It is widely known that correlation strategies applied to the popular pairs degenerate with time. As I can see from my personal experience there are two types of degeneration:1) gradual decrease of Sharpe ratio2) some kind of “phase transition” then strategy just stops work immediately (2-3 months scale)What can be the reason of such “phase transition”? Are there any research papers on the subj?

Does this degeredation apprear in backtesting? Or on a forward looking basis?I could see how if you base your trading rulesaround the correlation in the past year, a step change might occur when an earnings report comes out, or some fundamental change in the conditions of one equity realtive to another occurs.

>>Does this degeredation apprear in backtesting? Or on a forward looking basis?Backtesting and production. Of course you remove money from the pair if it stops working.The goal is to predict such transitions and to stop broken pair trading before the first losses.So I am looking for criteria and trying to find analogy with smth like "critical point" approach http://www.arxiv.org/abs/cond-mat/9510036

What sort of timeframe are you looking at? What is your stop-loss policy? Can you compare the historical deviation from the mean and then bail out when it exceeds that?QuoteOriginally posted by: olegs>>Does this degeredation apprear in backtesting? Or on a forward looking basis?Backtesting and production. Of course you remove money from the pair if it stops working.The goal is to predict such transitions and to stop broken pair trading before the first losses.So I am looking for criteria and trying to find analogy with smth like "critical point" approach http://www.arxiv.org/abs/cond-mat/9510036

There is a famous paper on just this subject ... I would search for all references to it on citeseer:They find #1 (gradual decrease over a period of years).Pairs Trading: Performance of a Relative Value Arbitrage Rule; EVAN G. GATEV, WILLIAM N. GOETZMANN, K. GEERT ROUWENHORSTThe obvious reason for #2 is that DE Shaw found it.

QuoteOriginally posted by: crowlogic What sort of timeframe are you looking at? Position ~ hours (up to 2 days)Correlation estimation ~ 1week - 1 monthQuoteWhat is your stop-loss policy?80% - time-interval (if you don't get mean revertion you close the position)20% - drowdownQuoteOriginally posted by:mdlm Pairs Trading: Performance of a Relative Value Arbitrage Rule; EVAN G. GATEV, WILLIAM N. GOETZMANN, K. GEERT ROUWENHORSTMany thanks

QuoteOriginally posted by: olegsMany thanksYou are welcome.I have a question for you: Given that pairs trading has been known for 20+ years, why do you think that relatively simple pairs strategies (of that sort that you would expect undergrads working on a homework assignment to find) still work?

And why they shouldn’t if the fundamental market causality exists?There are several ways to use it:1) search for new pairs (equities mostly);2) use exotic correlations: rank, etc.;3) use “pairs” instead of pairs: clusters etc.

Quoteolegs: And why they shouldn’t if the fundamental market causality exists?I'm not attacking. I'm curious.What is the fundamental market causality?

QuoteWhat is the fundamental market causality?It's different for different pairs The common pair is two equities from the same industry, but causalities could be rather exotic. See for example "Chernobyl-potato" effect from the Liars Poker.My point is that you should have a "physical model" to pick a profitable pair.Btw, this model is essential for the stop-loss policy.psI have a nice example in my mind but we trade it, so I can't make it public.

QuoteOriginally posted by: mdlmGiven that pairs trading has been known for 20+ years, why do you think that relatively simple pairs strategies (of that sort that you would expect undergrads working on a homework assignment to find) still work?Has it made any billionaires? Do quants maintaining the strategy make more than lawyers getting paid by the hour? Perhaps "work" equals risk-free rate plus just enough to pay off student loans.Anyway, I would expect the lag to be traded out of the correlation, before the correlation itself deteriorates. There are plenty of highly correlated things which you can't cover your costs trading against one another.

QuoteOriginally posted by: farmerHas it made any billionaires? David Shaw is pretty close. One of the pioneers of pairs trading was Nunzo Tartaglia at Morgan Stanley who is the father (or the uncle) of all alpha seeking quants.

QuoteDavid Shaw is pretty close.But we don't know precisely what David Shaw is doing, do we ?QuoteI have a question for you: Given that pairs trading has been known for 20+ years, why do you think that relatively simple pairs strategies (of that sort that you would expect undergrads working on a homework assignment to find) still work?I agree with this. Making money from pairs trading should involve a lot of work, as for any profitable strategy. Maybe with a model that deals with more observable quantities than just closing prices, like balance sheet figures ? Or a complex black-box that models lags and non-linearity ?

QuoteI agree with this. Making money from pairs trading should involve a lot of work, as for any profitable strategy. Sure it does.QuoteOr a complex black-box that models lags and non-linearity?All these "black-boxes" just improve the Sharpe ratio.In case of "pair phase transition" it looks like causality destruction.

Using only EOD data is certainly not going to get you anywhere.. all the guys I know making money consistently with blackboxes are very high frequency and highly quantitative.QuoteOriginally posted by: ShermanQuoteDavid Shaw is pretty close.But we don't know precisely what David Shaw is doing, do we ?QuoteI have a question for you: Given that pairs trading has been known for 20+ years, why do you think that relatively simple pairs strategies (of that sort that you would expect undergrads working on a homework assignment to find) still work?I agree with this. Making money from pairs trading should involve a lot of work, as for any profitable strategy. Maybe with a model that deals with more observable quantities than just closing prices, like balance sheet figures ? Or a complex black-box that models lags and non-linearity ?