Have a look here, best site I know for pair trading: https://www.pairtradinglab.com/

Which market are you looking at? I am interested in this topic and would be willing to share thoughts and ideas, we can connect offline.

I pairs trade stocks, mid-frequency, and initiate and close each trade with market orders. It's a lot of edge to give up, but I'm doing pretty well at it. I look at different things for trade setup than your classic pairs trader. And I look at different things for the pairs themselves.

http://www.tradestation.com/education/labs/analysis-concepts/market-neutral-pairs-trading

Pairs trading is pretty common in currency futures pairs. I know a futures trader who implementa a strategy using pairs I can introduce you to.

here is some good info that should be helpful:

http://www.tradestation.com/education/labs/analysis-concepts/market-neutral-pairs-trading

Mean reversion is a fascinating subject. Sadly, identifying a useful application was elusive if not completely futile. The deal breaker was always the same; determining and measuring the time slice. As everyone knows means move. Mean reversion may occur but may happen at levels unkind to the capital account. Microsoft offers a decent example at this time. The stock dropped 10% during late session trading January 26 and early trade January 27. Means in all time slices change as a result with the shorter slice lines showing the most movement. The sharp decline drops price well below a few mean lines. Here is the problem. The mean direction starts moving south. Price can revert to the mean but that contact may happen at a price no better or below the entry as the mean slides toward price during the declining price period. It seemed that identifying movements into the distribution tails was more useful and is arguably a derivative of regression analysis. However, trading the tail was only justifiable with support from supplementary data. Basing the trade simply on tail proximity was no better than good old seat of the pants floor trading.

if you can't get a strategy to work within 1 or 2 months (even just a rapid prototype to test the concept is profitable) then you should abandon it and try something else.

there was another guy who posted in one of the groups here he hired out a big development team to build machine learning automated strategies i guess because he assumed these must be very profitable but after 2 years or futile results he decided to fire the whole development team and was thinking to start again from scratch!

We could team up and put brains together with some development on https://www.quantopian.com/ smart guys and lot of good ideas there, but have not seen any medium-term pairtrading approach there so far.. let me know if interested either here or as PM.

thanks to all for the messages...

I think that at this point I'm leaving this kind of strategy on hold. Didn't get acceptable results with pair trading.

Compared to other (very)basic strategies, pair trading is still under-performing (at least what i have so far)

So, I'm perusing new strategies for now ;)

Thanks!

@Ariel,

I think that's a wise choice, because it's a crowded field with no really new insights. If you're approaching pairs trading the same way that 99.999999% of traders do, you will be disappointed in your results. There are other other profitable approaches to pairs trading, but it requires a paradigm shift in your foundational understanding of pairs.

Just saw this thread. We have been doing stat arb trading for a number of years, which is a generalization of pairs trading. It has been getting much more difficult as the field gets more crowded.

The one pointer I can give you is to not construct your pairs using correlations of historical data. This is just data mining and will lead nowhere. The pairs must have an economic underpinning and you must build your model using those economic relationships.

Good luck.

@Greg, I was looking for that kind of insights !

Thank you for sharing!

Comparing to S&P is... interesting. Unless you're taking on similar levels of risk, comparing rewards is pretty useless, IMO. I personally like comparing my strategies more to bonds (of course, I don't completely), for the simple fact that algo strategies should primarily be low risk, and then focus on rewards. If your strategy is too volatile, then it's too "chancy" and not "systematic". Again, these are simply my opinions. I'm sure others will disagree with me.

how about intraday cointegration with triples of stocks? Anyone tried?

starting to backtest in matlab