this is a thread on one type of mean-reversion strategy, http://elitetrader.com/vb/showthread.php?s=&threadid=134253 basically, it is any strategy where you are playing the probability that the spread between two correlated/cointegrated financial intruments that has diverged will eventually converge back to the mean. your profit is the spread's convergence, i.e. the short comes in more than the long goes down or the long goes up more than the short (as a % of the ratio of the two).
Sounds too risky to me. No such claims can be made solely on mathematical grounds. For example, you have two companies in the same sector, say energy exploration, which are correlated but one of them got contracts for exploring new fields for natural gas. On what basis do you think the spread has to converge back to the mean? One company has future potential and the other is in a stagnant state. IMO, math is not enough for such claims to be made and this is also the nature of the problem the caused the financial crisis. Math cortelation is a frozen picture of the past. Tells nothing about the future. I kept telling that to a guy I know who is now down 50% with his hedge fund. He never listened, as a matter of fact he played too smart. Now he is sorry.
You would take that into consideration before putting on the trade. If there is no fundamental reason for the two stocks to have diverge then maybe you would want to put on the trade. And one thing people have to realize with stat arb, its not all about correlation its about cointegration.
good point. the higher the correlation, the less opportunity to suck money out of the spread and the more skill required (or algorithmic arbitrage programs), as with a stock and its preferred which can be like 99+% correlation and all you can hope for is pennies. however, they should be correlated at least 60-70% as would be many pairs in the same industry, a stock and the sector ETF it is in or even an oil stock and USO. the cointegration is where your probability and opportunity come together, betting on the likelihood that the divergence will come back in. IntradayBill, i was just answering the guy's question. but yes, it is based on what the relationship of the financial vehicles you are using has been in the past but if the trader practices due diligence and know that there is no fundamental reason for the divergence(like your NG exploration contracts), then most likely, it will come back in. any play you make has to be based somewhat on the past, be it yesterday's guidance that crushed estimates or the resultant spike that may be due for a retracement or even the last 5 minutes' momentum. on stat.arb., based on how many times a pair has gone out and come back in, you can determine the mathematical probability of further cointegration of the two. as for your hedge fund buddy and stat.arb./pairtrading, in general, there are no guarantees. but then, in this game we play, there are NO guarantees for profit. anyone who says otherwise is spewing scat and/or selling something.
You still got problems. Cointegration implies correletion. In longer term trading, value investing is not only far more superior than this method but I should say that this method has high probability of failure. In short time frames you still have the problem of choosing a time period. It turns out that mathematically, at least, things may look different when you take into consideration 10,000 data points and much different when you take 20,000 data points. I thus wonder why people even discuss such methods.
i think this stat arb thing is interesting, if you buy somethin you gotta sell somethin . . . i'm not familiar with the mechanics, but the concept is interesting. i think i'm going to buy chan's book on algo trading to see what he says about this for now i'm trying to find a shortcut to pairs trading, by using various visualizations like: marketrac.nyse.com/mt/index.html the other one is finviz, its market maps based on the finviz visualization alone (not looking at anything else, like fundamental, earnings), i've made the following paper trade today: XLK long (worth $5K) technology etf TXN short (worth $5K) texas instruments i think if a stock has a daily spike of more than, say, 1.5-2 daily standard deviations, it may be a good trigger to short it against its peer group. the point is, i think, it has to be way out of balance vs its peer group, not just a 2-3% spike on some news. plus it shouldn't be a penny stock, otherwise the percentages don't mean much, probably. that's why i think visualizations are a good starting point for this. you can literally visually see, by eyeballing this, if there's any stock that's out of balance. this is just a guess, so i'm only beggining to make look at it now
dude, what is your data based upon that makes you say the method has a high probability of failure? from your touting "value investing" as superior to stat arb, it is clear that you have not looked into the rudiments of this method because one of the primary premises of some types of stat arb are to short the overVALUED and go long the underVALUED, in addition to the technical correlation and cointegration. and what are you trying to say about time frames and data points? i am easily confused so please explain. my guess as to why people discuss such methods is...now i am going to go out on a limb...but i think it might actually work. just a guess
i love the Marketrac image, if i could actually figure out a way to use it to help me, i would be stoked. on your paper pair, what do you have against TXN? i mean, why don't you short an unprofitable one like STM or a larger cap like INTC(slower to move if the market rips)? maybe a basket of all three against the XLK if you are down on the semiconductors, in particular? or maybe throw in a telecom that is bleeding out like S? good luck with the funky visualizations. never heard of using anything but a ratio balanced chart for such but hey, new technology is coming out all the time
Dude, you're asking good questions.... To start with, every calculation you make between time series depends on the time interval and number of data points involved. Let us not make this more complicated than it actually is. This type of trading is based on unfounded assumptions that can be proven fatal. You determine that a given time series (like the spread between two securities) is stationary but that is based (A) on a certian time interval and (2) NOTHING can assure you that the series will remain stationary. I agree with Taleb here that randomness is fooling you for a while and then a small black swab takes it all away from you. This is how markets treat wise asses. Every one of them has blown up.
Good reference. Mainly the step most people calculate is calculating how far away the mean is from the mean on a normalized basis. You take a difference of a difference and divide by the standard deviation to normalize your results. The process really refers to first differencing a time series, and is what my system at www.collective2.com/go/pairsqidqld is based on. Really ingenius, but the research that went into it dates back to the mid 80's when quants discovered pairs trading. They weren't fortunate enough to live in a day where ETF's could serve as eternal pairs and so didn't have to bother themselves with "finding" the pairs. We have perfectly negatively correlated pairs with 2 x leverage and 2 x leverage inverse ETF's. All good stuff. Stat arb produces results looking like a regular linear regression and mostly you'll find things are perfectly priced with very few mispricings. Incidentally, this is usually done in STATA, SAS, or Excel. One of my most recent ones took the log of the market cap and regressed it onto the log of revenue and about 27 other variables with results that looks like this with a high R^2 of 0.88 Source | SS df MS Number of obs = 500 -------------+------------------------------ F( 28, 471) = 133.61 Model | 495.498531 28 17.6963761 Prob > F = 0.0000 Residual | 62.3830638 471 .132448118 R-squared = 0.8882 -------------+------------------------------ Adj R-squared = 0.8815 Total | 557.881595 499 1.11799919 Root MSE = .36393 ------------------------------------------------------------------------------ logofmarke~p | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- logofrevenue | .8916765 .0154872 57.58 0.000 .8612439 .9221091 var4 | -.2091459 .0369259 -5.66 0.000 -.2817059 -.136586 var5 | -.0752378 .0314276 -2.39 0.017 -.1369935 -.0134821 var6 | .005468 .0396742 0.14 0.890 -.0724924 .0834284 var7 | .561904 3.969468 0.14 0.887 -7.238154 8.361962 var8 | .0037096 .0089385 0.42 0.678 -.0138546 .0212739 var9 | 2.968399 .633364 4.69 0.000 1.72383 4.212968 var10 | .012367 .0089514 1.38 0.168 -.0052227 .0299567 var11 | -.2553104 .3799248 -0.67 0.502 -1.001868 .4912469 var12 | .3935184 .1404802 2.80 0.005 .1174729 .6695639 var13 | -.8261104 .1359956 -6.07 0.000 -1.093344 -.5588772 var14 | -27.7728 2.798556 -9.92 0.000 -33.272 -22.2736 var16 | -.4961916 .1187044 -4.18 0.000 -.7294474 -.2629358 var17 | .3137915 .0928498 3.38 0.001 .1313404 .4962427 var18 | .440714 .1671892 2.64 0.009 .1121851 .769243 var19 | .7017086 3.978552 0.18 0.860 -7.116199 8.519616 var20 | -1.297751 .3029049 -4.28 0.000 -1.892963 -.7025383 var21 | 1.136644 .2618217 4.34 0.000 .6221605 1.651127 var22 | -3.173427 .92185 -3.44 0.001 -4.984874 -1.361979 var23 | (dropped) var24 | .0009775 .0061591 0.16 0.874 -.0111252 .0130802 var25 | .0974009 .0497012 1.96 0.051 -.0002626 .1950643 var26 | -.0144898 .0814906 -0.18 0.859 -.1746199 .1456403 var27 | -.1635945 .0436717 -3.75 0.000 -.24941 -.077779 var28 | .0018706 .0014 1.34 0.182 -.0008804 .0046217 var29 | .0029186 .0008508 3.43 0.001 .0012468 .0045904 var30 | .2466865 .0140288 17.58 0.000 .2191198 .2742533 var31 | -.1140925 .120899 -0.94 0.346 -.3516607 .1234757 var32 | .1789512 .1083369 1.65 0.099 -.0339323 .3918346 _cons | 1.956292 .3624543 5.40 0.000 1.244065 2.66852 ------------------------------------------------------------------------------