statistical arbitrage

nitro · Aug 27, 2003

Ok, there is some meat there. Let me digest it...

nitro

Quote from Satyrican:

Fair enough.

Yes, Cointegration measure long term stabiilty but you forgetting that you can measure the speed of mean reversion! Which is the heart of stats arb.

In addition with vector autoregression (VAR), you can measure the speed of mean reversion for Any set of variable, while at the same time, "holding" other variable constant.

With regard to correlation, you can show that correlation converges into a stationary process if you increase the number of observation. Spurious correlation. Thus very little value to traders.

You can also show that in the SHORT RUN, the time frame of interest, volatility dominate return ~white noise if you will. Thus correlation based on RETURN on a short time frame offers very little info. You can test this by running a simple Monte Carlo (sp?) setup with two randomized sets. In the short run, volatility, not return, influence correlation calculation. Cointegration is "established" as a mean of correcting this all important spurious correlation.

Compounding the problem is the fact of the length of computation for correlation. How do you decide?

Your comment:"Cointegration is a TWO step PROCESS: first any long-run equilibrium realtionships between prices are established using Ordinary Least Squares (OLS) regression (plus some statistics to verify), and THEN a dynamic CORRELATION model of returns is estimated. The Error Correcting Model (ECM), so called because short-term deviations from equilibrium are corrected."

Wrong. The "standard process" for stats arb is to run VAR and use the mean speed of reversion as your catalyst, not correlation dynamics. Why would you use correlation model? With VAR you can "hold" other parameters constants like interest rate fluctuation, implied volatility, etc. Correlation does not, it is very very crude. Unless your're talking about some nonlinear correlation - which I never heard of it - I dont see the sense of using correlation analysis.

Moreover, residual analysis would have been a better bet in of itself. If you already established with cointegration the cointegration vector, why would you need correlation???? You can simply use that cointegration vector and ECM, that would suffice. Where does correlation fall under?

----Satyrican

nononsense · Aug 27, 2003

Could somebody, please point me to a reference on "cointegration" in the sense that it is used here.

(I often use statsoft's "Electronic Statistics Textbook" searches for quick clarifications - it did not return anything, which is unusual)

bone · Aug 27, 2003

How is it that I trade all manner of inter and intra-market spreads and have no idea what the hell you guys are talking about? (OK, a little, and I do surf the Wilmot website from time to time, and I am an Engineer by training).

I am proudly bliss in my ignorance.

I revel in my stupidity.

If I get a decent correlation analysis on the Bloomberg and it charts real purdy I trade 'em.

Oh... the problem with statistical arbitrage is fat tails. Ask LTCM about those fat tails. Maybe Niederhoffer, too. (list is getting bigger)

FishSauce · Aug 27, 2003

Quote from bone:

How is it that I trade all manner of inter and intra-market spreads and have no idea what the hell you guys are talking about? (OK, a little, and I do surf the Wilmot website from time to time, and I am an Engineer by training).

I am proudly bliss in my ignorance.

I revel in my stupidity.

If I get a decent correlation analysis on the Bloomberg and it charts real purdy I trade 'em.

Oh... the problem with statistical arbitrage is fat tails. Ask LTCM about those fat tails. Maybe Niederhoffer, too. (list is getting bigger)

Good point. There is more than one way to skin a cat.

Caveat: I think you're spreading and not necessary stats arb ... technical definition.

nitro · Aug 27, 2003

Because you are a TRADER and not some quant sitting in a room looking for at 15% return year and you don't jump up and down in a stuppor if you return 15% a year.

Correlation analysis is fine for "short term" pair trading, and the stuff that Bloomberg returns is pretty useful and a similar analysis that I used to use when pair trading. IMHO, Cointegration is meant for those that believe in CAPM or APT and want "effcient regime" long/short portfolios over a longer time frame.

BTW, I credit you with getting me thinking about more "exotic" pairs, like the GE/ES trade - it opened up a whole new world to me.

nitro

Quote from bone:

How is it that I trade all manner of inter and intra-market spreads and have no idea what the hell you guys are talking about? (OK, a little, and I do surf the Wilmot website from time to time, and I am an Engineer by training).

I am proudly bliss in my ignorance.

I revel in my stupidity.

If I get a decent correlation analysis on the Bloomberg and it charts real purdy I trade 'em.

Oh... the problem with statistical arbitrage is fat tails. Ask LTCM about those fat tails. Maybe Niederhoffer, too. (list is getting bigger)

hermit_trader · Aug 27, 2003

Quote from nitro:

Because you are a TRADER and not some quant sitting in a room looking for at 15% return year and you don't jump up and down in a stuppor if you return 15% a year.

nitro:
Do you mean those quant only looking for 15% return p.a? Is this what a quant expected after studying so much math/finance?

nitro · Aug 27, 2003

Hermit_trader,

I am being really hard on them. Returning 15% a year on $1B is REALLY hard.

At 20% of profits for hedge funds, that's 20% of $150,000,000. That's a nice payday!

nitro

Quote from hermit_trader:

nitro:
Do you mean those quant only looking for 15% return p.a? Is this what a quant expected after studying so much math/finance?

hermit_trader · Aug 27, 2003

Quote from nitro:

Well,

I am being really hard on them. Returning 15% a year on $1B is REALLY hard.

at 20% of profits for hedge funds, that's 20% of $150,000,000. That's a nice payday!

nitro

How many quant got $1B to play with?

nitro · Aug 27, 2003

What about $250M then? 15% of $250M is $37M. 20% of that is still a nice payout! And don't forget the 2% management fee!

nitro

Quote from hermit_trader:

How many quant got $1B to play with?

bigbeech · Aug 27, 2003

seems to be dead now- thats why its being mass marketed