How Stambaugh Bias Affects Predictive Return Regressions

Summary

The document explains Stambaugh bias in regressions that use persistent lagged predictors to forecast stock returns. It distinguishes the issue from generic small-sample bias: the predictor follows an autoregressive process, and its innovation can be correlated with the contemporaneous return-regression error. That correlation links estimation error in the predictor’s persistence to bias in the estimated return-forecasting coefficient.

The explanation first describes finite-sample bias in an OLS estimate of an AR(1) coefficient, then relates the bias in that persistence estimate to bias in the predictive slope. The direction depends on the covariance between the predictor innovation and the return error, so it need not always be upward or downward. The text gives an approximate expression for AR(1) bias and a relationship for the predictive coefficient, but provides no derivation or empirical demonstration. It is a concise theoretical account; readers should consult the cited research for assumptions, refinements, and methods to correct inference.

Key ideas

• Stambaugh bias arises when predictor innovations correlate with contemporaneous return-regression errors.
• Estimation error in predictor persistence can transmit bias to the estimated return-predictability slope.
• The bias direction depends on the covariance between the predictor innovation and the return error.
• The discussion presents an approximation and encourages consulting the underlying research for details.

Tags

# What is the stambaugh bias? Why is it important for predictability regressions?


# What is the stambaugh bias? Why is it important for predictability regressions?












What is the Stambaugh bias? Why is it important for predictability regressions?

Can anyone explain it in simple terms?

## Answer by phdstudent (score 11, accepted)

https://quant.stackexchange.com/a/38806

The bias comes from the paper Stambaugh (1999) and has nothing to do with small sample bias. It has to do with point (1) below.

The argument goes as follows:

- Typical lagged explanatory variables for stock-return regressions are correlated with contemporaneous stock returns

- This contemporaneous correlation biases forecasting regressions

First review OLS bias of AR(1):

\begin{equation} x_t = \alpha + \rho x_{t-1} + v_t \end{equation}

\begin{equation} \hat{\rho} = \frac{\hat{Cov} (x_t, x_{t-1})}{\hat{Var} (x_{t-1})} \end{equation}

\begin{equation} \hat{\rho} = \rho + \frac{\hat{Cov} (v_t, x_{t-1})}{\hat{Var} (x_{t-1})} \end{equation}

Stambaugh shows that there is no analytical formula but as an approximation the bias is given by:

\begin{equation} E_t(\hat{\rho}) - \rho = - \frac{1+3\rho}{T} \end{equation}

Now assume that the predictor of stock returns follows the process $x_t$ above. If returns $r_t$ follows:

\begin{equation} r_t = \alpha + \beta x_{t-1} + u_t \end{equation}

Then you can see the bias of $\beta$:

\begin{equation} E(\hat{\beta}) - \beta = \frac{Cov(u_t, v_t)}{Var(v_t)}[E{(\hat{\rho})}-\rho] \end{equation}

Depending on the sign of $Cov(u_t, v_t)$ you get the sign of the bias.

I strongly recomment reading the reference above.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.