Forums  > Trading  > Pairs Trading - Selection Methodology?  
Page 1 of 1
Display using:  


Total Posts: 3
Joined: Jun 2017
Posted: 2017-06-08 09:52
Hey guys,

I know that pairs trading is a beaten topic that was researched far and wide, and returns are not what they used to be, however I would like to fish for some ideas on how to select the pairs beyond simple ADF statistics or Hurst exponent values.

I've been running some tests with mixed results and basically came to the conclusion that I need to minimize the number of pairs which have great in-sample stats but quickly diverge once tested out-of-sample and then asses every execution myself. So far I've tried filtering pairs based on their industry, ADF t-stat, Hurst exponent (if less than 0.4). There are some filters that I'm still working on like: difference in log(market cap), difference in P/E values, difference in variance of log prices, Kolmogorov-Smirnov test to check if samples are pulled from same distribution.

What can you suggest to tackle this specific problem of selecting the most stable pair that will not diverge, perhaps there are some statistical tests or bucketing methods to make the selection more robust? Should I move to more high frequency data?

Founding Member
Nonius Unbound
Total Posts: 12737
Joined: Mar 2004
Posted: 2017-06-08 15:04
I've found that ADf is kind of bullshit in practise.

from my experience stationarity is just an academic construction and in real life things aren't stationarity. the exceptions to this are, for example, cross exchange arb, or triangular arb (when you take two legs of the triangle and convert the cross to your base currency). not much else comes to mind. I think whats more important is handling the shift out of that quasi stationary state.

Chiral is Tyler Durden


Total Posts: 3
Joined: Jun 2017
Posted: 2017-06-08 17:19
Yes, I try to account for the unstable stationarity via using rolling mean of the spread and stdev's, it help to some extent... however when the spread suddenly start to trend this kills much of the returns, thus I figured the need for additional filters. Also one idea was to measure the slope of the spread's moving average for some window and trade only if that slope is less than some threshold, which should help to avoid the trending phases.


Total Posts: 336
Joined: Jan 2015
Posted: 2017-06-09 00:06
+1 @nonius

The problem doesn't stem from selecting the right pairs. The problem stems from pairs-trading being obsolete in DM equities at anything longer than an HFT horizon. To the extent that it used to work, it did so because stat-arb traders were significantly less sophisticated.

If Ford moves, low-latency algos will slam GM in about 1 millisecond. And anyone holding longer-term risk isn't trading GM against Ford, they're trading GM against S&P 500, the industrials sector, the auto industry, the US dollar and a whole galaxy of multi-colored exotic style factors. Plus they're conditioning on news, earnings, analyst activity, order flow, and microstructure.

It doesn't matter if you can pick the optimally best twin every single time. A single co-pairing will always have way less information. It's like you're trying to win the Tour de France by making the most nutritious trail mix, when all your competitors are on HGH and EPO.

Good questions outrank easy answers. -Paul Samuelson


Total Posts: 361
Joined: May 2006
Posted: 2017-06-09 00:06
@arti, you might be missing the point.

If you are trying to train a blind robot on blind data you don't understand, that is pretty much what you should expect - overfitting in sample, with zero ability to predict anything out of sample.

"There is a SIX am?" -- Arthur
Previous Thread :: Next Thread 
Page 1 of 1