviation of 10. Each next deviation is \(0.8\) times the previous one:
The residual at the first post-event time is \(10-0.8(0)=10\). At the second time it is \(8-0.8(10)=0\). The next three are \(6.4-0.8(8)=0\), \(5.12-0.8(6.4)=0\), and \(4.096-0.8(5.12)=0\). The initial shock is followed by a decaying response.
Case B: Permanent level shift. The five post-event deviations are \(10,10,10,10,10\). The AR(1) residuals are \(10-0.8(0)=10\), followed by four values of \(10-0.8(10)=2\). The repeated residual value of 2 arises because the model predicts only 8 from a previous value of 10, while the shifted process remains at 10.
Interpretation and check. The AR response decays by a factor of 0.8 at each step; the level-shift sequence remains constant. Under the stated calculation, the persistent AR model has zero residual after the initial shock, while the model applied to the step leaves four residuals of 2 after the first residual of 10. The common mistake is to call both patterns “high persistence” because later observations remain elevated. Their paths and residual patterns are different.
Assessing a Trend Change with Forecasts
For a suspected trend break, compare the direction and size of increments on either side of the candidate change. A stable trend extrapolated through a changed slope can create forecast errors that grow systematically with horizon. A random-walk-style forecast may adapt to a newly observed level, but it does not by itself explain why the trend changed. Candidate models should be evaluated using only information available at each forecast origin; using later observations to choose or estimate a model for an earlier forecast would leak future information.
Worked Example: Forecast Errors After a Change in Trend
A hypothetical quarterly claim count is observed as \(52,54,56,58,60,70,71,69\) at times \(t=1,\ldots,8\). The first five observations follow the simple trend \(50+2t\). At time 6, a change is suspected. Compare forecasts for times 7 and 8 from (A) continuing the original trend and (B) using a post-change flat level equal to the observed value at time 6. Treat these as explicitly specified candidate forecasts, not as fitted models with estimated uncertainty.
Step 1: Forecast by continuing the pre-change trend. The original rule gives \(50+2(7)=64\) and \(50+2(8)=66\). The actual values are 71 and 69, so the forecast errors, defined as actual minus forecast, are \(71-64=7\) and \(69-66=3\).
Step 2: Forecast by holding the newly observed level. The level at time 6 is 70, so the flat-level forecasts are 70 and 70. The errors are \(71-70=1\) and \(69-70=-1\).
Step 3: Compare errors. For the original-trend forecasts, the mean absolute error is \((|7|+|3|)/2=5\), and the mean squared error is \((7^2+3^2)/2=(49+9)/2=29\). For the flat-level forecasts, the mean absolute error is \((|1|+|-1|)/2=1\), and the mean squared error is \((1^2+(-1)^2)/2=1\).
Check and interpretation. The post-change level forecast has smaller errors at both stated horizons. This supports adapting to a change rather than blindly extrapolating the old trend, but two forecast errors are not enough to establish the future process. In particular, the example does not show that a flat level will remain appropriate indefinitely. The common mistake is to compare only the first post-change error or to select a model using information from times 7 and 8 before making the forecasts. Here the candidate forecasts are defined using observations through time 6, and the later values are used only for evaluation.
A Practical Diagnostic Sequence
When a fitted ARIMA model appears highly persistent, use a sequence of checks rather than immediately adding or removing differencing. The earlier material on ARIMA identification and residual diagnostics provides the tools for evaluating dependence; the goal here is to check whether an apparent dependence pattern could instead be associated with a change in level or trend.
Inspect the level plot for a sudden shift or a sustained change in slope. Note whether the apparent transition is abrupt, gradual, or uncertain.
For a level shift, look for a large increment near the change followed by a new baseline. For a trend change, compare average increments before and after. Check whether residuals show runs, a nonzero average, or a change in spread.
Assess whether stable ARIMA dependence describes the pattern, or whether a level or trend change needs to be represented. Do not interpret a high persistence estimate in isolation.
Compare forecast errors at matching horizons before and after the candidate change. Keep the forecast origin fixed to the information available then, and interpret short evaluation periods cautiously.
If the change time was selected after searching the whole series, apparent evidence for a break can be overstated: the same data were used both to locate the change and to assess it. In a forecasting assessment, preserve the time order and avoid allowing observations from after a forecast origin to influence model selection for that origin. A genuine process change may also have several causes, and a model that represents one observed break does not guarantee that the future will follow the new regime.
Common Exam Traps
- Equating a lasting high level with AR persistence. A permanent level shift and a slowly decaying stochastic response are not the same pattern. Inspect the path and the increments.
- Treating one large difference as proof of a break. An unusual observation can occur without a structural change. Look for a sustained change in the level or trend and supporting residual or forecast evidence.
- Using a single residual diagnostic as a verdict. Residual autocorrelation may reflect dependence or omitted structure; residual plots and context are also needed.
- Ignoring forecast horizon. Errors at different horizons answer different questions. Compare like with like before attributing a change in performance to a break.
- Using future observations to set an earlier forecast. A break detected using the full series cannot be treated as if it were known at an earlier forecast origin.
The diagnostic principle is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus: distinguish model structure from apparent pattern, justify the interpretation using available evidence, and state what the data do and do not establish.
Check Your Understanding
For each problem, distinguish the evidence for a structural change from evidence for persistence. Answers are given at the end of each item.
- A monthly index is \(40,41,39,40,46,47,45\). Calculate its first differences and identify the pattern most consistent with the values. (Answer: Differences are \(1,-2,1,6,1,-2\); the isolated increase of 6 followed by values near 46 is more consistent with a level shift than a gradual decay.)
- A centered AR(1) model has \(\phi=0.5\) and an initial shock of 12, with no later innovations. What are the next three expected deviations after the initial deviation of 12? (Answer: \(6,3,1.5\); the effects decay geometrically.)
- A candidate model forecasts 82 and 84 for two successive periods; actual values are 88 and 89. Calculate the mean absolute error and state what the errors suggest, without claiming proof of a break. (Answer: Errors are \(6,5\), so MAE is \(5.5\); consecutive positive errors suggest underforecasting and could be consistent with a level or trend change, but do not prove one.)
- A series follows a pre-change trend with slope 3. After a suspected change, its first differences are \(3,0,1,0\). Is a new slope of 3 still a good summary of the post-change increments? (Answer: No; the post-change average increment is \((3+0+1+0)/4=1\), suggesting a lower average slope, subject to sampling variation.)
- A level shift of 8 occurs at time \(\tau\), with no other changes. What contribution does the shift make to first differences at \(\tau\) and at later times? (Answer: It contributes 8 at \(\tau\) and 0 at each later time.)