Page 1 of 1

PCA questions

Posted: April 3rd, 2014, 6:09 pm
by theRedBaron
I want to use PCA to trade a set of related fixed income products, for intraday trading with short holding times (2mins - 30 mins)1) Do I populate the data matrix with yields or prices, and what difference does it make?2) What is a reasonable sampling rule/frequency/number of samples/lookback window if I am aiming for holding times of a few minutes?3) Suppose I sample every 5 minutes, and my last sample was 1 minute ago. How do I compute the fair prices now, without inserting a new sample in my dataset, i.e. using the same "coefficients", but applied to the most recent state of the market?4) When it comes to hedging, is it customary to be neutral with respect to the first component only, or the first two? 5) Related to 1), is it Ok to populate the input matrix with first differences(p(t+1)-p(t))? Those will look stationary, and anyway on top of that I would apply normalizations(mean=0, stdev=1)

PCA questions

Posted: April 3rd, 2014, 8:16 pm
by katastrofa
If you want to hold a few minutes, I am not sure if sampling every 5 minutes is frequent enought. You will be betting on a single return from your distribution going in the right direction.

PCA questions

Posted: April 3rd, 2014, 8:18 pm
by theRedBaron
The example given in part 3) is specific to that part, not to the general setting. Any ideas on the other topics?

PCA questions

Posted: April 3rd, 2014, 8:25 pm
by katastrofa
"When it comes to hedging, is it customary to be neutral with respect to the first component only, or the first two?"Don't fix the number of components hedged, rather find out what the cost of not hedging a component will be (e.g. a'la VaR) and compare it to the cost of hedging it (bid/ask spread?). Then make a decision.

PCA questions

Posted: April 3rd, 2014, 8:34 pm
by theRedBaron
Thanks a lot.How about the other issues? I am really curious as to what do I put in the matrix of data? Textbooks warn against prices, and suggest, for example, first differences. But then, how do I find fair prices, when my matrix doesn't contain prices? I can only find fair first differences.Also, if I have first differences in it, will the loadings be still useful for computing hedging coefficients? For example, can I still use the ratios between the entries in the first loading as weighs for hedging against the first component?