i'd like to start a thread to meet others interested in the topic. a few questions to start: can anyone explain in english, what are the common approaches to multifactor stat-arb? i'm running into limitations on testing in excel 100 stocks, 10 years daily data, anyone know of any reasonable platforms to test on? matlab? last year was a horrible year for most pairs models. this year, seems to be pretty rough on stat-arb as well. i've heard of some funds giving back their year on the march rally. any thoughts?
decimilisation was the first big hit for the stat. arbs.. many models do not work anymore. pairs trading lost a lot of edge after 2000. last year slightly up was the best for most funds. this year in pairs trading was tough. low dispersion is key problem. I think most people operate on prop. software. peace
Excel should be fine for backtesting. I have 250 stocks with multiyear end-of-day closing data. I don-t regress every stock against each other, but only wthin industry groups, that saves a lot of processing power. Not sure why some people would want to trade MO and IBM (example) as a statistical pair, based on historical correlation. I think some common sense need to be included in a statistical pair arbitrage model. Oliver
MO and IBM have just about 10% of historical correlation. There is no reason to trade that pair (speaking of common sense). However, if you'd test all the stocks against each other and check the results you'd be surprised of what you'd see. There is no correlation btw having stocks in the same industry group and the profit you'd get. And common sense also suggests that money go from one sector to another - just one reason why pairs from different industry sectors could be profitable. I'm talking not about day trading though.
Statistical arbitrage is very, very quantitative ... we're talking about cointegration, integrated I(2) into I(0), nonlinear combination, ARIMA, EGARCH, etc.. It's a tough tough arena. Using simple technical analysis will put you behind the eight ball.
Satyrican, Tantalizing post. Do you do stat arb? Either way, can you elaborate more? I am not a quant, but stat arb has always fascinated me -- I love the idea of being mkt neutral over a well diversified portfolio. From my (very crude) testing on the subject, I have concluded that one cannot just simply play regression or diversion on a purely mechanical basis (as you hinted, profitable undertakings of stat arb are very complex). If regression or diversion is being played, there has to be some discretion involved (which defeats my purposes). Also, in my humble opinion, it seems that people are too eager to play spurious correlations between fundamentally unrelated pairs...which may have worked well historically, but opens one up to fatter than fat tails in the future. Additionally, people are under the misconception that high correlation between stocks = large opportunity for profit. I would be interested in hearing more about your experiences... Cheers, MYD
The math is actually pretty straight forward. The problem is, unless you are able to do many pairs at once (the "law" of large numbers) the math doesn't really matter much... I have started to get back into this after maturing as a trader and armed with some fresh ideas...we'll see how it goes. nitro
Kinda tough to explain statistical arbitrage .... the math is somewhat rigorous. The "stats arb" that traders used on this forum is not really stats arb. Technical analysis on pairs is not statiscal arb. Moreover, correlation is a horrible, horrible measurement... You hit the hammer right on the nail. Spurious correlation WILL give you a higher correlation factor. Cointegration is more appropriate. I used to run an arb desk via a hedge fund. I currently dont trade stats arb as the outlook is not vary favorable. Momentum usually kills stats arb. High volatility is king for stats arb.