Please delete this post.

Pravin, thanks for sharing! Is there any reason why you're not using all of the sectors?

Hey Pravin,

Looks interesting. Can you explain the main thesis behind this algorithm a little more? I unfortunately don't currently have the time to infer it from the code. What are you trying to do and how is the regression helping you get there?

Please delete this post.

Please delete this post.

I thought this was quite an interesting algo so I removed some duplicate stuff and created a version where you can use sector ids.

I'll go through the paper and implementation later and try to isolate what drives the performance.

Here is backtest with all sectors with improved sharpe and reduced drawdown.

Awesome Mikko.

Hello Pravin,

A few thoughts here:

• This algorithm could likely be far more elegantly expressed using our pipeline API. Let me know if it's not clear how.
  
• Evaluating quality of on-the-fly regressions such as the ones you're doing is difficult. One problem can be if you don't use enough of a historical window to compute the betas, your betas can vary a ton day over day as they won't converge. If your residuals are stationary that's a good sign, as it indicates that you're explaining the non-stationarity in prices with equivalent non-stationarity in factors. One thing you shouldn't do is trust any estimates of p-values or parameter standard error that come out of these regressions, as those are very wrong on non-stationary data.
  
• In this case you are using regression more as part of a forecast. You probably want to validate the overall model rather than worrying about validating each regression you run. If you think that the regressions are capturing important dependencies between factor and price values, I would think about ways to validate that overall hypothesis rather than worrying about each regression you're running.
  
• What you are constructing is effectively a multi-factor alpha model with dynamic alpha combination. The question is whether there is alpha left over. I might try to use pipeline for this and then feed it into alphalens to get a better sense.
  
• I ran a tearsheet on your backtest and it seems like it makes a good chunk of its returns during 2008-09. This worries me because usually that's an indication that your strategy is overfit (pricing model fit to N=1 event). I would run it post 2010 and see how it looks during your analysis.
  
• At the end of the day consistency is the best check. After you've done a few more analyses, stop developing this strategy and let it accumulate 3 months out of sample while you work on other stuff. Then come back to it and use pyfolio to run an out of sample tearsheet and ensure that the behavior and distribution of returns match historically and in your new window.

Hope this helps, it's difficult for me to say more without reading the paper, which I unfortunately don't have time to do right now. Overall this is a very interesting strategy and there are known examples of being able to trade on noise left over once you've removed principle components from the market.

Thanks,
Delaney

Thanks Delaney for the very useful feedback. I will comb through it and try to incorporate your feedback into the analysis.

I ported the algorithm to the new "Q1500" and tried it using pipeline. Two issues:

1. It is painfully slow. 3 hours and backtest has not completed.
   
2. All profits are eaten up transaction costs. Don't know what is so different about this.
   

Try setting the universe to the intersection of the top 1500 largest by market cap with the top 1500 by dollar volume. That should get rid of most of the illiquid stuff, and be fairly quick to run.

I've updated the algo to use pipeline with Q1500US and optimize but there is a problem. For some reason its failing when I use optimize but it doesn't fail on the same stock when using the original ordering. The stock optimize is failing on is DVMT, the day after Dell switched over.

To see the algo work fine on the original ordering, change line 75 to:
useOptimize = True

Any idea whats going on or how to fix it?