Time-Series, Stationarity, and Forecast Choice

How Level II tests trend choice, autoregression, stationarity, seasonality, ARCH, and forecast evaluation in time-series problems.

Time-series questions at Level II are about choosing a usable forecast structure, not about admiring a historical chart. The exam wants to know whether the series is stable enough to model, whether mean reversion is plausible, whether seasonality matters, and whether the chosen forecast actually outperforms alternatives.

Why This Lesson Matters

Candidates often make two mistakes here:

  • they use a familiar model before checking whether the data structure supports it
  • they interpret a forecast without asking how the model behaves out of sample

Level II is much more demanding. It expects model choice to follow the series.

Start With The Problem Structure

    flowchart TD
	    A["What kind of time-series problem is this?"] --> B["Deterministic trend dominates"]
	    A --> C["Series may follow an autoregressive process"]
	    A --> D["Volatility dynamics matter"]
	    A --> E["Several time-series variables may be linked"]
	    B --> F["Choose linear or log-linear trend"]
	    C --> G["Check stationarity, unit roots, and seasonality"]
	    D --> H["Consider ARCH-type volatility modeling"]
	    E --> I["Check nonstationarity and cointegration before regression"]

That sequence is more important than memorizing isolated vocabulary.

Trend Models Work Only For The Right Kind Of Series

A deterministic-trend model may be written as:

$$ \hat{Y}_t = b_0 + b_1 t $$

or, in log-linear form,

$$ \ln(Y_t) = b_0 + b_1 t $$

Trend choiceWhen it tends to fit better
Linear trendAbsolute changes are more stable
Log-linear trendPercentage growth is more stable

The exam often tests whether the candidate knows when a log-linear model is better because growth compounds rather than rises in equal absolute steps.

Autoregression And Mean Reversion Need Stationarity

An AR(1) process can be written as:

$$ Y_t = b_0 + b_1 Y_{t-1} + \varepsilon_t $$

If (|b_1| < 1), the process is covariance stationary and the mean-reverting level is:

$$ \frac{b_0}{1-b_1} $$

ConceptWhy it matters
Covariance stationaritySupports stable mean, variance, and autocovariance structure
Unit rootSignals nonstationarity and a weak basis for standard AR interpretation
Mean reversionMakes long-run level interpretation possible

A nonstationary series can make a respectable-looking regression or forecast deeply misleading.

In-Sample Fit And Out-Of-Sample Accuracy Are Different

The exam often contrasts models that fit history well with models that forecast better. A common accuracy measure is root mean squared error:

$$ \mathrm{RMSE} = \sqrt{\frac{1}{n}\sum_{t=1}^{n}(Y_t-\hat{Y}_t)^2} $$

ComparisonWhy Level II cares
In-sample versus out-of-sampleGuards against overfitting historical noise
Lower RMSE versus prettier theoryForecast usefulness matters more than elegance
Stable coefficients versus drifting coefficientsTime-series relations can break across regimes

Seasonality, ARCH, And Cointegration Change The Model Choice

IssueWhat it changes
SeasonalityForecasts may need seasonal lags or seasonal adjustments
ARCH behaviorThe variance process itself becomes forecastable and relevant
CointegrationMultiple nonstationary variables may still have a stable long-run relation

This is where Level II becomes practical. The right answer is usually the model that respects the data-generating behavior, not the model that looks easiest to compute.

How CFA-Style Questions Usually Test This

  • by asking whether a trend should be linear or log-linear
  • by giving an AR process and asking for a forecast or mean-reverting level
  • by asking whether a unit root invalidates the intended analysis
  • by comparing in-sample fit with out-of-sample forecasting accuracy
  • by forcing you to recognize seasonality, ARCH, or cointegration before running a regression

Mini-Case

Two models forecast monthly volatility. One fits the historical data more tightly, but the residual pattern shows volatility clustering and the out-of-sample error is worse.

A weak answer still prefers the tighter historical fit.

A stronger answer recognizes that the forecast problem is about usable forward variance behavior, which makes the volatility structure and out-of-sample evidence more important.

Common Traps

  • fitting an AR model before checking stationarity
  • choosing linear trend when compound growth behavior suggests log-linear
  • trusting in-sample fit more than out-of-sample evidence
  • ignoring seasonality or cointegration because the headline forecast looks reasonable

Sample CFA-Style Question

What is the strongest reason to test for a unit root before using a time series in an autoregressive model?

Best answer: A unit root indicates nonstationarity, which can invalidate the usual interpretation and forecasting framework of the AR model.

Why: Level II often tests model choice and diagnostic logic before computation.

Continue In This Chapter

Revised at Thursday, April 9, 2026