CFA Level II Exam · Model Misspecification

Time-Series Misspecification in CFA Level II Regression

Updated 7 October 2026 · Fact-checked

Time-series misspecification is a model-building error specific to data observed over time. The curriculum's main cases are a lagged dependent variable with serial correlation, a function of the dependent variable used as a regressor (forecasting the past), independent variables measured with error, and nonstationarity. To solve questions, name the error, then state its effect on coefficients and tests.

Understand Time-Series Misspecification

A regression is misspecified when its form or variables do not match the process that generated the data. Misspecification can make coefficients biased or inconsistent, and it can make standard errors unreliable. Then t-tests and F-tests mislead you.

Time-series data add their own traps, because observations are ordered and often related to their own past. The curriculum's time-series misspecifications are these.

1. Lagged dependent variable with serial correlation. If you include the lagged dependent variable (Y(t-1)) as an independent variable and the errors are serially correlated, the regressor is correlated with the error term. This breaks the assumption that regressors are uncorrelated with errors. The estimates are then inconsistent (biased even in large samples).

2. Using a function of the dependent variable as an independent variable (forecasting the past). This happens when an independent variable is built from information not known at the time the dependent variable is observed, so it is in effect a function of the dependent variable. A common case is using a year-end value, such as year-end book value, to explain returns that were earned during that year. The fit looks strong, but the model could not have been used to forecast. Fix it by using only data available at the start of the period (for example, lagged values).

3. Independent variables measured with error. If a regressor is a proxy or is measured with noise, it is correlated with the error term in the regression. The slope estimate is biased, and in the simple one-variable case it is biased toward zero (attenuation). It stays inconsistent in large samples. It is hard to predict the direction of bias with several regressors.

4. Nonstationarity. If the dependent or independent variables have a trend, a unit root, or a changing mean or variance, the relationship can be spurious. Regression of one random walk on another can show a high R-squared and significant t-statistics with no true link. Test for a unit root, and difference the data (or use a cointegration approach) if it is present.

Key formulas to remember

Lagged dependent variable model: Y(t) = b0 + b1·X(t) + b2·Y(t-1) + ε(t); Problem arises when ε(t) is serially correlated: Y(t-1) is then correlated with ε(t), so estimates are inconsistent.
Measurement error in one regressor (simple regression): Estimated slope is biased toward zero (attenuation); Holds for one regressor measured with classical noise. With multiple regressors the direction of bias is not predictable.
Random walk (unit root): x(t) = x(t-1) + ε(t), so b1 = 1 in x(t) = b0 + b1·x(t-1) + ε(t); Nonstationary: mean-reversion level undefined, variance grows over time. First difference is stationary.
First difference: Δx(t) = x(t) - x(t-1); Common fix for a unit root before regressing.

How to solve Time-Series Misspecification questions

Use this routine for any vignette that describes a time-series regression problem.

1Read the vignette for how each variable is defined and when its data are observed.
2Check timing: does any independent variable use information not available at the start of the dependent variable's period? If yes, it is forecasting the past.
3Check whether the dependent variable's own lag is a regressor and whether the exhibit shows serial correlation in the errors (for example, a significant Breusch-Godfrey result). If both, the estimates are inconsistent.
4Check whether any regressor is a proxy, estimate, or noisy measure. If so, expect an inconsistent and (in the one-variable case) attenuated slope.
5Check whether the series trend or look like random walks, or whether a unit-root test fails to reject. If so, the regression may be spurious.
6State the consequence: biased or inconsistent coefficients, or unreliable standard errors and tests.
7State the remedy: use lagged information, better measures or instruments, correct for serial correlation, or difference the data.
8Match your answer to the option that gives both the correct error and the correct effect.

Quickest way: Four-question scan

When to use it: When time is short and the options name the error and its effect.

Ask: is any regressor known only after the period ends? Forecasting the past.
Ask: is Y(t-1) on the right side with serial correlation? Inconsistent estimates.
Ask: is a regressor a proxy or estimate? Bias, toward zero for a single regressor.
Ask: do the series trend or have a unit root? Spurious regression; difference the data.
Eliminate options that mix the error with the wrong effect.

Common mistakes in Time-Series Misspecification

Saying a lagged dependent variable always causes misspecification.
Students remember the lag as a problem without the condition.
Fix: Remember the trouble comes from the lag combined with serially correlated errors. If the errors are not serially correlated, including a lag does not create this problem.
Saying measurement error in a regressor only inflates standard errors.
It is confused with multicollinearity or heteroskedasticity.
Fix: Measurement error makes the regressor correlated with the error, so the coefficient itself is biased and inconsistent.
Assuming bias is always toward zero.
The one-variable attenuation result is over-generalized.
Fix: Say toward zero only for a single mismeasured regressor. With several regressors the direction is uncertain.
Missing forecasting the past because the model has a high R-squared.
A strong fit looks like a good model.
Fix: Check dates. If a regressor is observed at or after the end of the dependent variable's period, the fit is not usable for forecasting.
Trusting a high R-squared and significant t-stats on trending series.
Students treat statistics as proof of a relationship.
Fix: Test for a unit root first. Two unrelated random walks can give spurious results; difference them.

Worked examples

Example 1

An analyst regresses a fund's annual return on its year-end portfolio price-to-book ratio and on last year's return. Exhibit: a Breusch-Godfrey test shows significant serial correlation in the residuals. Q1: What forecasting error is present? Q2: What problem does the lagged return create?

Show the solution

Q1: The year-end price-to-book ratio is known only at the end of the year in which the return is earned.
So the regressor uses information not available at the start of the period. This is forecasting the past.
Q2: Last year's return is a lagged dependent variable. With serially correlated residuals it is correlated with the error term.
That violates the assumption that regressors are uncorrelated with errors, so the coefficient estimates are inconsistent.

Answer: Q1: Forecasting the past, because year-end data explain returns earned during the year. Q2: Lagged dependent variable with serial correlation, giving inconsistent estimates.

Example 2

A researcher regresses a stock's return on a single proxy for market sentiment that is measured with substantial noise. Separately, she regresses one index level on another index level; both trend upward, show R-squared of 0.94, and both fail a unit-root test. Q1: What is the likely effect of the noisy proxy on the slope? Q2: What is the concern with the second regression, and what is the remedy?

Show the solution

Q1: A single regressor measured with error is correlated with the regression error term.
The estimated slope is biased toward zero and remains inconsistent in large samples.
Q2: Both series fail the unit-root test, so both are nonstationary.
A regression of one nonstationary series on another can give a high R-squared and significant t-statistics with no true relationship, which is spurious.
Remedy: first-difference the series (or test for cointegration) and re-estimate.

Answer: Q1: The slope is biased toward zero and inconsistent. Q2: Spurious regression from nonstationary data; difference the series or test for cointegration.

Exam tips

Match the error to a trigger phrase: year-end or later data means forecasting the past; proxy or estimate means measurement error; lagged Y plus serial correlation means inconsistency.
Check the exhibits for unit-root test results before trusting any R-squared or t-statistic in a time-series regression.
When an option says bias toward zero, confirm there is only one regressor.
Questions often ask for the effect and the remedy together, so know both.
There is no penalty for wrong answers, so always pick an option after eliminating the clearly wrong ones.

Time-Series Misspecification in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

Time-Series Misspecification: frequently asked questions

What is forecasting the past in a regression?

It means an independent variable uses information that was not available when the dependent variable's period began. For example, year-end book value used to explain that same year's return. The model fits well but cannot be used to forecast.

Why is a lagged dependent variable a problem with serial correlation?

The lagged variable depends on past errors. If errors are serially correlated, the lag is correlated with the current error. This violates a key assumption and makes the coefficient estimates inconsistent.

How does measurement error affect regression coefficients?

A regressor measured with error is correlated with the error term. The coefficient is biased and inconsistent. With one regressor the bias is toward zero; with several, the direction is not predictable.

How does nonstationarity cause misspecification?

If the series have trends or unit roots, regressions can show a strong but false relationship, called spurious regression. Test for a unit root and difference the data, or test for cointegration.