nevermind

I dont know about how "sensible" it is .. but I would imagine that the p+l is going to be volatile.In addition to correlation you should include the variance (or covariance) or look at how close the tracking is.

QuoteOriginally posted by: daveangelI dont know about how "sensible" it is .. but I would imagine that the p+l is going to be volatile.In addition to correlation you should include the variance (or covariance) or look at how close the tracking is.What would be the correct formula to use?

QuoteOriginally posted by: brianhough24QuoteOriginally posted by: daveangelI dont know about how "sensible" it is .. but I would imagine that the p+l is going to be volatile.In addition to correlation you should include the variance (or covariance) or look at how close the tracking is.What would be the correct formula to use?for covariance ? or tracking error ? tracking error is the standard deviation of the difference of returns of the two assets - asset 1 is your stock, asset 2 is the basket

double post

QuoteOriginally posted by: brianhough24QuoteOriginally posted by: daveangelQuoteOriginally posted by: brianhough24QuoteOriginally posted by: daveangelI dont know about how "sensible" it is .. but I would imagine that the p+l is going to be volatile.In addition to correlation you should include the variance (or covariance) or look at how close the tracking is.What would be the correct formula to use?for covariance ? or tracking error ? tracking error is the standard deviation of the difference of returns of the two assets - asset 1 is your stock, asset 2 is the basketNo, I meant, what would be the proper equation for pricing of the obscure stock when taking into account correlation, covariance, etc?the hedge ratio between the long and short positions to minimise the variance is given byh = - Covar(A,B)/Var(A)or if your write the Covar() in terms of correlation(covar(a,B) = rho*volA*volB)h = - rho * volB/volA

One caution is that covariance from close-to-close data will aggregate both day time (normal trading hours) and night time events but the estimator you are trying to construct is for the night time. Given the reality of timezones and different companies having different exposures to events in different countries, the correlations between stocks will be different at night than they are for total 24 hour period. Thus, you might want to use close-to-open data to estimate the off-hours correlations.

The "proper equation" will be highly biased toward your view of the market, which is your model, and which is always wrong.The linear relations that you have identified in the last two years will probably not be valid that day (like one of the stocks going wild etc)..It makes sense to start with a linear model with gaussian returns assumption, which puts your model into the CAPM/APT realm, but I suggest to understand their shortcomings.In that case see: http://en.wikipedia.org/wiki/Capital_as ... cing_model and http://en.wikipedia.org/wiki/Arbitrage_pricing_theory

Thanks for the links

Hi Brian,Your question was clear actually, but the answer depends on your assumptions. The correlation and covariance are both related statistical measures that summarizes your data. You can see correlation as the adjusted(normalized) covariance. I think the figure below explains what I was trying to explain quite well:The image is from wikipedia, and was initially created by Francis Anscombe. As the associated page underlines, all four series have the same mean (7.5), standard deviation (4.12), correlation (0.816) and regression line (y = 3 + 0.5x). Let's assume now that they are your successive X-Y return observations, and the right-most X value is the last observation which you're trying to estimate. The correlations will give an estimation on what your error rate will be if you do a linear regression(proof needed-)). But as you can see, your estimation error rate will inevitably also depend on the time-evolution.So if you want a short answer, use linear regression on the X-Y pairs. This means that you assume gaussianity of errors on both of your series. If you want a more robust approach, with time-variant relationships, use a kalman filter, which also assumes gaussianity but updates the underlying linear model recursively. Depending on how much complexity you want to introduce, you can use extended kalman filter or other methods under non-linearity assumptions.Whatever method/model you use, it is important to understand about their assumptions, and verify that they hold for your data.I hope this helps.