Filtering Candidate Pairs for Mean-Reversion Trading
Summary
The document discusses narrowing a large universe of candidate equity pairs after an Augmented Dickey–Fuller test and half-life calculation leave many candidates. Suggested filters include liquidity, trading costs, backtesting with cross-validation, and statistical significance. These checks address whether a pair can be traded economically and whether its apparent behavior is supported by evidence beyond the initial screen.
Other suggestions focus on economic and structural relationships. Grouping assets by sector or other characteristics, then examining correlation within those groups, can reduce the search space before applying cointegration tests. Correlation is presented as a screening step, while cointegration is used for final pair selection. The discussion is a list of starting points rather than a validated ranking method; it provides no comparative results or rules for choosing thresholds. A low estimated half-life or a passed stationarity test alone does not establish that a pair is robust, liquid, or profitable after costs.
Key ideas
- Liquidity and the cost of entering trades can help screen candidate pairs.
- Backtesting with cross-validation can assess whether pair behavior persists beyond the selection sample.
- Sector, market capitalization, and trading volume can guide economically meaningful grouping.
- Correlation may narrow candidates before cointegration testing, but does not by itself establish a tradable relationship.
- Statistical tests and estimated half-life should be considered alongside execution costs and robustness.
Tags
Full text
# what are the criteria to select pairs? # what are the criteria to select pairs? I'm new to this forum, this is the first question I posted. I have many candidate pairs and I've used ADF test to make a first selection. There are more than 800 selected. The pairs are absolutely too many. I'm thinking of other criteria to eliminate some of them. I've calculated the half-life and I want keep those who have low half-life, but all of them have a half-life less than 30 days(since all of them have passed the ADF test). Are there any other criteria which could select paris who have good mean reversion property? Thanks in advance. ## Answer by meh (score 2) https://quant.stackexchange.com/a/19356 In no particular order here are some ideas to get you started. - Liquidity (ADV, # of shares, etc) - Cost Basis (Cost to put on a trade) - Back test / Cross validation - P-values ## Answer by Ashish Garg (score 0) https://quant.stackexchange.com/a/43317 Cointegration as tested by ADF and other tests was developed to test the economic dependence of a time series on another. So, it would help you to think on similar lines. Bucketing your time series which can be economically dependent on each other would certainly help you eliminate more pairs. ## Answer by vibhu_singh (score 0) https://quant.stackexchange.com/a/45119 If you have too many stocks in your bucket. The first step to filter the stocks would be to do a qualitative analysis of the stocks. You can segregate the stocks on the basis of a sector, market cap, daily traded volume, etc. Then check for correlation between the stocks in each segregated groups. After running correlation you will be left with less number of stocks. On the remaining stocks, you can run the cointegration test. Finally, select the pairs which are cointegrated.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.