I am performing an analysis on energy time and crack spreads. Normally I use
a return whenever calculating a correlation. Does it make sense to take a
return when correlating times series which go negative frequently.
For instance if the spread goes from -.5 to -.6, is this a +20% change? What
about when the spread moves from 0 to .5?
Other than cross correlation is there any other statistic in R which may
help to reveal lead/lag relationships?
In most case you might be calculating the correlation between that spread which is stationary in nature (quite logical to assume the stability/stationarity of a spread series) with some other price which may often be realization from some unit root process. In such case you might think the joint evolution of that bivariate system as a bivariate integrated VAR process (with the maximum order of integration of that system is 1). Assuming there is no co-movement exist between them, you need to take just the 1st difference of the entire system i.e. you are differentiating once both price and spread series. And estimate the correlation between them from the estimated VCV matrix of the Residuals.
On the other way, if that spread is not actually just tradable then, you can think that spread as a portfolio of 2 non-stationary asset series. Hence in such case you will have a trivariate VAR system with all variables are non-stationary, unlike the 1st case. Here again you can estimate the required correlation from estimated VCV matrix (which is of order 3 now) with little straightforward mathematics.
Hence if you think your problem of estimating correlation from a VAR system perspective, your problem will become quite simplified.