Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I am seeing a regression model which is regressing Year-on-Year stock index returns on lagged (12 months) Year-on-Year returns of the same stock index, credit spread (difference between monthly mean of risk-free bonds and corporate bond yields), YoY Inflation rate and YoY index of industrial production.

It looks thus (though you'd subsitute the data specific to India in this case):

SP500YOY(T) = a + b1*SP500YOY(T-12) + b2*CREDITSPREAD(T) +    

SP500YOY is the year-on-year returns for SP500 index To compute this, monthly average of SP500 values is computed and then converted to year-on-year returns for each month (i.e. Jan'10-Jan'11, Feb'10-Feb'11, Mar'10-Mar'11, . . ). On the explanatory variables side, a 12-month lagged value of the SP500YOY is used along with the CREDITSPREAD at time T and INFLATION and INDUSTRIALPRODUCTION two period AHEAD. The INFLATIONASYMM is a dummy for whether the Inflation is above a threshold value of 5.0%. The index in the parenthesis shows the time index for each variable.

This is estimated by standard OLS linear regression. To use this model for forecasting the 1,2 and 3-months ahead YOY returns of SP500, one has to generate 3,4 and 5-month ahead forecasts for Inflation and the Index of Industrial Production. These forecasts are done after fitting an ARIMA model to each of the two individually. The CreditSpread forecasts for 1,2 and 3 month ahead are just thrown in as mental estimates.

I'd like to know whether this OLS linear regression is correct/incorrect, efficient/inefficient or generally valid statistical practice.

The first problem I see is that of using overlapping data. i.e. the daily values of the stock index are averaged each month, and then used to compute yearly returns which are rolled over monthly. This should make the error term autocorrelated. I would think that one would have to use some 'correction' on the lines of one of the following:

  • White's heteroscedasticity consistent covariance estimator
  • Newey & West heteroscedasticity and autocorrelation consistent (HAC) estimator
  • heteroscedasticity-consistent version of Hansen & Hodrick

Does it really make sense to apply standard OLS linear regression (without any corrections) to such overlapping data, and more so, use 3-period ahead ARIMA forecasts for explanatory variables to use in the original OLS linear regression for forecasting SP500YOY? I have not seen such a form before, and hence cannot really judge it, without the exception of correcting for the use of overlapping observations.

share|improve this question
Please don't cross-post. – Joshua Ulrich Mar 16 '11 at 19:43
Josh, which exchange would be better for this question? – Vishal Belsare Mar 17 '11 at 19:30
I think it's better suited here, since it's about statistics, not programming. – Joshua Ulrich Mar 18 '11 at 14:38
up vote 7 down vote accepted

Here are a couple of articles that deal with this subject:

Britten-Jones and Neuberger, Improved inference and estimation in regression with overlapping observations

Harri & Brorsen, The Overlapping Data Problem

share|improve this answer
Thanks for this. – Vishal Belsare Sep 15 '11 at 17:59

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.