Tue Aug 31, 04 12:15 PM This may be a very elementary question but sometimes one needs to go back and clear the very basics!I'm trying to determine the variability in the dependant variable (y=stock price) that is explained by changes in the independant variable (x 1 = spot price of a commodity).One way i can do this in excel is by calculating the daily returns and simply calculating the correlation coefficient and squaring it.If i use the Regression tool in Excel and regress the absolute stock price against the absolute spot price of the commodity, I obviously get a different answer since we are regressing the absolute price and not daily returns.My question is this: What is the correct way to do a regression analysis in excel using two different sets of prices e.g. trying to determime how much variability in the stock price is explained by changes in the spot price of a commodity. Do you use the returns or the absolute price? And obviously, regression is a complicated process so let's ignore problems with serial correlation etc. for now.Thanks

Usually people always use return. The only time I heard people use price is when they use cointegration.

So what if I had a multiple regression equation that is forecasting futures pricesThe dependent variable is (y=absolute futures price).The independant variables could be demand, supply, some economic indicator AND the price of another commodity.Question: Should i be using the absolute price of the other commodity OR the daily retruns as the independant variable (x4).

we don't use absolute prices because research has empirically shown that most price data (CPI, commodity price, GDP) are usually NON-STATIONARY, unless you have a 100 years of data or something.Jasemin's comment on cointergration, is correct in that when two variables are indiviudally non-stationary, the cointegrated element can be stationary. Without stationarity you would not know for sure whether your results are 'true' or may be a false pattern having occured by random chance. if you want to use price data - You can try differencing the price data, or taking logs, growth rates or some sort of transformation to ensure stationarity. I prefer logs or growth rates, because the interpretation is more intuitive.

QuoteOriginally posted by: KL Without stationarity you would not know for sure whether your results are 'true' or may be a false pattern having occured by random chance.I don't know where you got this notion from, but basically it's wrong. Regression and stationarity are pretty independent concepts...

hmmm, well the statement is a bit sloppy. But it's not that wrong is it ?Statistical test tells you whether it occurs by chance or not.Stationarity tells you whether there exists a unit root, because if it does exist then interpretation of model results may be doubtful.A regression on a non-stationary series can lead to spurious results with misleading DW, T-stats, and R^2.Thats the bottom line on why you need to transform it to ensure stationarity. If you difference the variable- you may lose useful information, but if you don't then your results are doubtful(may be spurious). Thats where ECM and cointegration come in.

just to make sure, can you elaborate, what form of stationarity you are referring to?

Ok, I think we are talking about different things here. If you talk about regression models as commonly used in econmetrics you need some form of stationarity, if you talk about in regression in general (like speed v time in an accellerating system) stationarity, unit roots etc. don't belong there...