Factor Effectiveness over Time

Research
1

I was playing around with an idea about why and when the factors that used to work seemed to have stopped and of course I wanted to test just how true that was. There has been a fair bit of talk about this over the last few months. Mostly with respect to regime change and value is lost. There is some truth to this from what I see.

So here’s what this crude graph shows:

The data points (not sp500) are the SLOPE of the ranking for a single factor with default direction (lower is better in this case). The test used the SP500 universe with a 1yr period. A positive slope (above 0) means the histogram has a positive relationship to the factor over the past 12mths. In that case it also tells me that investors are VALUE seeking.

Therefore a negative slope (below 0) means that the histogram has a negative relationship with the factor over the past 12mths. In that case it tells me that investors are GROWTH seeking. As I’m typing this out it makes me think I should try this with P/E.

  • SP500 of course
  • 1yr period, 6mth rolling pr2SalesTTM
  • 1yr period, 6mth rolling pr2CashFlowTTM

There are a few things that I get out of this graph and maybe some of you can see more or possibly this is meaningless.

  1. The obvious one is that the linear regression line is a negative slope with an X intercept at the beginning of 2014. The date of the X intercept is somewhat arbitrary due to being start date dependent but approximate year is good enough for me. So these factors are currently on a path that puts growth over value and maybe we should think about flipping the direction to higher is better in the future??
  2. There seems to be some mild correlation between the run up before a correction. When growth is heavily favored over value, let’s call that a negative data point <-1. It seems to indicate a previous year of exuberance only to follow with some kind of correction. The exception to this rule is RIGHT NOW. So I’m not sure how to interpret that at the moment. Correction coming soon perhaps?
  3. During a correction investors seem to seek VALUE again. A positive data point doesn’t necessarily mean that the factor is profitable however it does show that value is a safer place to bet than growth in down markets.

2

Barn,

Have you considered the possibility that “value” outperforms after corrections because stocks which were previously not value stocks become value stocks until they are bought back up? Buying value after a correction means buying companies that are undervalued from the correction but still have good growth and risk prospects. The market basically overreacted to them. Buying pure “value” any other time without attention to growth vs risk prospects is just buying “Value traps” which are underpriced because they are companies with poor growth and risk. Another scenario is that after a recession market dynamics change improving the growth and risk prospects of previously unproductive “Value” stocks or readjusting expectations downward of “growth” stocks. Marc Gerstein talks about that in model design strategy.

Jeff

3

Price = Earnings / (Risk - Growth)

Stocks with combined highest earnings, highest growth, and lowest risk will being priced the highest by the market.

The characterization of stocks as being either “growth” or “value” is an oversimplification or stock categorization. Higher P/E or P/FCF companies can be better value than “value” stocks if the future projected value of that growth stock is not currently priced in.

Personally I also think seeking value in unloved industries is more likely to underperform as well. It’s hard for a “value” stock in a sector or industry with poor risk and growth trends to buck that. Think financials and energy. There’s a lot of “value” stocks in those sectors but the growth prospects aren’t very good at the moment. Energy companies are in riskier positions than they were in the 2000s now that energy prices are lower and the drive to renewables. With low interest rates and flatter yield curved financials are in less desirable positions as well.

4

Thanks Jeff and as much as I agree, it really wasn’t the point I was trying to make with this study. How about we ignore the labels of value and growth, since that debate is just semantics. What I wanted to show was that when we are making a ranking system and use a long period during testing, that we are likely ignoring the in and out of favor periods. As long as the tested slope is positive (above 0) then we put that in the keep or at least “maybe” pile. The long test duration blinds us from the periods of out of favor. This is why creating a ranking system using multiple periods is so important. Another point is that it may be wise to make 2 ranking systems for a trading system. 1 that is for market down turn when investors are seeing safer bets and one for up markets where investors are willing to pay higher premiums. In the example you would want to reverse the direction of the rank or use neutral ranking and either pick from the top or pick from the bottom.

Another point was that this example may also work as a type of market timer. When high premiums have been in vogue there tends to be market turbulence soon after. With the exception of right now, which I have no answer for other than to keep this pattern there would need to be a market correction soon.

5

Barn,

Reversing rankings based on market conditions sounds like a dangerous strategy. I personally stay away from market timing. Diversification of strategies is much easier to execute. Most market timing models fail.

Jeff

6

As evidence in the DM’s most models fail anyhow. As my example shows, the effectiveness of using certain factors in a ranking system have a tendency to be counter productive anyhow and you end up picking statistically out of favor stocks. I believe this to be one of the reasons models fail. High ranks for long term statistically proven factors to hold a positive slope is not a good measure for using them in a ranking system. I’m as guilty as anyone for doing this. Until this test I didn’t realize how rocky the ride was, nor the magnitude at different points in the market cycle. Again this was only 2 factors and only executed on the sp500 but I suspect there are much worse examples of factors and universes that some of us are guilty of grouping together.

As for the timing aspect. I agree that timing is hard but long term trending is something that many of us do. When the market is historically pumping up the valuations without the sales or cashflow then it’s only a matter of time before it needs to correct. Once the historical TTM reverses again then go back to growth ranking. Really it requires 2 separate ranking systems because you would normally have many factors and I can’t imaging they would all hold the same reversal characteristics.

Until I make a model to test this it’s a lot of pie in the sky theory but I thought it was a bit of an out of the box idea that I wanted to share. Right now I think we need some more out of the box ideas because a lot of the tools in the box just aren’t cutting it. BTW that’s no disrespect to P123, I think you guys have some of the best modelling tools available to retail investors on the internet.

7

Barn,

Most DM models fail because they are overfit or misfit. This occurs for the following reasons:

  1. Lack of underlying stock mis-pricing theory or poor underlying theory. Marc has been trying to beat it in our heads that we need to stop thinking about choosing the right factors or ratios and instead look at the big valuation picture. He’s right. It is the interplay of factors in market conditions that leads to stock price appreciation. You are trying to make a dynamic model to anticipate this but I question how well your approach will work. Many investors already employ factor rotation strategies. Factor investing can help achieve lower volatility returns or higher returns at times, but companies are complex entities that cannot be explained by a few valuation ratios or a style factor. So why do we delude ourselves in believing we can be outperform with those approaches? Marcs models do work. They might not grossly outperform or always outperform but they do work. We have to understand why we model. We are trying to automate the process of examine a company’s financials a market perception to pick a company that is a good investment. This is where we actually struggle. What makes a company a good investment?

  2. Over-optimization of variables to maximize returns and minimize drawdowns in backtesting. Most models can’t realistically achieve achieve 40-50% returns year after year without virtually no drawdowns. In a lot of cases this over-optimization is fitting to noise which won’t repeat and cause your model to fail horribly when you try to trade it. These models either don’t have good underlying financial theory for why they should work or are developed by excessive trial and error. If the basis of your model is trying every variable and tweaking each to see what works then you’ve violated my first point. Variables should only be specified which strengthen the theory of the model.

  3. Market timing. A lot of the designer models do market timing hedges. These worked well in backtesting. They worked horribly thereafter. Why? The future never is exactly as the past and investors watch out for reasons that lead to crashes or corrections in the past past. Notice how the only individuals who predict crashes at the right time are those who always predict them? The market timing fails for a more insidious reason though. It creates false signals which push you in and out of investments at the wrong times. That’s what really kills your returns rather than if you just accepted the downside risk.

  4. Lack of diversification. This has been said but maybe not enough. Just like it would be bad to put all your money in one stock it’s probably likely a bad idea to invest in one strategy. Every reasonable strategy will go through periods of underperformance. Poor strategies will eventually fail. Some good strategies will stop working. By diversifying strategies you can tolerate these outcomes in your portfolio. But you do have to make a good effort to select good strategies and not just the ones that appear to have performed the best.

As for the matter of time argument the market can stay irrational longer than you can stay insolvent. Ask pure value investors how long they have been waiting to make money on value investments. Market dynamics are always changing and evolving. Yes valuations are very frothy but they can continue indefinitely. You’d be hard pressed to accurately predict when this valuation reaches the tipping point because it will be different every time because it depends on so many variables. Ask the dot com investors how they knew when to get out of high P/S growth. When everybody heads for the exits they jam up.

Jeff

8

Barn has no disrespect for P123.

I have none for the forum.

But sometime between the development of a new idea, the discussion of the idea in the forum and the request for implementation of that idea…….your idea will be dead in a box and buried.

Never out of the box and implemented. Perhaps, with some in the forum being glad that they have, perhaps in some small way, discouraged the development of a new idea.

Who’s fault is this? The data providers. And for the record, I do not really care that some in the forum have never liked as single new idea, ever, except their own.

I used to be angry. If I get angry again it will be at the data providers for slowing the actual development of ideas. For making us pass every single new idea that involves significant amounts of data through the forum for consideration.

Marco, and I assume the rest of the P123 staff, are trying to remedy the situation. Looking at DataRobot and probably a whole host of great ideas. No promises that any of those ideas will pan out.

P123 makes the situation better. Gives us as much data as possible. As many downloads as possible. And a very good method of making models (using sims, ranks, rank performance etc).

And P123 is better than most sites. Probably better than all of the rest of the sites for retail investors. I agree with Andreas that Quantopian is not so good, for example. Zacks is not so good.

No complaints. Never should have been. But for most newer methods for developing models: Well, dead in the water if not already buried.

P123 just does not have the capacity to develop all of the good ideas at this time.

I think we will have to think mostly inside the box. At least a little while longer.

-Jim

9

Jim,

I have to disagree. More data is not necessarily better if the way we develop models is not sound or improving. If we just get more data people will hunt for more relationships in the data without sound theory to back it up. That will lead to worse models.

I would suggest an alternate approach for Barn. If you wish to find value that can outperform perhaps look at SectorInspector. He picks industries where he believes there are good growth prospects and then screens and backtests for value in those sectors. His strategies have been pretty successful. But he’s looking for value in growth industries with quality.

Jeff

10

Jeff,

I understand you know I would have no good use for more data.

Kind of makes my point I think.

Thank you for making my point so crystal clear, in fact.

And again, thank you P123 for the data and models that you are able to provide while working with the data providers.

I write this while Georg is working (in another thread) to get around the restrictions involving the listing of stocks in an ETF. He is able to get around it but he is good and he has to work hard.

-Jim

11

Jim,

Not sure how that makes your point clear. I’m confused now.

Jeff

12

How could you know enough to disagree with my simple desire for more data?

How could you have enough information to disagree with my belief that more data would be helpful for me.

If you were a psychic maybe you could know this. Also if you were a hacker and had put a key logger on my computer perhaps you could know this.

I am sure that you are not a hacker.

But I wonder how you know enough to disagree with me when I say I could use more data?

Do you think you are psychic or something?

More likely you think all of our decisions should be passed though the forum. You are happy to lead the discussion of what others need helping to determine how legitimate their needs might be.

That was my point and you seem to agree. You seem to agree with the point I was making.

I actually think you are not bad at that if we need an additional person to do that on the forum. I will vote for you if it becomes an official position.

-Jim

13

Jim,

I disagree because without a sound and rationale stock mispricing theory more data leads to more data mining. Data mining is precisely our problem. That’s why nutrition, medical, and psychology are in disarray. They fit models to data. Marc keeps saying this. He’s right. Now if you have a sound model with a strong reliance on that data then by all means. For example population growth projections in a utility market might offer growth information for that utility. Although I would suspect this information might already be captured in analyst sentiment.

Mind you that I am generalizing. I said more data isn’t necessarily good. I didn’t say it can’t be good or used to achieve better models but it can be dangerous if the framework behind a model is flimsy. Also no data point is perfect. Data is noisy which is another challenge we face.

I can withhold my comments in the future. It just seems to me like the answers are in front of us and have been the whole time. Now improvements can be made but realistically we are still going to face the same challenges. Creating good stock models IS hard and that won’t change. If it was easy everybody could do it and outperform.

Jeff

14

Jeff,

People have their own ideas.

Remember, I was one (of the few) who supported your—clearly correct—belief that random noise added to the data can be helpful at times.

I understand that P123 has priorities and I would actually agree that P123 might not want to make that feature a top priority. I would prefer that they didn’t make adding random noise to the data a top priority in fact (I know you were not asking for this).

But members of the forum were wrong to criticize your idea.

Just plain wrong to try to get into your head, figure out what you were thinking and be so sure about what might be helpful to you.

I like your posts. Even if you still disagree with what I said here. I like your posts and appreciate your ideas.

And I am still going to vote for you;-)

Best,

-Jim

15

Jim,

A lot of users complain about the designer models not working. Perhaps there needs to be a really thorough re-evaluation about how these models get vetted.

But that I believe is why you are suggesting ANN and I was suggesting noise injection and why we are talking about sample groups in backtesting.

Jeff

16

All good ideas.

Except, me being in charge of (or on a committee) vetting someone else’s models.

I do think adding noise, making ANN possible etc. are good thoughts

P123 just made what Georg does a little easier (a good thing). P123 is accomplishing a lot at a good pace. Especially, considering their arrangement with the data providers.

The only strong belief that I have brought to the table is that maybe lots of feature requests could be solved at once with Python.

I need to (and have) backed off from that. I will leave that to the programers, business people, negotiators etc at P123. There is more to that decision than I first thought. I apologize for pushing too hard on that.

Do count be out on vetting other people’s Designer Models, however.

-Jim

17

Jeff,

I have to agree with Jim and say that more data is always better. If you check out the success of the Medallion Fund managed by Renaissance Technologies with annualized return of 66% for the past 30 years, they digesting all kinds of information and operates as a data sponge, soaking up a terabyte, or one trillion bytes of information annually, buying expensive super computers to clean, digest, store, and analyze the data looking for reliable patterns. In fact, the fund reveals its secret manatra as “There’s no data like more data”.

I hope that Portfolio123 will work with the data providers and start to support using the holdings of ETFs and mutual funds as “universe” in the near future, something Georg is working very hard right now to get the results for “one” fund.

Regards
James

18

And to be clear. ANN makes Jim Simon’s early computers (at Renaissance Technologies) seem like……well, just machines;-)

Primitive machines at that. That is not a boast for ANN. It is just a fact. A fact because of the progression of computer technology and human knowledge. Google had not created TensorFlow when Jim Simons started. Everyone is familiar with Moore’s law on the doubling of computer power every 18 month. Simons started a while ago.

Yes, more data can be good, IMHO. I might be wrong but that is my opinion.

-Jim

19

If that’s support for thematic models, then I agree. Two of my public models follow that strategy. Consumer indulgences looks for value in a niche space. Even dividend income is thematic in that it’s tilted towards companies with lower capex requirements. It think it was Charlie Munger’s idea to avoid companies that have their operating profits tied up in capex. Also, capex isn’t much of a barrier to entry anymore given that some countries are willing to subsidies capex intensive industries (thinking of China a its large state-owned steel companies).

I’m hopeful that the ETF following stratgey that both SteveA and Georg have put forward, and that p123 now supports, will lead to more models beating their benchmarks. We’ll see.

Walter

20

Every time is see the word “vetting”, I think of liability.

I would like to see p123 extend the incubation time to perhaps a year. But, on the other hand, subscribers can to that now with a bit of patience.

Walter