Now one mistake that happens often with Asset Allocation, Asset Rotation and Pair switching models is ignoring the dynamic nature of correlations between the assets in the portfolio. I think often one makes this mistake in one of two ways -
The approach this study takes is to utilize Pearson correlation of assets (range -1 to 1) for the last 13 weeks and then enter into positions only when this dynamic correlation between pairs is below a threshold. The correlation threshold used here is 0.25.
- One considers two assets as having low correlation just because they have different asset names and appear conceptually belonging to two different segments. Examples: US & International equities; Large Cap & Small Cap. Now these assets do have low correlation some times but correlation numbers in recent years tell a different story from popular wisdom.
- Another way one can make a mistake is assuming correlations are static and will continue to be that way. Actually many probably many not even investigate correlations and just accept popular knowledge as truth. Example: Stocks & Bonds. One has to just look at Spain to see how correlated they are (i.e., both stocks and bonds going down together) recently. Similarly for US, in recent years the Stocks & Bonds are less correlated (due to Risk On-Risk Off game?). But I think in 90s Stocks and Bonds had high correlation for some periods.
The approach this study takes is to utilize Pearson correlation of assets (range -1 to 1) for the last 13 weeks and then enter into positions only when this dynamic correlation between pairs is below a threshold. The correlation threshold used here is 0.25.
System Rules:
- Each week, check if index performance over last 13 weeks is better than the bond performance (TLT) AND the correlation between the index & bonds for last 13 weeks is below 0.25. If that is the case then sell the bond position (if any) and buy the index.
- Do similar for bonds i.e., if bonds performance for last 13 weeks is better than index performance and their correlation is below 0.25 then sell the index position (if any) and buy the bond.
- Now if only the performance part meets the criteria but pairs have high correlation for last 13 weeks then sell any existing positions and go to cash till correlations come below threshold.
Now, applying the dynamic correlation filter would reduce slightly the system performance compared to not applying the filter. On other hand, this helps one not require to know about how assets will correlate with each other in future. I think any method that requires one to figure future often leads to lot of complexity in the methodology, requires lot of efforts from trader/investor and also provides a false sense of security.
Stats & Graphs:
Additional notes are on the graphs itself.
Caveats
Results are cumulative. They are friction less (i.e., no commission, no slippage). Test duration: 2002-Current. The graphs are based on closed equity. The parameters are not optimized. My intent was more to share the concept of using dynamic correlations as a filter and get feedback/new ideas from the readers.
Questions to readers
I am sure there are many other ways one can compute correlation between asset pairs. Similarly there are multiple ways one can choose assets/construct model to take advantage of these dynamic correlations. I don't know that much about statistics and portfolio theory. So I am interested in your thoughts/suggestions on
- What would be a better correlation measure between asset pairs? Or how do you do it currently?
- In what other ways one can use this knowledge of dynamic correlations in trading/portfolio construction besides the approach taken by above study?
- Do you find the studies interesting/useful? Any changes/studies you would like to see?
Wish you all good health and good trading!
Note: This post (or for that matter any information on this blog) is not an advice. In trading, one can lose more than they think. So please do your own due diligence.
Note: This post (or for that matter any information on this blog) is not an advice. In trading, one can lose more than they think. So please do your own due diligence.
2 comments:
Insightful post! Correlation between assets are ever changing and past correlation doesn't provide a way of predicting the future correlation. There is quite a bit of similarity between pairs trading (stat arb) and pairs switching, in terms of possibly ways of managing the risk. In pairs trading they use co-integration as oppose to correlation; in pair switching why not try to minimize co-integration between two assets?
I am also very excited to the idea that pair switching has such low correlation to benchmark. Great things can happen when you combining multiple uncorrelated return streams. You've just added a big chunk of to-do things on my research list!
Hi Michael,
Thanks for your insights. I will check out what co-integration means and see if I can construct a study. My background is more of a discretionary price action trader dabbling in Quant concepts. Regards
Post a Comment