We need a risk model

[hemmerling]

Yuvall,

I am going to stay out of this conversation from here on. I just wanted to defend my own position that I was against Betas. That’s not the case. And my issue with this is a fundamental one based on correlations.

Suppose I trade turnips from my garden. Then we compare this to the price of Apple. Suppose they appear to have high trailing correlation to each other. We then calculate the Beta with the method of dividing the StdDev of my turnip prices by Apple’s StdDev. And multiply this by the correlation. The question is - what is the real link between my turnips and the price of Apple? Zero. The trailing prices may have high correlation which is a component of Beta. But in reality, there is no casual link between them. My high turnip to Apple Beta factor is not useful. If Apple releases a new Iron Man suit and goes up 10x, my turnips are not going to go up 20x regardless of the trailing correlation and comparison of standard deviation of returns. Now if I knew I was selling Apple all my turnips and they were extracting turnip juice to power the Iron Man suit…then suddenly my turnip to apple beta is meaningful. But you have to make a causal connection in order to say this and not just looking at correlation and volatility.

My issue with this thesis is that you can have a portfolio of stocks that have no actual ‘value risk’ but it may look like you have high beta exposure to the value factor. And if I do not have value stocks in my portfolio, how can someone say I have high exposure to value risk just based on trailing price returns?

This is why I was advocating an approach that involves looking at actual value characteristics to determine exposure to the value factor. Perhaps this could somehow be modified to include market beta (which I use). Thus, I would feel confident in saying that I have a strong value tilt and I also high market beta. But I wouldn’t try to make some connection of a beta to the value premium using portfolio returns to value premium returns. Seems like a lot of assumptions are made in the middle. Too much is read into correlation and standard deviation without ever looking at the actual factors.

[Jrinne]

Kurtis,

You are right. This should all be done by looking at the financial data.

But if the price of eggs and apples and cereal and….everything else is going up at the supermarket I am not going to short port bellies (is that still a thing).

Admittedly, better if I check on the supply of pork bellies, price of feed etc before making a decision. Not the type of thing I do every morning after seeing if the data is up to date at P123, however.

Sometimes I am not ever sure what the name of the stock is or its sector—assuming I can remember the tickers for all of the stocks I have bought. Just for me personally, I am going to say I use some correlations no matter how you look at it.

EVERYTHING we do with ranks is correlation based. Value metrics “correlate with the price-action” for reasons we can only speculate about with behavioral economics.

But you have a point.

Jim

[hemmerling]

Jrinne,

Ok. I lied. I am going to reply again. I agree with you. And the illustration you gave is good. However, you named the products and pork bellies is related to food. That’s the fundamental and causal link. But suppose you were blindfolded and you had no idea which products or services were under discussion. You didn’t know if it was food or pet rocks or tickles. You would have a harder time making the link.

And correlations are fine as long as they are paired with something. And it is also the reason why a hyper-fitted ranking system with no logic is not worth trading. All correlation and no underlying reason.

At best, I think the risk meter could be a simple tool to say, ‘hey maybe you have too many value stocks here and also high beta stocks. Take a look’. And then I take a look and maybe it is true and maybe not. If so, I can remove a few value stocks based on characteristics. If that is my goal. Unless I want it to be a value portfolio. But I probably already know that I have a value tilt and have high market beta just by looking at the average factor rank for beta and value.

The risk model could be a tool to simply suggest further inspection if you are concerned about unintended factor tilts. But I would imagine the more active your strategy is, the less meaningful this will be. If I am trading one-week price reversions, I wouldn’t put too much stock into an aggregate price to book benchmark.

[yuvaltaylor]

Kurtis -

How could you possibly have any meaningful correlation between the price of turnips and the price of AAPL? Correlation doesn’t just happen randomly. If there’s no actual relationship, there’s going to be no correlation between the returns.

For example, the returns of mining stocks are going to be more correlated with those of chemical stocks than with those of financial stocks. A portfolio based on price to sales will be more correlated with a porftolio based on EV to EBITDA than with a portfolio based on earnings growth or price momentum. Correlation studies provide evidence that factors are related to each other. OK, my examples are obvious and trivial. But I have found some rather surprising relationships by studying correlations. Keep an open mind!

[Jrinne]

Kurtis. My analogy cannot be taken too far. But my car cost me a lot more than even the sticker price when I bought it.

Turned out the price I paid for my car was related to pork bellies in some weird way. But surely something is not correlated to pork bellies……hmmm, seems like I pay a lot for everything now but……uh, something.

Better if you can figure it out based on principles but that is truly hard too. Remember "transient "inflation? All COVID still?

I read 100 reasons for things in the news every day. Some happen to be correct and a few of those only by chance.

Correlation has its limits and I do not really disagree with you I think. I am just saying I use it every trading day in the morning. That’s all: Not to be totally discounted.

[hemmerling]

Yuvall,

I think this is being taken out of context. Correlations are fine. But say nothing of causation. Small people drive small cars and big people big cars. But I won’t grow just because I buy a bigger car.

Correlations are great. But now we are using correlations and volatility to imply that the two are linked and one is directly tied to the risk of the other. Without looking any further under the hood. We need to go one step further.

I am making the case for using Betas, correlations, covariances or whatever paired with something that attempts to make a logical connection between the two. I am not saying throw it out. I am saying we could probably make it better if we find a way to establish a higher likelihood of their being an actual relationship.

This reminds me of pairs trading. It makes a great narrative in that you find two stocks which are co-integrated and they move around some central mean with an equilibrium. Then you look at the stocks and they are so totally unrelated. And the co-integration falls apart with no profit. But if you can find 2 stocks which are fundamentally similar, in the same industry group where returns are being made from the same drivers, then there is a higher chance that the variance in returns might come back to some mean. You marry the two concepts.

I have a very open mind. But there are reasons why this type of risk model has had so many changes over the past 10 - 12 years. I think with the brains here we can actually design something new and better. And that has some forward risk mitigation. But if we want to just make a copy of an older risk model…as I said before…that’s fine. I just want to see more effort going into a causation. So far, that hasn’t been addressed. Adding in factor characteristics into the equation takes us one step closer to that.

edit:
To be honest, I never meant for this to get so out of hand. I am truly sorry for that. I am not some Stone Age villain trying to prevent technological progress. I do want us to build a better risk model. And for real this time, I will leave this thread. You have nothing to gain from my input. It seems like I am just creating a debate which I didn’t mean to. Sorry if I come across combative. It is hard to properly express the emotion behind the words in posts. I meant no harm.

[Jrinne]

Kurtis,

Maybe I should apologize. I do use correlations and sincerely believe it is pretty much all I do with regard to investing at P123—even using P123s ranking system.

I do not understand how P123’s method is not like a regression or correlation of future returns to factors or factor ranks. I have no ability to see it otherwise. Given that I simply cannot understand how something being a correlation automatically disqualifies it as a method.

It is a sincere view. Using my car and pork bellies as examples was not meant to be personal in any way.

Jim

[judgetrade]

I think it is crazy to calculate risk of a factor based on factor statistics.

That is trying to determine the risk of a car loaded with factors not seeing the weather (tornado, snowstorm or sunny weather?) or the street (racetrack or bad road). And there is no car that will drive you through a tornado without risk. Correlations are fine until all correlations = 1 (in a crash like 2008 no factor in the world would have saved you from at least a 40% drawdown long only).

Risk is not driven by factors, risk is driven by the macro regime which favors factors or dislikes factors and the valuation of the factor to itself historically (2000 value was cheap, 2007 it was expensive etc.).

Important in that regard is that you have a balanced ranking system that diversifies over a lot of factors, so your timing of what macro likes does not have to be perfect.

Also you need an alpha source that can not be easily arbitraged away (therfore micro caps and
small cap where the big institutions can not trade due to liquidity reasons).

Take your best strategies, put them in a book, one long only and one with a short ETF (which correlates to the factors you use in your book) and compare the capital curves and decide in which macro regime you choose which book you use (long only or hedged).

I you design your book with a 50% short, and you see no alpha start over (because then you
have a beta book!).

Book with a 50% Short

Also, it is very important to know which factors come back strong and fast after a 50% drawdown if
your timing was wrong (e.g., you did not choose the hedged book in a severe macro regime).

High beta can take a decade to come back and make new highs, small cap value momentum takes max 1-2 years (good systems below 1 Year!) to make new highs.

Andreas

[Jrinne]

TL;DR: Here is where we are in the forum with medicine as an analogy (and where I was when I finished medical school). I.e.,before statins (cholesterol lowering agents became mainstream): “Hey, did you read that article that said cholesterol levels are correlated with heart disease?” There are still serious questions about this but progress can be found in most texts on the subject.

BTW, they may still be wrong about many things but they were definitely wrong about cholesterol then. Correlation not being causation was discussed a lot with cholesterol as well as with smoking causing cancer. It is always just a lot of handwaving and gesticulating early on. But my main point is you cannot become a doctor without text-level understanding of many subjects, I think. And rightly so.

@judgetrade You said: “Important in that regard is that you have a balanced ranking system that diversifies over a lot of factors, so your timing of what macro likes does not have to be perfect.”

I have a simple question: Isn’t this exactly what Korr123 is trying to accomplish on the portfolio level? I understand that Korr123 may also be wanting to balance his shorts but I think that would only emphasize that the purpose is to balance and “diversify over a lot of different factors.”

BTW, I really do not care what P123 does with this. I think I will end up staying with Portfolio Visualizer. If nothing else the discussions in the forum are going to slow adoption of many useful techniques in the near future.

But I do not really fully understand what Korr123 is doing and want to learn. I would like to understand how this is used and how effective it is at accomplishing its intended use.

Has Ricccardo studied this in any detail in school so we do not have to reinvent the wheel on this STARTING WITH THE USEFULNESS OF CORRELATIONS? Some of this has been fully discussed with much of this in the 5th or 6th edition of a text, I would guess.

I am pretty sure that if correlations are discussed in a text about Finance at all, the general usefulness of correlations is established (or discounted) in an earlier chapter. Real sure actually.

But getting back to Andreas’ quote, isn’t that exactly what Korr123 is trying to measure and accomplish at the portfolio level? For long only portfolios, wouldn’t it balance diversification over a lot of factors which you said is desirable? Or isn’t that the goal, at least, with the question of how effective it is at doing that a separate question that could be answered empirically?

Note, I have no interest at all in how one would use this to create short positions but others might have a legitimate interest in that.

The only other comment I would have is mainly about my own comments (but not exclusively my comments). Pork bellies, car prices or how good cars are–when loaded with factors–at avoiding tornadoes will only go so far. I think I will be looking at some text books and papers on the subject at some point. There is certainly a lot about beta in any serious text. I think what Korr123 is interested in implementing is talked about in the texts in detail also. I have read about it, I believe, but it is a more complex subject and my understanding was limited on the first read-through. This is not in my priority list now (did I say I don’t really care what P123 does with this but I would like to learn about the technique as long as people are posting something with some useful information) but I don’t mind learning about it in the forum now.

Marco presented a paper in this regard earlier: @Marco: Thank you for advancing the discussion a little better than I have been able to so far. Real research is much appreciated.

While this is a complex topic that I do not fully understand, I still have the simple question about whether this is trying to accomplish exactly what judgetrade says is important. Or not.

To be sure, the above should reiterate that I do not know enough to make a recommendation about where this should be on anyone’s priority list. I am only saying this discussion might not be the final word for me. I will probably ignore the discussions about correlation—although you should check out the great insights about pork bellies above

Jim

[korr123]

@Jrinne

I still have the simple question about whether this is trying to accomplish exactly what judgetrade says is important. Or not.

From my understanding of @judgetrade’s post, his brings up two risk management methods, both used widely by investors with quantitative strategies. My initial post focused on one of them.

The first, as @Jrinne points out, is a factor based risk model. And that is precisely what I introduced - so yes! I read @judgetrade’s post as first saying it’s crazy (“I think it is crazy to calculate risk of a factor based on factor statistics”) but then discussing how one needs alpha, not beta.

“Take your best strategies, put them in a book, one long only and one with a short ETF (which correlates to the factors you use in your book) and compare the capital curves and decide in which macro regime you choose which book you use (long only or hedged). I you design your book with a 50% short, and you see no alpha start over (because then you have a beta book!).”

So I think thinking through what is alpha and what is beta would be productive in that context. Perhaps the remarks were meant to discuss how one’s risk model should be built - and in his case, in discrete time frames? I believe you will still come to the realization that a factor based risk model is necessary though I’d love to hear how you’d quantify that future risk. And to Jrinn’s point, that is the beauty of PV.

The second, is scenario analysis, or as @judgetrade put it, macro risk. That is, for example, when correlations go to one. That is the equally important, imo. The intent of my post in this thread is that you cannot do anything other than looking at the past via rolling simulations or something similar (with P123’a current tech), to do this exercise in any numerical way. I’m suggesting a way to improve this.

However, if you were to understand your strategy’s sensitivity to a return stream (i.e beta), you would be able to quantify that, at least partially. So, in his example, you’d want to look how a portfolio of stocks respond to a “regime”. Well, you could run a simulation based on a trimmed time scale - that is his approach (which comes with the assumption that if that regime returns, he will be holding the same securities). But I maintain that one must still quantify that regime to come up with a profit or loss for that environment. Otherwise, I think one is left throwing their hands up in the air leaving it without numerical results and once again looking in the rear view mirror for a risk event that has transpired. How accurate your result depends on the quality of your research - no different then your research on the various factors with which you invest.

The purpose of my request had two implicit benefits which may seems to have gotten lost in this thread: you cannot know your alpha without knowing your beta. And a risk model that shows your presumed alpha is actually beta, is a good thing to know! I’m not saying this a holy grail but a very large step in the right direction to help manage risk and improve returns.

To be clear, scenario analysis is different than a factor based risk model. And to the best of my knowledge, comprehensive scenario analysis cannot be done without the risk model I have described.

Hope this helps.

[korr123]

@yuvaltaylor @marco

Looks like this hasn’t progressed for whatever reason. As a workaround, is it possible for us to upload risk factor data and use either the regression or betafunc factors to call those custom series? (Some assistance would be helpful since I’ve had trouble doing it).

I think that will at least enable members to calculate their own risk factor betas and include that into models for analysis. No UI work on your end.

Separately, was there any thought on enabling users to do ranking analysis on alpha returns (calculating the alpha from the risk factor models) vs total returns, which is what we have now?

Thank you.

[marco]

I moved this to the roadmap where voting is enabled. We definitely want to do something in this area but right now we’re trying to get the AI factors launched. We will be discussing the roadmap in July where we’ll all get together face to face (probably in Milan in case someone wants to meet us)

[korr123]

I understand.

But it wasn’t explicit as regards to my other question:

Is it possible to use any of the current current regression function or betacalc functions using “get series” on an uploaded custom series?

[enisbe]

Marco on AI factors lunch what is your timeline expectation?