Anyone want to collaborate on doing research on Industry specific and/or sector specific ranking systems?

So I have finished my ranking system for the stock universe and I am fairly happy with it so far. We aren’t as concerned with creating a trading system, but more so to create a ranking system that acts like a screener. To use the highest ranked stocks and then do traditional qualitative equity research to improve alpha even further. My system works pretty well across a lot of industries and sectors but I know that a lot of my factors work well with some industries, but not others. For example, FCF doesn’t really make sense to use on banks, or some growth metrics make more sense on certain industries etc. So my plan is to develop ranking systems for specific industries/sectors, to make the screening process even better. This will obviously be more time consuming as I have to test all of the hundreds of factors over again for each industry. I was wondering if anyone wanted to collaborate on this and share what factors work well in industries/sectors, and which don’t. In addition, someone to bounce ideas off of and discuss whether the factors make theoretical sense.

My whole theory on using the backtesting and ranking systems is to look for consistent slope across the buckets, have the system maintain that slope across time, sectors, industries, and size. I really don’t care if the top bucket is 10% or 50% annualized, because all I really want out of this is 3%-5% extra alpha when picking stocks. My time horizon on stocks is 6mo-5 years, so I don’t care much for factors that perform well on very short rebalancing periods. I want the system to tell me as much information about when to sell a stock, as its does about which stocks to buy.

I also want to have some of you guys take a look at my system and critique it, without releasing it to the whole world.


Screen Shot 2016-03-19 at 9.51.23 AM.png

Hi,
I too had thought about trying to create more Sector/Industry specific systems as well but ran into these issues:

  1. From my experience, the GICS codes don’t always bucket companies very well by Industry or even Sector. So companies that are truly in the same industry and competing with each other, for comp analysis, does not work well. For example, a common analysis is to see if a company is gaining market share on its competitors. But if the companies are not in the correct industry, that will be erroneous.
  2. My guess is that after you do Sector/Industry optimization, you will want to have some kind of rotation strategy so you are ‘usually’ in the right Sector/Industry at the right time. Timing this is hard and you can be wrong footed at the wrong time.
  3. Using the information in the three financial statements to come up with industry specific measures (like sales per store in retail), can be either very difficult or impossible. Since analysts use this data and they influence where investments occur, then if you get this wrong, it may hurt you. (I don’t even attempt to derive this information in my models). Getting this data is usually part of the qualitative analysis. So yes, coming up with a broad ranking system that does a top level sort on good candidates and then doing more qualitative analysis is a good idea. Trying to use the backtester for this will be difficult however because you are not modelling your industry-specific metrics. I just do generic, non Sector/Industry specific models.
  4. If you have not already, I would recommend going through Marc Gerstein’s recent tutorials on how to build models, from a top down perspective. I found it enlightening and useful.
  5. It sounds like you do know the basics about how the finance sector can vary from service/operating companies in terms of potential predictive factors/formulas. From my experience, if you create a model for service/operating company specific concepts, and match that with a universe which has only service/operating companies in it, then that is a good generic start. My work and academic experience has been with service/operating companies (technology) and not Financial sector companies so I am reticent about including Financial sector companies in my models. I do know that ROE and even ROA along with P/B can be used with Financial sector companies but as you say, free cash flow can be misleading.
  6. As a start, I would run your system as a live port and paper trade (no real money) it for a year by matching the system with individual universes for the 9 sectors (all but Finance) and see how each does out of sample. Do this in parallel with your Sector/industry specific work. Running OOS for at least 6 months is enlightening. I would use the State Street sector ETFs as my benchmarks.

My 2 cents. Good luck.

  1. Yea it’s crazy how messed up industry/sector classification is, even from the best data providers. Would creating custom universes for each industry introduce survivor-ship bias to the system?

  2. I’m not so concerned about a rotation strategy, I like to pick my sector weights on my own.

  3. Yes, it would be nice to have those industry specific metrics. I do think that even with more simple metrics, it helps to favor different factors for different sectors. For example, Financials and Industrials favor earnings based valuation factors and Materials favor cash flow based valuation factors (see attachment). My idea would be to have a base model that works on the universe and then tilt each sector towards the factors that work a little better for that sector.

If anyone is interested, that is from a pretty good analysis from JPM that can be found Here.

  1. Yes, I have read through his tutorials. I love the way Marc describes the process. Very well thought out and easy to understand.

  2. I’m curious as to how you define, or what you meant by “service/operating companies”. Did you mean just technology companies or non financial companies? Have you thought of any other ways to group up different types of companies? A couple of ideas: Companies that require inventory, companies with high capex requirements, companies with high sales expense, or # of employees relative to the company size. Just recently I was looking at factors for “TECH”, and realized that the factors that worked on the software and services subsector were completely different than the factors that worked for the more industrial-like tech subsectors (hardware and semis).

  3. That’s a good idea, I haven’t converted my sector simulations into live yet. I will do that.


Hi, My email is david.vornholt@icloud.com. If you want to chat further about this, let me know. We can do by email or via phone. I am in Hawaii so could be a large time difference.

The problem is that there is no objective way to do this because companies touch many different businesses (even seemingly single-industry companies) and change over time. Your best bets would be to (i) work with GICS classifications as is, (ii) packages of classifications like what was done with the Smart Alpha themes, or (iii) customizing on your own with Lists and being willing to tolerate some survivorship bias (which really isn’t necessarily a bad thing if you understand its there and if it came about as a tradeoff for some desirable thing, like the sort of custom group you can only create with a list; this kind of survivorship bias would seem benign, since it’s not associated with a deceptive choice, as for example, with dividend aristocrat lists that systemically eliminates bad outcomes - dividend cuts).