EViews.com

Hello,

I have data on credit spreads for Goldman Sachs (2007-2009, daily). I want to be able to forecast what the spreads would have been had the recent financial crisis had not occurred using macroeconomic data and firm specific data. For example, I run the regression:
GoldmanSpreads = c(1)+c(2)ROE+c(3)PB+c(4)DE+c(5)GDP
where ROE=return on equity (firm specific)
PB= price to book value (firm specific)
DE= debt to equity ratio (firm specific)
GDP=GDP growth (macroeconomic variable)

The output of this regression is great - I get very high (~0.9 adjusted Rsquared) but the Durbin Watson stat is low (0.4). There is obvious serial correlation going on.

However, if I add an AR(1) into this equation:
GoldmanSpreads = c(1)+c(2)ROE+c(3)PB+c(4)DE+c(5)GDP+AR(1)
the coefficient on AR(1) is so large that it makes the other variables that were originally significant to become NOT significant (according to t-stats).

I know that ROE, PB, DE, GDP growth are utilized and makes sense in a model of credit spreads. But, I want to correct for the low Durbin Watson Stat WITHOUT including AR(1) or something of the sort that takes away the original predictive power of the variables like ROE, etc.

Does anyone have suggestions?!

I am a senior undergrad writing my econ thesis and ANY inputs would be wonderful.

thanks in advance.

Try including a lagged value of the dependent variable AND AR(1). (It's not okay to leave out the AR(1)).

I've posted a similar question in another board:

I have specified a partial adjustment model (so I have a lagged dependent variable), but there is some serial correlation (looks like AR(1) disturbances).

If I include an AR(1) in the estimation line...does this change my interpretation of my partial adjustment model (and the coefficients that I'm backing out)? How do I interpret the coefficient on the AR(1) coefficient that shows up in the estimation output?

u can run ur model without AR(1) but include Garch (1,1) and check the results. u do not have to report the Garch effect.
the other point is is the unit of measurement for all the variables the same. did u run unit root test?
hope that helped

u can run ur model without AR(1) but include Garch (1,1) and check the results. u do not have to report the Garch effect.
the other point is is the unit of measurement for all the variables the same. did u run unit root test?
hope that helped

This is not good advice. Estimating with a lagged dependent variable in the presence of serial correlation leads to inconsistent estimates unless you correct for serial correlation, which including a GARCH term does not do.

ok startz