Public Custom Europe Multi-Factor Ranking System

[hemmerling]

I decided to put together a multi-factor model for Europe which anyone with subscriptions to Europe data can use freely.

https://www.portfolio123.com/app/ranking-system/410234

Here is a factor backtest with weekly rebalancing since 2005 with 20 buckets in the easy to trade Europe and UK data set.

Here is a 10 year factor backtest with 5 portfolio sorts.

Use at your own risk. Please know that there was some data-mining in the creation of the system. I try to balance sensible factors with ‘what has worked’. This has not been optimized for turnover or weighting between factor categories. You may very well find that weighting value or momentum and lowering weight to some other category works well.

I made all factor categories equal-weight. And factors within the categories equal-weight. There are about 57 factors when I counted quickly.

I would be interested to see what you come up with based on this.

[yuvaltaylor]

A word of warning. A large number of these factors are based on quarterly income statements. This doesn’t really make sense in Europe without some modification. A very large number of European companies report semiannually, so for those companies SalesQ will be twice as large as for those that report quarterly.

The easiest way to fix this is to use TTM and Annual numbers instead of quarterly. If you really want to use quarterly numbers, there is another way around this. Anytime you use a quarterly figure from the income or cash flow statement in your numerator but not your denominator, preface your factor with this: Eval(InterimMonths(0)=3,1,0.5)*. So, for example, AstTurnQ should be Eval(InterimMonths(0)=3,1,0.5)*AstTurnQ.

There’s no need to worry about most growth factors unless you’re using PQ (which you aren’t), ratios that are unmixed between balance sheet items and income/cash flow items, or any periods greater than a quarter. TTM and annual figures should work fine, as should balance sheet items. So the items that really need fixing in your system are AstTurnQ, InvTurnQ, DbtS2NIQ (because income is in the denominator here, the code would be Eval(InterimMonths(0)=3,1,2)*), SalesPerEmpQ, ROE%Q, ROI%Q, ROA%Q, Pr2CashFlQ (=3,1,2), PEExclXorQ (ditto), and Pr2SalesQ (ditto).

[Victor1991]

Thank you, Hemmerling.

Here’s my 2 cents. From my experience the pre-slippage ranking returns in Europe are mostly ‘similar or better’ than equivelent ranking systems used in combination with a US based universe.

However, variable slippage in simulations usually takes a large chunk out of the those returns in Europe, as a larger percentage of stocks have relatively low volume. Of course, the returns in a ‘real money’ live strategy could be much better, as the slippage approximation of P123 is only that, an approximation.

[hemmerling]

Thanks Yuval. I had no idea on the semi-annual reporting. I changed it to TTM or annual as that seemed to be the simplest solution. I appreciate you sharing that.

[Whycliffes]

Thank you Hemmerling!!

I think it is great that more people are publishing more systems.

I tried to put together a system based on some different nodes I’ve looked at. I’m the amateur here, so use with caution. The system can certainly be optimized more, and there are certainly some nodes that consider the same things, like “vma”.

I have not considered or examined this system against the post from Yuval.

But:

• all nodes are weighted equally
• all factor folders are weighted equally
• I have added the “Size factor”
• there are 85 nodes in total
• Percentile NAs Neutral

https://www.portfolio123.com/app/ranking-system/410431

[hemmerling]

Thank you for this. I will look it over with interest. The one thing that Yuval mentioned with mine and I think it applies to yours as well is having to modify any quarterly factor due to many companies reporting every 6 months.

But there is another interesting anomaly in Europe I would like everyone’s opinion on. With higher yielding stocks in Europe, higher volatility actually creates a simple and effective single factor ranking system. I haven’t looked too deeply as I am on the road this weekend but it works so well it defies logic.

First, here is a 12 month volatility factor (higher is better) in the Easy to Trade Europe.

Higher volatility is better to a point and then it sharply drops off a cliff. But notice what happens when I restrict it to stocks with a yield of 5% or more. And the liquidity filter is increased to $500K as well.

Suddenly, the more volatility risk you take the higher the return.

But here is the interesting thing…your buy rules are simply yields above 5% and high volatility. But your sell rule can be something as simple as when the yield drops below 5%. So if the stock rockets up and the yield drops, you sell. It seems like any other ranking rule in the sell side actually makes it worse. And I tried it with all kinds of different start dates (because sell rules this loose would obviously be affected more by starting points). Seems to work pretty well with random start dates.

I looked at it market neutral and the returns are pretty consistent year over year being about 12% CAGR outperformance above benchmark. Here is my sim made public. Would appreciate more viewpoints. It seems a bit illogical to me. I am also wondering if the yields reported are correct as in many instances they are crazy high. Again…I will check more deeply once I am back in my office. But I value your opinions here.

https://www.portfolio123.com/port_summary.jsp?portid=1725576

[Jrinne]

Kurtis,

Also just my quick takes to be taken with more than a grain of salt.

2 different ways of looking at this:

1. the Fama and French models clearly allow for increased returns with increased risk. You point out that this is not generally true. But perhaps it is true for more liquid stock that the finance industry is more likely to follow. I.e., for the investors who have read the text books that include Fama and French’s it works.

2. but there is a better way to look at this in my opinion. An only SOMEWHAT more complex statistical method. If you divided the returns by the inverse variance I bet the EFFECT would disappear. In other words the EFFECT SIZE (which is often the reciprocal of the variance) MIGHT be the same. I say might because I have not looked at your data.

This is consistent with Meta-analysis which is a heavily used technique for combining studies in the medical literature (which divides the means by the inverse variance to combine studies).

Anyway, I looked forward to other takes. But ultimately I think it will come down to the same—seemingly complex—reason.

Here is something about inverse weighting from Wikipedia (and one could google meta-analysis for more information):Inverse-variance weighting

Effect size would be another thing to Google—assuming what I say does have any relation with is a bit of speculation I admit.

I will have more time when I am goofing of in the office A sincere quick-take on your observation for now.

Jim

[hemmerling]

Some good points here. But wouldn’t inverse-variance weighting just be pretty much the opposite of the volatility ranking? The whole idea of CAPM and higher beta equating into higher returns just hasn’t worked out. The low volatility factor premium is in defiance of that.

And overall, lower volatility to a point is better. Until you look at higher yielding stocks…in Europe. Seems odd that a factor works in one universe but not the other.

It might be similar to the idea that higher risk has higher returns once you remove the ‘lottery like’ risk stocks where people price them poorly. For instance, for negative FCF stocks in microcaps…the value premium is negative and gets worse with deeper value overall. But in positive FCF stocks it is exactly the opposite.

Maybe higher yielding stocks in some way signals lower risk? So higher volatility is good? This also doesn’t make sense though because you would think that very high yield is higher risk due to depressed prices.

Still thinking on this…

[Jrinne]

Kurtis,

To be clear I was talking about dividing the returns by the volatility (and providing a rationale for doing that),

The main question, I think, would be whether my rational can be justified or whether it is a bunch of after-the-fact handwaving.

I am fine with the idea that it is all after-the-fact justification better known as: BS. It remains a serious candidate as explanation in my mind (I am fine if it is just me). But clearly, I could be just plain wrong about any cause or effect. Maybe it is just a spurious coincidence that the math works.

But I think coincidence or not the math does work. iI you look at the buckets the y-axis is the returns while x-axis is the volatility. Better- known as the slope. And things look pretty linear. I.e. there is a clearly define linear relationship with one well-defined slope. The it just a fact that is immediately obvious, I believe.

The returns over the variance (the slope in this case) is also known as the effect size as it turns out.

So all of the buckets have the same effect size which was actually the only point I was making. As I mentioned justification for that actually meaning anything is clearly debatable.

It is a fair question as to whether I could have predicted this ahead of time or whether it is an ad-hoc, hand-waving, after-the-fact explanation.

Probably the latter I admit. But I think the math is correct. I will keep looking too. I have to BS about something out of the office

Seriously, effect size is an important concept and I have posted about it before. From my iPhone on the weekend the math seems to be correct or interesting at a minimum.

Not statistics so much, but science is always happy to find constants (especially in physics).

I appreciate your post whatever the answer is. Very interesting!!! I learned something even if it only causes me to have confirmation bias in thinking the effect size may be useful in some situations.

Jim

[Whycliffes]

Thanks Victor! I have not thought about the effect variables have on the return itself and how it therefore gives you advantages in reducing turnover.

If variables do not present a realistic picture of costs, you can easily forcing turnover down unnecessarily, which will have a negative impact on the return because you have taken into consideration a cost in turnover that is not essential.

What are your best tips for reducing turnover?

Feeding off the last bullet point, include equal in the compare like AvgRec <= AvgRec4WkAgo. This keeps stable recommendations in your set which helps a lot with limited analyst coverage and addresses the left boundary condition of 1.00 as as unchanged 1.00 < 1.00 would be excluded. Cheers, Rich

Here is the difference in one of my systems, with and without “variables”:

Both have “Force Positions into Universe”

[Jrinne]

This is an interesting topic and I asked ChatGPT3.5 about this to try to learn more.

It is interesting that it has an understanding of the meaning and importance (or lack of importance perhaps) of effect size in this context without any prompting from me.

Here was my question: “Why is the effect size for stock returns constant?”

FWIW, here was the response from ChatGPT 3.5: "The effect size for stock returns is not necessarily constant. In fact, stock returns can be highly volatile and unpredictable, and their effect size can vary depending on a number of factors, such as market conditions, company performance, economic indicators, and geopolitical events.

However, some studies have shown that over long periods of time, the average effect size of stocks tends to be relatively consistent. This is known as the equity premium puzzle, and it refers to the fact that stocks have historically provided higher returns than safer investments like bonds, even though they are riskier.

One explanation for this phenomenon is that investors demand a premium for taking on the risk of investing in stocks, and this premium remains relatively constant over time. However, there is still ongoing debate among economists and financial analysts about the underlying causes of the equity premium puzzle, and it is not yet fully understood."

Edit: I am not sure that ChatGPT got that entirely correct. If I understand the equity premium puzzle (EPP) correctly it attempts to explain why the effect size is DIFFERENT for stocks and bonds. Not the same as it seems ChatGPT was, incorrectly, implying here.

Despite ChatGPT’s obvious error regarding EPP (and bonds) it does leave open the possibility that the effect size for stocks is fairly constant over longer periods.

And clearly ChatGPT got that the fact that the broader topic is widely debated with no clear answers (a puzzle not just for us or me at least). I don’t mean to imply that I have added much to the debate as far as finding an underlying cause for all of this. I have found the effect size useful in other contexts.

BTW, since ChatGPT 3.5 seemed to be incorrect I upgraded and asked ChagGPT 4 the same question. Notice ChatGPT 4 does leave out any reference to the equity premium puzzle which was the main error I identified. In addition, I think relating this to the efficient market hypothesis (EMH) is more accurate—especially since it calls into question whether the EMH is always correct and says there are probably anomalies or exceptions to the EMH. It is not dogmatic in its support of the EMH.

ChatGPT 4: "The assertion that the effect size for stock returns is constant may not be entirely accurate. Stock returns are influenced by various factors, which may change over time. These factors include economic conditions, market sentiment, company performance, geopolitical events, and investor expectations, among others. As these factors are dynamic and subject to change, it is challenging to assume that the effect size for stock returns remains constant.

The Efficient Market Hypothesis (EMH) posits that stock prices incorporate all available information and that it is impossible to consistently outperform the market. However, the EMH has been challenged by various anomalies and market inefficiencies, which suggest that stock returns may not be entirely predictable, and their effect size may not be constant.

There are different approaches to measuring the effect size for stock returns, such as the Capital Asset Pricing Model (CAPM) and the Fama-French Three Factor Model. These models attempt to explain stock returns using a limited number of factors. However, the stock market is a complex system with a multitude of factors that are constantly changing. As a result, it is difficult to generalize and claim that the effect size for stock returns is constant.

In summary, the effect size for stock returns may not be constant due to the dynamic nature of the factors that influence stock prices. Market anomalies and inefficiencies also challenge the notion of a constant effect size in stock returns."

This mere mortal does not have much to add to that I think it is not bad.

Jim