Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Discuss all general (i.e. non-personal) investing questions and issues, investing news, and theory.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

triceratop wrote: Fri Dec 15, 2017 7:50 pm
rkhusky wrote: Fri Dec 15, 2017 7:47 pm The above is incorrect. The factors are constructed such that they are zero for the market and involve long and short positions. There is no small factor or value factor in the standard formulation. See. e.g., http://mba.tuck.dartmouth.edu/pages/fac ... loped.html
I fail to see your point. You absolutely can get exposure (with a loading strictly less than 1) to SmB and HmL using long-only (index, too) funds. People call these the Small and Value factors. What is your specific objection?
I've never said that one could not get factor exposure from a given fund. Confusion arises when the correct terms are not used. There is no small factor or value factor in the standard Fama French formulation. The factors are SmB and HmL.
User avatar
patrick013
Posts: 3301
Joined: Mon Jul 13, 2015 7:49 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by patrick013 »

dbr wrote: Fri Dec 15, 2017 5:02 pm
nisiprius wrote: Fri Dec 15, 2017 4:50 pm One of the things that I don't quite understand... I had originally thought that factors were extracted by a mathematical technique called factor analysis. I've never been deeply enough into it to know offhand just how that works. It's similar, but not identical to principal components analysis, which I have done (i.e. written computer programs to do it). Principal components analysis takes a swarm of measurements of many different things. In the case of stocks it could be, say, the 600 monthly total return numbers of each individual stock for the last fifty years, treats them as points in multidimensional space, and asks "are these really scattered across 600 different dimensions?" Often they aren't, and by rotating the axes around you can arrange them so that one axis goes through the longest spread of the swarm, then you sweep the second axis around, staying at 90% to the first axis, until it transects the widest spread it can, and so forth. With luck, it may turn out that the whole swarm is really only one-dimensional, or two-dimensional, or three-dimensional, with a fairly small amount of noisy spread in the other directions. The first few dimensions capture almost all of the total variance of the swarm.

Now that sounds a lot like the way the factor mavens talk, but there's an important difference. The procedure for principal components analysis keeps the axes at 90° to each other. That is, the axes are constructed in such a way that there is always zero correlation between the projection of each point on the axes.

Apparently that's not what the financial factor investigators do. Apparently they simply define the factors in some plausible ad-hoc way, and then observe what the correlations are; and they aren't zero. They aren't independent. That is to say, if I have my head screwed on right, the momentum factor itself has a loading on the value factor, and so forth.

I don't know why they do it that way. That's just a statement. I'm not saying they're wrong, I'm saying I don't understand why they do it that way.
Right about factor analysis which this is not.

In this area of finance it is pretty clear factor just means "that which predicts return" in the sense of being a variable in a regression model that has some meaningful degree of explanatory power. This is just the word factor in the sense that all the Xi are factors in a formula like

Y = Ao + Sum (AiXi) for i = 1,n

When you do a regression like that one issue is what is the smallest set of factors that gives the best explanatory power. This can even be found out by the technique of stepwise regression. Another issue is the issue of selecting among factors that are correlated. The best thing is for the factors one ends up with not to be much correlated with each other and a lot of finagling with the model can occur in doing this. This is part of that business of why dividends, for example, don't usually get picked as factors. But unlike factor analysis, there is no attempt to find a new basis set of combinations of initial factors that gives a good explanation with factors that are all computed to be uncorrelated. At least that would be my understanding of it. I think the whole thing is less rather than more than meets the eye.
If I had 20 years of history of the RUSSELL 2000 and the same for a small
cap fund and did an ordinary least squares regression I would get a number
for beta. If I did the same for a LC vs. the 2000 beta should be smaller,
proving a SC premium exists for SC funds when regressed with their data.
Or something like that. Am I close ?

Just looking into this a little bit.
age in bonds, buy-and-hold, 10 year business cycle
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

triceratop wrote: Fri Dec 15, 2017 8:08 pm
rkhusky wrote: Fri Dec 15, 2017 8:07 pm
nedsaid wrote: Fri Dec 15, 2017 8:01 pm Well, there is a Morningstar video out there where the example of the disappeared Small-Cap effect re-emerging with the S&P 600 Small-Cap Index is discussed. I didn't make this up. This was also discussed in another thread.
The small cap effect is not the same as the SmB factor in the Fama French factor model, although it is related. The SmB factor involves the difference in return between small and big stocks.
The small cap effect is about outperformance relative to the market, no? Please explain how they differ in kind, not in degree.
SmB is more specific than the small cap effect:
SMB =1/3 (Small Value + Small Neutral + Small Growth)
- 1/3 (Big Value + Big Neutral + Big Growth).
User avatar
triceratop
Posts: 5838
Joined: Tue Aug 04, 2015 8:20 pm
Location: la la land

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by triceratop »

I think most people here who know about SmB and HmL understand that small factor refers to SmB. The reverse is not true: if I talk about HmL and SmB very few will understand what I am saying.

Given the discussion has tended towards definitional semantics I'll bow out.
"To play the stock market is to play musical chairs under the chord progression of a bid-ask spread."
User avatar
nedsaid
Posts: 19249
Joined: Fri Nov 23, 2012 11:33 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nedsaid »

rkhusky wrote: Fri Dec 15, 2017 8:07 pm
nedsaid wrote: Fri Dec 15, 2017 8:01 pm Well, there is a Morningstar video out there where the example of the disappeared Small-Cap effect re-emerging with the S&P 600 Small-Cap Index is discussed. I didn't make this up. This was also discussed in another thread.
The small cap effect is not the same as the SmB factor in the Fama French factor model, although it is related. The SmB factor involves the difference in return between small and big stocks.
Well, that is why I have overweighted Small stocks because in theory they outperform Large ones. I want to capture that difference in returns. As far as I know, the two effects are the same.

Here is the Investopedia definition of the Small Firm effect:
A theory that holds that smaller firms, or those companies with a small market capitalization, outperform larger companies. This market anomaly is a factor used to explain superior returns in the Three Factor Model, created by Gene Fama and Kenneth French - the three factors being the market return, companies with high book-to-market values, and small stock capitalization. A theory that holds that smaller firms, or those companies with a small market capitalization, outperform larger companies. This market anomaly is a factor used to explain superior returns in the Three Factor Model, created by Gene Fama and Kenneth French - the three factors being the market return, companies with high book-to-market values, and small stock capitalization.
Here is the Investopedia definition of Small minus Big effect:
Small minus big (SMB) is one of three factors in the Fama and French stock pricing model. SMB accounts for the spread in returns between small- and large-sized firms, which is based on the company's market capitalization.

This factor is referred to as the "small firm effect", as smaller firms tend to outperform large ones.


My untrained eye tells me that the Small Cap effect and Small Minus Big are the same thing.
A fool and his money are good for business.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

patrick013 wrote: Fri Dec 15, 2017 8:14 pm
If I had 20 years of history of the RUSSELL 2000 and the same for a small
cap fund and did an ordinary least squares regression I would get a number
for beta. If I did the same for a LC vs. the 2000 beta should be smaller,
proving a SC premium exists for SC funds when regressed with their data.
Or something like that. Am I close ?

Just looking into this a little bit.
Actually, the regression is completely separate from the determination that there is any sort of premium. The regression just tells you how the returns of your sample portfolio compare to the returns of the baseline portfolio. If there is high correlation, then you will know that if the baseline portfolio does well, then your sample portfolio will also do well. If the baseline portfolio does poorly, then your sample portfolio will also do poorly. The correlation factor tells you the degree to which the returns match. For example, if the correlation factor is 0.5 and the baseline portfolio rises by 2%, then your sample portfolio will rise by 1%. Similarly for losses.
lack_ey
Posts: 6701
Joined: Wed Nov 19, 2014 10:55 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by lack_ey »

triceratop wrote: Fri Dec 15, 2017 7:50 pm I fail to see your point. You absolutely can get exposure (with a loading strictly less than 1) to SmB and HmL using long-only (index, too) funds. People call these the Small and Value factors. What is your specific objection?
As an aside to the actual point being made, I don't think that's quite right. You can get size loadings above 1 in a long-only portfolio by using concentrated micro caps smaller than the typical stock used on the long side of the SmB formulation. Or by choosing small stocks that I suppose are economically significant in character to what drives small cap performance, and would thus generate high loadings even without being quite as small.*

*similar to the idea where you can have a stock that behaves more like a value stock, regressing on value, despite being in growth territory (negative value) by P/E and other metrics

For example
https://www.portfoliovisualizer.com/fac ... sion=false
Last edited by lack_ey on Fri Dec 15, 2017 8:30 pm, edited 1 time in total.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

triceratop wrote: Fri Dec 15, 2017 8:15 pm I think most people here who know about SmB and HmL understand that small factor refers to SmB. The reverse is not true: if I talk about HmL and SmB very few will understand what I am saying.

Given the discussion has tended towards definitional semantics I'll bow out.
Unfortunately, you just cause confusion when you talk about a small factor, leading people to think that Total Stock Market contains the small factor because it contains small stocks.
User avatar
nedsaid
Posts: 19249
Joined: Fri Nov 23, 2012 11:33 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nedsaid »

rkhusky wrote: Fri Dec 15, 2017 8:12 pm
triceratop wrote: Fri Dec 15, 2017 7:50 pm
rkhusky wrote: Fri Dec 15, 2017 7:47 pm The above is incorrect. The factors are constructed such that they are zero for the market and involve long and short positions. There is no small factor or value factor in the standard formulation. See. e.g., http://mba.tuck.dartmouth.edu/pages/fac ... loped.html
I fail to see your point. You absolutely can get exposure (with a loading strictly less than 1) to SmB and HmL using long-only (index, too) funds. People call these the Small and Value factors. What is your specific objection?
I've never said that one could not get factor exposure from a given fund. Confusion arises when the correct terms are not used. There is no small factor or value factor in the standard Fama French formulation. The factors are SmB and HmL.
The standard usage by Larry Swedroe is Size and Value. I said "small" to emphasize that smaller stocks are supposed to outperform larger stocks. That is the way I communicate, a lot of people get lost with SmB and HmL, in fact I had to look them up. But I understand the language of Market, Size, Value, Momentum, and Profitability/Quality as this is what Larry uses in his books. Easier to understand and easier to communicate.
A fool and his money are good for business.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

nedsaid wrote: Fri Dec 15, 2017 8:16 pm
rkhusky wrote: Fri Dec 15, 2017 8:07 pm
nedsaid wrote: Fri Dec 15, 2017 8:01 pm Well, there is a Morningstar video out there where the example of the disappeared Small-Cap effect re-emerging with the S&P 600 Small-Cap Index is discussed. I didn't make this up. This was also discussed in another thread.
The small cap effect is not the same as the SmB factor in the Fama French factor model, although it is related. The SmB factor involves the difference in return between small and big stocks.
Well, that is why I have overweighted Small stocks because in theory they outperform Large ones. I want to capture that difference in returns. As far as I know, the two effects are the same.

Here is the Investopedia definition of the Small Firm effect:
A theory that holds that smaller firms, or those companies with a small market capitalization, outperform larger companies. This market anomaly is a factor used to explain superior returns in the Three Factor Model, created by Gene Fama and Kenneth French - the three factors being the market return, companies with high book-to-market values, and small stock capitalization. A theory that holds that smaller firms, or those companies with a small market capitalization, outperform larger companies. This market anomaly is a factor used to explain superior returns in the Three Factor Model, created by Gene Fama and Kenneth French - the three factors being the market return, companies with high book-to-market values, and small stock capitalization.
Here is the Investopedia definition of Small minus Big effect:
Small minus big (SMB) is one of three factors in the Fama and French stock pricing model. SMB accounts for the spread in returns between small- and large-sized firms, which is based on the company's market capitalization.

This factor is referred to as the "small firm effect", as smaller firms tend to outperform large ones.


My untrained eye tells me that the Small Cap effect and Small Minus Big are the same thing.
The key difference is that the SmB factor is formulated to be zero for the market, being Small minus Big. Therefore, the SmB factor is zero for the market, even though the market contains small cap stocks. Using terminology like a small factor can lead to people mistakenly thinking that the market has exposure to the small factor because it has small stocks.
User avatar
Taylor Larimore
Posts: 32839
Joined: Tue Feb 27, 2007 7:09 pm
Location: Miami FL

Factor-Based Investing With Additional Funds ?

Post by Taylor Larimore »

Bogleheads:

It is very difficult to "beat the market" although nearly everyone in the investment industry claim they can do it by using "past performance" which is so dangerous that even the government requires mutual fund advertisements to warn us: "Past performance does not guarantee future performance."

Before complicating your portfolio with additional funds, remember what these two very knowledgeable investment authorities have to say:
Michael Edesses, author of The Big Investment Lie: "As a mathematician I know when mathematical-sounding analyses are little more than elaborate sales pitches, designed to thoroughly obscure the simple fact that smart investing is non-mathematical and accessible to everyone."
Jack Bogle: "The beauty of owning the market is that you eliminate individual stock risk, you eliminate market sector risk, and you eliminate manager risk. -- There may be better investment strategies than owning just three broad-based index funds but the number of strategies that are worse is infinite."

I'll also add a quote from the common sense wisdom of my co-author:
Laura Dogu, Ambassador to Nicaragua: "A simple portfolio is actually the ultimate in sophistication. It almost always lowers cost (including taxes), makes analysis easier, simplifies rebalancing, simplifies tax-preparation, reduces paper-work and record-keeping, and enables caregivers and heirs to easily take-over the portfolio when necessary. Best of all, a simple portfolio allows the investor to spend more time with family and friends."
Bogleheads can read what other investment experts say about "Simplicity" by using the link at the bottom of this window.

Best wishes.
Taylor
"Simplicity is the master key to financial success." -- Jack Bogle
User avatar
nedsaid
Posts: 19249
Joined: Fri Nov 23, 2012 11:33 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nedsaid »

rkhusky wrote: Fri Dec 15, 2017 8:33 pm
The key difference is that the SmB factor is formulated to be zero for the market, being Small minus Big. Therefore, the SmB factor is zero for the market, even though the market contains small cap stocks. Using terminology like a small factor can lead to people mistakenly thinking that the market has exposure to the small factor because it has small stocks.
I do understand that to capture factors, you need to tilt away from the market. If you want to capture the Value factor, you need a greater proportion of Value stocks than in a market portfolio. If you want to capture the Size factor, you need more than a market weighting of small stocks. But even tilting won't capture everything as most investor portfolios are long only and most portfolios don't contain extreme tilts like the "Larry portfolio" which uses only Small Value Stocks in the equity portion of the portfolio.
A fool and his money are good for business.
User avatar
patrick013
Posts: 3301
Joined: Mon Jul 13, 2015 7:49 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by patrick013 »

rkhusky wrote: Fri Dec 15, 2017 8:21 pm
patrick013 wrote: Fri Dec 15, 2017 8:14 pm
If I had 20 years of history of the RUSSELL 2000 and the same for a small
cap fund and did an ordinary least squares regression I would get a number
for beta. If I did the same for a LC vs. the 2000 beta should be smaller,
proving a SC premium exists for SC funds when regressed with their data.
Or something like that. Am I close ?

Just looking into this a little bit.
Actually, the regression is completely separate from the determination that there is any sort of premium. The regression just tells you how the returns of your sample portfolio compare to the returns of the baseline portfolio. If there is high correlation, then you will know that if the baseline portfolio does well, then your sample portfolio will also do well. If the baseline portfolio does poorly, then your sample portfolio will also do poorly. The correlation factor tells you the degree to which the returns match. For example, if the correlation factor is 0.5 and the baseline portfolio rises by 2%, then your sample portfolio will rise by 1%. Similarly for losses.
The Factor Model I was looking at was a general model, not FF exactly.
My first read was the factor load was simply beta, pending further analysis.
R2 comes into play but just for the model accuracy. So the stats made a
little more sense like that.
age in bonds, buy-and-hold, 10 year business cycle
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

nedsaid wrote: Fri Dec 15, 2017 8:40 pm
rkhusky wrote: Fri Dec 15, 2017 8:33 pm
The key difference is that the SmB factor is formulated to be zero for the market, being Small minus Big. Therefore, the SmB factor is zero for the market, even though the market contains small cap stocks. Using terminology like a small factor can lead to people mistakenly thinking that the market has exposure to the small factor because it has small stocks.
I do understand that to capture factors, you need to tilt away from the market. If you want to capture the Value factor, you need a greater proportion of Value stocks than in a market portfolio. If you want to capture the Size factor, you need more than a market weighting of small stocks. But even tilting won't capture everything as most investor portfolios are long only and most portfolios don't contain extreme tilts like the "Larry portfolio" which uses only Small Value Stocks in the equity portion of the portfolio.
Sure. And by capturing part of the factor, you are capturing either part of the future out-performance or part of the under-performance of the factor in the same proportion.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

patrick013 wrote: Fri Dec 15, 2017 8:43 pm The Factor Model I was looking at was a general model, not FF exactly.
My first read was the factor load was simply beta, pending further analysis.
R2 comes into play but just for the model accuracy. So the stats made a
little more sense like that.
That sounds like the model that preceded FF, which I believe was called CAPM, and did not address size or book/market. The procedure for CAPM is the same as FF, just simpler. CAPM also just provided the correlation, it didn't tell you whether or not the market was going to go up or down or outperform the risk free investment.
User avatar
in_reality
Posts: 4529
Joined: Fri Jul 12, 2013 6:13 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by in_reality »

In trying to pick funds to do factor investing, I became disillusioned and just found measurement of factor investing to be too fuzzy.

For example, take MOM.

If you look at two AQR momentum funds (AMOMX - large cap, ASMOX -small cap), you will find they underperformed a market portfolio over the course of their existence (Aug 2009 - Nov 2017) (1).

ASMOX - 14.11% CAGR
AMOMX - 14.43% CAGR
VTI - 15.05% CAGR

I understand it's a short time, and factors don't always pay a premium but it seems MOM did pay a premium in that time and according to portfolio visualizer, MOM had a 4.22% Annualized Return (2).

Also, factor regressions show ASMOX and AMOMX did have strong MOM loadings [0.35% and 0.33%] (3).

So, what I see is that the actual funds (not backtests) of a leader in the field, and maybe the leader for MOM as least judging by recommendations here, underperform during a period when exposure is said to have paid a premium. [You can note that other MOM funds such as MTUM did do much better, and that MOM paid the second highest factor premium during that time second only to BAB (bet against beta)]

This leads to my questions:

1) How can I know which fund to choose if loadings don't correspond to factor premium returns?
2) If 8 years is too short to gauge the actual loading on ASMOX and AMOMX, then really is it measuring anything at all?
3) If MOM actually didn't pay a premium during that time, but the risk premia measurements says it should have, then isn't something in the model broken?
4) If ASMOX and AMOMX did truly load on MOM, and MOM did actually return a premium but was dragged down by other exposure (i.e. negative QMJ or BAB), then is it really even possible to create a fund that targets your selected factor? I mean you can, but doing so might create other factor exposure that effects returns even more.
5) Could this simply be a reflection of costs where funds with a MOM loading in a time of MOM outperformance couldn't overcome their costs?
6) If actual returns are so seemingly dependent on fund construction and methodology, isn't it a huge leap of faith to assume that any particular fund will be able to capture what the academic backtests say is there?


Look, I understand 7 years of data is too short to answer the question of whether MOM over time will outperform. I do believe that factor investing can say whether or not MOM paid a premium in those 7 years and whether or not a particular fund had exposure to that factor.

7) If factor investing needs 15-20 years to explain if a factor paid a premium in that time, and if a fund had exposure to it, is it really explaining anything at all? Surely fund loadings will change overtime and what is to say that the fund methodology hasn't resulted in differing exposure. (this ties back into 4 above).
8) Is it wise to stick to the belief that a factor will eventually pay a premium if one can't be sure that choosing a fund that loads on that factor doesn't actually return the premium for other reasons (effects of other loadings resulting from the funds targeting, costs, whatever). How would you know whether it's fund construction or a period where the premium is absent.
9) Is factor investing like active management where it's certain that some funds will outperform, but you can't pick which ones in advance?



(1) https://www.portfoliovisualizer.com/bac ... ion3_3=100

(2) https://www.portfoliovisualizer.com/fac ... F15%2F2017

(3) https://www.portfoliovisualizer.com/fac ... sion=false
User avatar
patrick013
Posts: 3301
Joined: Mon Jul 13, 2015 7:49 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by patrick013 »

rkhusky wrote: Fri Dec 15, 2017 8:47 pm
patrick013 wrote: Fri Dec 15, 2017 8:43 pm The Factor Model I was looking at was a general model, not FF exactly.
My first read was the factor load was simply beta, pending further analysis.
R2 comes into play but just for the model accuracy. So the stats made a
little more sense like that.
That sounds like the model that preceded FF, which I believe was called CAPM, and did not address size or book/market.
Well additional regressions can be created with the model(s) to address
different concerns. It was a multi factor viewpoint but somewhat different.
Funny low volatility does so well as higher correlation or volatility(beta) would
give a better relationship to a specific factor.

Well someday as time allows. Really don't understand FF right now.
But I don't think they're the only ones doing it.
age in bonds, buy-and-hold, 10 year business cycle
User avatar
in_reality
Posts: 4529
Joined: Fri Jul 12, 2013 6:13 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by in_reality »

rkhusky wrote: Fri Dec 15, 2017 8:12 pm
triceratop wrote: Fri Dec 15, 2017 7:50 pm
rkhusky wrote: Fri Dec 15, 2017 7:47 pm The above is incorrect. The factors are constructed such that they are zero for the market and involve long and short positions. There is no small factor or value factor in the standard formulation. See. e.g., http://mba.tuck.dartmouth.edu/pages/fac ... loped.html
I fail to see your point. You absolutely can get exposure (with a loading strictly less than 1) to SmB and HmL using long-only (index, too) funds. People call these the Small and Value factors. What is your specific objection?
I've never said that one could not get factor exposure from a given fund. Confusion arises when the correct terms are not used. There is no small factor or value factor in the standard Fama French formulation. The factors are SmB and HmL.
One question on trying to read it right.

Looking at Exhibit 9 - Performance of Momentum Strategy Based Portfolios: U.S. 1926–2016 and U.K. 1900–2016, it shows winner-minus-loser (WML) portfolio returns equal to 7.4% per year (or 586 times as much). Wow!

Yet, they also cite "two significant caveats" - cost and volatility. In terms of volatility, they state:
The returns from a WML strategy are volatile, with the occasional very bad year when markets sharply reverse direction. This happened in 2009, when the dip in the long-run WML series for the U.K. represents an annual loss from the WML strategy of 25.4%. Looking back in time, investors would have been similarly whiplashed at the turning points in 1975, 2000, and 2003, when annual losses of similar magnitude occurred. These were all strong years for long-only equity performance. The relative performance of a pure momentum manager would thus have looked woeful and would have severely tried the patience even of long-term investors.
How does the rosy picture in Exhibit 9 [7.4% per year (or 586 times as much)] from using WML reconcile with their subsequent bleak statement on volatility?
grog
Posts: 622
Joined: Sat Jul 20, 2013 1:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by grog »

Looking at the straight CAPM in portfolio visualizer ...

NAESX (Vanguard Small Cap Index) has data going back to 1974. Portfolio visualizer gives me a beta of 1.12 and an alpha of -0.31%/yr (not significant). The rolling regression (36 mo) shows generally elevated beta and extended periods of positive and negative alpha with no apparent pattern. (Incidentally, if I throw in size and make it a two factor model, the beta drops all the way down close to one. The excess beta/volatility is just going to "size.")

VISVX (Vanguard Small Cap Value Index) only goes back to 1998. It shows beta=1.04 and alpha=2.20%/yr. Okay. But VISGX (which also started in 1998) also did pretty well with beta=1.15 and alpha=1.75%/yr. NAESX also shows alpha over that more limited period. The bulk of the outperformance seems to be concentrated in the early 2000s. Shifting to 2003-2017, VISVX has beta=1.22 and alpha= -0.62%/yr.

At a glance, I'm not seeing what all the fuss is about. On the other hand, if high beta tends to underperform more generally (I know that empirically CAPM is generally too steep, but I don't know if that's the case over this modestly elevated range), maybe small caps achieving zero alpha with a beta of 1.2 is something of an accomplishment. But then why not just tilt to low beta and possibly leverage up?
User avatar
triceratop
Posts: 5838
Joined: Tue Aug 04, 2015 8:20 pm
Location: la la land

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by triceratop »

grog wrote: Sat Dec 16, 2017 3:00 am Looking at the straight CAPM in portfolio visualizer ...

NAESX (Vanguard Small Cap Index) has data going back to 1974. Portfolio visualizer gives me a beta of 1.12 and an alpha of -0.31%/yr (not significant). The rolling regression (36 mo) shows generally elevated beta and extended periods of positive and negative alpha with no apparent pattern. (Incidentally, if I throw in size and make it a two factor model, the beta drops all the way down close to one. The excess beta/volatility is just going to "size.")

VISVX (Vanguard Small Cap Value Index) only goes back to 1998. It shows beta=1.04 and alpha=2.20%/yr. Okay. But VISGX (which also started in 1998) also did pretty well with beta=1.15 and alpha=1.75%/yr. NAESX also shows alpha over that more limited period. The bulk of the outperformance seems to be concentrated in the early 2000s. Shifting to 2003-2017, VISVX has beta=1.22 and alpha= -0.62%/yr.

At a glance, I'm not seeing what all the fuss is about. On the other hand, if high beta tends to underperform more generally (I know that empirically CAPM is generally too steep, but I don't know if that's the case over this modestly elevated range), maybe small caps achieving zero alpha with a beta of 1.2 is something of an accomplishment. But then why not just tilt to low beta and possibly leverage up?
Is there a particular reason you are using strictly CAPM rather than a multifactor model? How much of the return does each explain?
"To play the stock market is to play musical chairs under the chord progression of a bid-ask spread."
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

in_reality wrote: Fri Dec 15, 2017 9:19 pm
rkhusky wrote: Fri Dec 15, 2017 8:12 pm
triceratop wrote: Fri Dec 15, 2017 7:50 pm
rkhusky wrote: Fri Dec 15, 2017 7:47 pm The above is incorrect. The factors are constructed such that they are zero for the market and involve long and short positions. There is no small factor or value factor in the standard formulation. See. e.g., http://mba.tuck.dartmouth.edu/pages/fac ... loped.html
I fail to see your point. You absolutely can get exposure (with a loading strictly less than 1) to SmB and HmL using long-only (index, too) funds. People call these the Small and Value factors. What is your specific objection?
I've never said that one could not get factor exposure from a given fund. Confusion arises when the correct terms are not used. There is no small factor or value factor in the standard Fama French formulation. The factors are SmB and HmL.
One question on trying to read it right.

Looking at Exhibit 9 - Performance of Momentum Strategy Based Portfolios: U.S. 1926–2016 and U.K. 1900–2016, it shows winner-minus-loser (WML) portfolio returns equal to 7.4% per year (or 586 times as much). Wow!

Yet, they also cite "two significant caveats" - cost and volatility. In terms of volatility, they state:
The returns from a WML strategy are volatile, with the occasional very bad year when markets sharply reverse direction. This happened in 2009, when the dip in the long-run WML series for the U.K. represents an annual loss from the WML strategy of 25.4%. Looking back in time, investors would have been similarly whiplashed at the turning points in 1975, 2000, and 2003, when annual losses of similar magnitude occurred. These were all strong years for long-only equity performance. The relative performance of a pure momentum manager would thus have looked woeful and would have severely tried the patience even of long-term investors.
How does the rosy picture in Exhibit 9 [7.4% per year (or 586 times as much)] from using WML reconcile with their subsequent bleak statement on volatility?
Perhaps they are saying that, in the past, a WML strategy did pay off long term, but that the returns were volatile and investors following the strategy had to be very patient. Practitioners of factor investing, like Larry Swedroe, say this all the time. You may have to endure a decade of underperformance before the strategy pays off, if it ever does in your investing lifetime. In order to get extra return you have to take extra risk. And sometimes the risk shows up. If factor investing guaranteed higher returns over any given time period, it wouldn't be risk, would it.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

Caveat: I'm not a practitioner of factor investing, so have not gotten into the nuts and bolts of actually trying to set up a portfolio.
in_reality wrote: Fri Dec 15, 2017 8:52 pm This leads to my questions:

1) How can I know which fund to choose if loadings don't correspond to factor premium returns?
Loadings do not correspond to premiums, they are completely separate calculations.
2) If 8 years is too short to gauge the actual loading on ASMOX and AMOMX, then really is it measuring anything at all?
The methodology works for any time period, you just need to test the model on out-of-sample data to see how well it is really describing reality. The r2 statistic and others tell you how well it is even modeling the in-sample data
3) If MOM actually didn't pay a premium during that time, but the risk premia measurements says it should have, then isn't something in the model broken?
4) If ASMOX and AMOMX did truly load on MOM, and MOM did actually return a premium but was dragged down by other exposure (i.e. negative QMJ or BAB), then is it really even possible to create a fund that targets your selected factor? I mean you can, but doing so might create other factor exposure that effects returns even more.
If the model has other factors, then you can't disregard them when looking at fund performance, unless the loadings for the other factors are zero
5) Could this simply be a reflection of costs where funds with a MOM loading in a time of MOM outperformance couldn't overcome their costs?That's possible. One could try to find fund performance before costs.
6) If actual returns are so seemingly dependent on fund construction and methodology, isn't it a huge leap of faith to assume that any particular fund will be able to capture what the academic backtests say is there?
There is always the risk that the future will not replicate the past.

Look, I understand 7 years of data is too short to answer the question of whether MOM over time will outperform. I do believe that factor investing can say whether or not MOM paid a premium in those 7 years and whether or not a particular fund had exposure to that factor.
Factor analysis can tell you if a particular fund has exposure to a factor and the extent to which the model explains the return data. Return analysis tells you whether a factor returned a premium over a given time period.
7) If factor investing needs 15-20 years to explain if a factor paid a premium in that time, and if a fund had exposure to it, is it really explaining anything at all? Surely fund loadings will change overtime and what is to say that the fund methodology hasn't resulted in differing exposure. (this ties back into 4 above).
Factor analysis does not say anything about whether a factor paid a premium, and even less on whether the factor will pay a premium in the future. Out-of-sample testing provides a level of confidence in the methodology.
8) Is it wise to stick to the belief that a factor will eventually pay a premium if one can't be sure that choosing a fund that loads on that factor doesn't actually return the premium for other reasons (effects of other loadings resulting from the funds targeting, costs, whatever). How would you know whether it's fund construction or a period where the premium is absent.
Good question. Something each investor will have to answer for themselves or trust the experts.
9) Is factor investing like active management where it's certain that some funds will outperform, but you can't pick which ones in advance?
Practitioners point to behavioral explanations for why some factor premiums will continue to exist and also look at other markets and out-of-sample testing to provide additional confidence. But in the end, investing involves risk, and sometimes the risk shows up and you don't receive the long term averages over your investing lifetime.
User avatar
nisiprius
Advisory Board
Posts: 52105
Joined: Thu Jul 26, 2007 9:33 am
Location: The terrestrial, globular, planetary hunk of matter, flattened at the poles, is my abode.--O. Henry

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nisiprius »

I posted this once before and didn't think I got a full answer, so I'll try again. Very likely I have done something wrong, but I don't know what it is.

1) I downloaded the data for the HmL factor from the Ken French data library website, while for "large stocks" I used the SBBI series, which is the S&P 500 and its predecessors. (I also tried using the Fama-French "Benchmark Portfolio: B/M" which gives almost identical results). The range of years covered is 1927 through 2015. 1927 is the earliest year in the Ken French data set, 2015 is the year in which I did the work. I'll post the actual data sets below so that there's no question about exactly what I used.

HmL is the factor itself, the difference, not an asset class; conceptually it is a long-short portfolio, without any costs, that extracts the value factor in a pure form--it has the highest possible loading on the value factor.

The source of the data I used is on this web page, http://mba.tuck.dartmouth.edu/pages/fac ... brary.html
The data set I used is named, and the link to it, is: Fama/French Benchmark Factors -- Annual

2) I investigated the possible portfolios formed by mixing different percentages of large stocks and the pure value factor, by calculating the mean, standard deviation, and correlation between large stocks and the value factor, calculating the efficient frontier (the "Markowitz bullet"), and the tangent portfolio.

Here are the results. Although the correlation between large stocks and the pure value factor is low, 0.163, adding the pure value factor does not improve the portfolio, and the optimum percentage over that stated time period would have been a mix consisting of 100% large stocks and 0% value factor. That is, based on the data I used, adding the "pure value factor" to a portfolio of large stocks made the portfolio worse.

Image


The reason is that the return of the value factor was much lower than the return of large stocks, 3.566% compared to 11.951%. Although its standard deviation was also lower, it wasn't enough lower to matter. In fact, the return of the value factor, 3.566, was virtually the same as that of the riskless asset (I used Treasury bills), with the result that the Sharpe ratio was only 0.006 (!) for the value factor, versus 0.424 for large stocks. In general, in order for a diversifier to improve a target asset, it isn't enough for the correlation coefficient to be less than 1.0, or "low," has to be lower than result of dividing the lower Sharpe ratio by the higher one. In this case, it needed to be lower than 0.006 / 0.424 = 0.014, and it wasn't.

To put it another way, if an diversifier isn't earning more than Treasury bills, it can only improve a target asset if it has negative correlation.

As I say, my assumption is that I'm missing something, misinterpreting something, or using the wrong French data library series... but I'm not sure what it is.

The data:
Value factor (HmL column)

Code: Select all

          Rm -Rf          SMB            HML
  1927      30.85        -2.88          -5.53
  1928      34.88         3.97          -9.16
  1929     -17.90       -28.68           12.8
  1930     -30.77        -4.84         -13.63
  1931     -44.80         4.45         -15.22
  1932      -9.45         4.21           5.64
  1933      56.08        50.72          15.77
  1934       4.42        25.45         -29.97
  1935      44.35        12.56           6.88
  1936      32.44        15.46          28.87
  1937     -34.83       -13.92          -4.89
  1938      27.95        11.66          -12.8
  1939       2.95         5.28         -17.25
  1940      -6.93         1.38          -1.61
  1941     -10.14        -4.95          12.16
  1942      15.75         5.54          19.39
  1943      28.23        30.65          34.36
  1944      21.12         16.9          17.17
  1945      37.72        26.25          13.72
  1946      -6.08        -4.01           2.52
  1947       3.22        -7.51           9.14
  1948       1.33        -8.97            3.3
  1949      19.23         3.17          -3.54
  1950      28.26         2.01          26.63
  1951      19.09        -4.98          -4.84
  1952      11.89        -6.55            3.1
  1953      -1.15        -1.35          -7.78
  1954      49.28        -1.94          24.88
  1955      23.76        -6.25           5.39
  1956       5.83        -0.67          -2.41
  1957     -13.27        -2.44          -6.25
  1958      43.46        14.27          12.56
  1959       9.79         5.64           0.84
  1960      -1.73        -2.41          -5.22
  1961      24.88         0.42           5.56
  1962     -12.28        -9.25          10.33
  1963      17.92        -6.09          15.42
  1964      12.66        -1.78           9.72
  1965      10.58        22.82           5.79
  1966     -13.54         2.78          -0.94
  1967      24.00        50.61          -9.15
  1968       9.02        24.23          18.06
  1969     -17.20       -13.99         -10.71
  1970      -5.65       -11.47          21.58
  1971      11.65         6.26         -11.39
  1972      12.97       -12.24           2.09
  1973     -24.69       -23.94          17.98
  1974     -35.70        -0.61           9.36
  1975      32.35        15.04           8.63
  1976      21.90         13.8          24.01
  1977      -8.19        23.24           7.79
  1978       1.04        14.16           0.47
  1979      13.33         20.4          -2.21
  1980      22.10         5.59         -24.56
  1981     -18.05          7.3          24.57
  1982      10.78         8.79          13.16
  1983      14.41        13.94          18.86
  1984      -4.76        -8.61          18.63
  1985      24.63        -0.92           1.16
  1986      10.41          -10           9.99
  1987      -3.51       -10.39          -2.54
  1988      11.55         6.72          13.78
  1989      20.51       -12.01          -5.64
  1990     -13.84        -14.4          -10.6
  1991      29.10        16.51         -15.08
  1992       6.41         7.78          23.05
  1993       8.37         7.47          16.95
  1994      -4.11         0.39          -0.08
  1995      31.04        -6.94          -3.46
  1996      16.25        -1.86           0.22
  1997      26.07        -3.73          11.14
  1998      19.41       -23.29         -15.03
  1999      20.21        11.66          -39.4
  2000     -16.71        -5.69          21.39
  2001     -14.78        28.41          27.24
  2002     -22.91         4.35           3.71
  2003      30.74        28.08          15.14
  2004      10.69         6.32          13.21
  2005       3.21         -2.7           3.71
  2006      10.58         1.04          11.91
  2007       0.83        -7.01         -21.55
  2008     -38.39         0.16          -9.08
  2009      29.05        17.74          23.66
  2010      17.68        12.58           4.31
  2011       0.97        -5.21          -7.95
  2012      16.01        -0.16           7.72
  2013      35.15         7.27          -1.83
  2014      11.35        -7.86          -2.35
  2015      -0.20        -5.53         -11.49
Large stocks:

Code: Select all

	Large
	company
Year ending	stocks
1925	
1926	11.62%
1927	37.49%
1928	43.61%
1929	-8.42%
1930	-24.90%
1931	-43.34%
1932	-8.19%
1933	53.99%
1934	-1.44%
1935	47.67%
1936	33.92%
1937	-35.03%
1938	31.12%
1939	-0.41%
1940	-9.78%
1941	-11.59%
1942	20.34%
1943	25.90%
1944	19.75%
1945	36.44%
1946	-8.07%
1947	5.71%
1948	5.50%
1949	18.79%
1950	31.71%
1951	24.02%
1952	18.37%
1953	-0.99%
1954	52.62%
1955	31.56%
1956	6.56%
1957	-10.78%
1958	43.36%
1959	11.96%
1960	0.47%
1961	26.89%
1962	-8.73%
1963	22.80%
1964	16.48%
1965	12.45%
1966	-10.06%
1967	23.98%
1968	11.06%
1969	-8.50%
1970	4.01%
1971	14.31%
1972	18.98%
1973	-14.66%
1974	-26.47%
1975	37.20%
1976	23.84%
1977	-7.18%
1978	6.56%
1979	18.44%
1980	32.42%
1981	-4.91%
1982	21.41%
1983	22.51%
1984	6.27%
1985	32.16%
1986	18.47%
1987	5.23%
1988	16.81%
1989	31.49%
1990	-3.17%
1991	30.55%
1992	7.67%
1993	9.99%
1994	1.31%
1995	37.43%
1996	23.07%
1997	33.36%
1998	28.58%
1999	21.04%
2000	-9.11%
2001	-11.88%
2002	-22.10%
2003	28.70%
2004	10.88%
2005	4.91%
2006	15.79%
2007	5.49%
2008	-37.00%
2009	26.46%
2010	15.05%
2011	2.11%
2012	16.00%
2013	32.39%
2014	13.69%
2015	1.25%
Last edited by nisiprius on Sat Dec 16, 2017 9:42 am, edited 2 times in total.
Annual income twenty pounds, annual expenditure nineteen nineteen and six, result happiness; Annual income twenty pounds, annual expenditure twenty pounds ought and six, result misery.
User avatar
siamond
Posts: 6003
Joined: Mon May 28, 2012 5:50 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by siamond »

nisiprius wrote: Sat Dec 16, 2017 9:07 am I posted this once before and didn't think I got a satisfactory reply, so I'll try again. Very likely I have done something wrong, but I don't know what it is.
Er, the factor as you defined it (HmL) is a difference (a premium if you wish), not an absolute return. Hence your problem. If large stocks return 10% and small stocks return 13%, then the difference is 3%, which is way lower than 10%. And yet the small stocks investor would have been way richer.

PS. I wouldn't mix Fama-French data with SBBI data. They define 'size' in a very different manner.
User avatar
nisiprius
Advisory Board
Posts: 52105
Joined: Thu Jul 26, 2007 9:33 am
Location: The terrestrial, globular, planetary hunk of matter, flattened at the poles, is my abode.--O. Henry

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nisiprius »

PS. I wouldn't mix Fama-French data with SBBI data. They define 'size' in a very different manner.
I get essentially the same results, as I stated earlier, if I use the Fama French Benchmark Portfolio B/M series instead of the SBBI Large Stocks series. See below.
siamond wrote: Sat Dec 16, 2017 9:19 amEr, the factor as you defined it (HmL) is a difference (a premium if you wish), not an absolute return. Hence your problem. If large stocks return 10% and small stocks return 13%, then the difference is 3%, which is way lower than 10%. And yet the small stocks investor would have been way richer.
The tenor of much of the discussion above is that factor-based investing has moved on and is no longer based on long-only portfolios, but that factor does not mean using a stock subclass that is rich in a desired factor, it means using a long-short portfolio--such as are embodied in most? all? of the AQR factor-oriented funds. So in practice, in the real world as I understand it, investors who use these funds really are diversifying their traditional stock fund with funds that realize "the factor as you defined it"--a difference between the earnings of two different stock portfolios.

Here's a repeat of the analysis above, using the Fama-French Benchmark Portfolio B/M series in place of SBBI Large Stocks.

Image
Annual income twenty pounds, annual expenditure nineteen nineteen and six, result happiness; Annual income twenty pounds, annual expenditure twenty pounds ought and six, result misery.
User avatar
siamond
Posts: 6003
Joined: Mon May 28, 2012 5:50 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by siamond »

nisiprius wrote: Sat Dec 16, 2017 9:26 am
PS. I wouldn't mix Fama-French data with SBBI data. They define 'size' in a very different manner.
I get essentially the same results, as I stated earlier, if I use the Fama French Benchmark Portfolio B/M series instead of the SBBI Large Stocks series.
In this specific case, you're fine, but if you were to do the same with small stocks, you'll go in a world of troubles. So I *always* try to keep my data sources consistent. Too many differences in core definitions.
nisiprius wrote: Sat Dec 16, 2017 9:26 am
siamond wrote: Sat Dec 16, 2017 9:19 amEr, the factor as you defined it (HmL) is a difference (a premium if you wish), not an absolute return. Hence your problem. If large stocks return 10% and small stocks return 13%, then the difference is 3%, which is way lower than 10%. And yet the small stocks investor would have been way richer.
The tenor of much of the discussion above is that factor-based investing has moved on and is no longer based on long-only portfolios, but that factor means a long-short portfolio, just as are embodied in most of the AQR factor-oriented funds. So in practice, in the real world as I understand it, investors who use these funds really are diversifying their traditional stock fund with funds that realize "the factor as you defined it"--a difference between the earnings of two different stock portfolios.
Yes, it's fine to define 'factor' as a difference, but then you start using this quantity as an absolute return for your efficient frontier math, and well, it's not. In other words, you seem to be mixing up relative and absolute quantities as if they were both absolute. This doesn't work. In the real world, investors get the factor premium PLUS the base return. Or did I completely misunderstand what you are doing?
User avatar
nisiprius
Advisory Board
Posts: 52105
Joined: Thu Jul 26, 2007 9:33 am
Location: The terrestrial, globular, planetary hunk of matter, flattened at the poles, is my abode.--O. Henry

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nisiprius »

siamond wrote: Sat Dec 16, 2017 9:35 am...In the real world, investors get the factor premium PLUS the base return. Or did I completely misunderstand what you are doing?...
Not sure about the terminology. I'd say:

A long-only portfolio of large blend stocks, e.g an S&P 500 fund or (if you want to be purer) the Vanguard Large-Cap Stock Index Fund, gives you the base return of large-cap stocks without any factor premium.

A long-only portfolio of selected large value stocks, e.g. the Vanguard Value Index Fund or the DFA U.S. Large Cap Value Portfolio gives you the base return of large-cap stocks plus a portion of the value premium, depending on how "value-y" the selection is.

A long-short portfolio that is long on large value stocks and short on large growth stocks does not give you the base return of large-cap stocks, because it's been subtracted out. It gives you the only the difference, the amount by which the return of large-cap value stocks exceeded that of large-cap growth, the value premium. It represents the pure factor itself.

You can get a partial amount of the value premium in two ways: by an allocation to a subset of stocks that has a loading on the value premium, or by allocating part to large-cap blend and part to a long-short portfolio with value on the long side and growth on the short side. In theory, these should be the same. Vanguard Value Index, VIVAX, has an 0.23 loading on value, according to PortfolioVisualizer, so I think that means that if you had a perfect pure value factor long-short large-cap fund, a portfolio of 100% VIVAX, and a portfolio of 87% VIVAX/23% pure value would actually be essentially the same.

The reason for wanting the long-short portfolios would be a belief that a value tilt is advantageous, and that it is better to have a larger tilt than you can achieve in a long-only portfolio.

However, if adding the value factor (a idea long-short value portfolio) to large stocks doesn't improve it, then that raises some questions.
Annual income twenty pounds, annual expenditure nineteen nineteen and six, result happiness; Annual income twenty pounds, annual expenditure twenty pounds ought and six, result misery.
User avatar
siamond
Posts: 6003
Joined: Mon May 28, 2012 5:50 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by siamond »

nisiprius wrote: Sat Dec 16, 2017 10:03 amA long-short portfolio that is long on large value stocks and short on large growth stocks does not give you the base return of large-cap stocks, because it's been subtracted out. It gives you the only the difference, the amount by which the return of large-cap value stocks exceeded that of large-cap growth, the value premium. It represents the pure factor itself. [..]

The reason for wanting the long-short portfolios would be a belief that a value tilt is advantageous, and that it is better to have a larger tilt than you can achieve in a long-only portfolio.
Ah ok, sorry, I missed the long-short part of your post. Well, it seems to me that all what you've shown is that a long-short portfolio is a terrible idea for a long-term investor. Even in presence of a nice premium between the two parts of the portfolio. This only makes sense, right? The market goes up on average, actually pretty steeply due to compounding, even the not-so-great asset classes like SCG. It's NOT a good idea to bet against the market on the long term. I would see a 'short' investment as a tactical short-term speculative move, nothing else. This being said, I have zero experience with shorting, and never really thought about it. So maybe I'll shut up and let better qualified people speak up! :wink:
lack_ey
Posts: 6701
Joined: Wed Nov 19, 2014 10:55 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by lack_ey »

nisiprius wrote: Sat Dec 16, 2017 9:07 am 2) I investigated the possible portfolios formed by mixing different percentages of large stocks and the pure value factor, by calculating the mean, standard deviation, and correlation between large stocks and the value factor, calculating the efficient frontier (the "Markowitz bullet"), and the tangent portfolio.

Here are the results. Although the correlation between large stocks and the pure value factor is low, 0.163, adding the pure value factor does not improve the portfolio, and the optimum percentage over that stated time period would have been a mix consisting of 100% large stocks and 0% value factor. That is, based on the data I used, adding the "pure value factor" to a portfolio of large stocks made the portfolio worse.

https://s8.postimg.org/u7739y005/Captur ... 4_a._m.png


The reason is that the return of the value factor was much lower than the return of large stocks, 3.566% compared to 11.951%. Although its standard deviation was also lower, it wasn't enough lower to matter. In fact, the return of the value factor, 3.566, was virtually the same as that of the riskless asset (I used Treasury bills), with the result that the Sharpe ratio was only 0.006 (!) for the value factor, versus 0.424 for large stocks. In general, in order for a diversifier to improve a target asset, it isn't enough for the correlation coefficient to be less than 1.0, or "low," has to be lower than result of dividing the lower Sharpe ratio by the higher one. In this case, it needed to be lower than 0.006 / 0.424 = 0.014, and it wasn't.

To put it another way, if an diversifier isn't earning more than Treasury bills, it can only improve a target asset if it has negative correlation.

As I say, my assumption is that I'm missing something, misinterpreting something, or using the wrong French data library series... but I'm not sure what it is.
Here's the problem:

The HmL series is a differential return and involves 0 net investment, unlike stocks or most other assets you mix using these efficient frontiers. You're penalizing allocations using HmL.

For illustration, consider the 100% HmL portfolio in our frictionless world. You get a 4% return from the long/short formulation, but you haven't actually invested in anything. The money to invest in the long side comes from selling things on the short side. So maybe this asset allocation is really +100% long stocks, -100% long stocks (i.e. 100% short stocks), 100% cash. So 100% - 100% + 100% = 100%. That checks out. Without that 100% cash at the end, this is only 0% allocated.

On the curve you present, an [80% market, 20% HmL] portfolio is 80% stocks, 20% stocks, and -20% stocks for 80% net investment, with somehow 20% being thrown into the void and counting for nothing. You need to give it a 20% cash allocation. Or better yet, in an actual portfolio, you could use 20% bonds and get some term premium that's also mostly uncorrelated with equities and on average over the period had excess returns above cash.

Normally when you're mixing assets on and off the efficient frontier, these are long investments that represent actual allocations of money. With 50% stocks, 40% bonds, 5% gold, 5% bitcoin (lol?), you have put money in different places and have nothing left over. A differential return like HmL is on the other hand "free" in terms of budgeting your 100% to dole out. In practice there are implementation costs to an actual HmL formulation, though you can get value loadings for free in a sense with a long-only allocation (to a certain extent). My 80% stocks with 20% HmL is realistic in the sense that you can easily get a stock allocation that's 1 market beta and 0.25 value.
grog
Posts: 622
Joined: Sat Jul 20, 2013 1:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by grog »

triceratop wrote: Sat Dec 16, 2017 4:31 am
grog wrote: Sat Dec 16, 2017 3:00 am Looking at the straight CAPM in portfolio visualizer ...

NAESX (Vanguard Small Cap Index) has data going back to 1974. Portfolio visualizer gives me a beta of 1.12 and an alpha of -0.31%/yr (not significant). The rolling regression (36 mo) shows generally elevated beta and extended periods of positive and negative alpha with no apparent pattern. (Incidentally, if I throw in size and make it a two factor model, the beta drops all the way down close to one. The excess beta/volatility is just going to "size.")

VISVX (Vanguard Small Cap Value Index) only goes back to 1998. It shows beta=1.04 and alpha=2.20%/yr. Okay. But VISGX (which also started in 1998) also did pretty well with beta=1.15 and alpha=1.75%/yr. NAESX also shows alpha over that more limited period. The bulk of the outperformance seems to be concentrated in the early 2000s. Shifting to 2003-2017, VISVX has beta=1.22 and alpha= -0.62%/yr.

At a glance, I'm not seeing what all the fuss is about. On the other hand, if high beta tends to underperform more generally (I know that empirically CAPM is generally too steep, but I don't know if that's the case over this modestly elevated range), maybe small caps achieving zero alpha with a beta of 1.2 is something of an accomplishment. But then why not just tilt to low beta and possibly leverage up?
Is there a particular reason you are using strictly CAPM rather than a multifactor model? How much of the return does each explain?
1) The R^2 for CAPM was about 0.75 for NAESX, VISVX, and VISGX. The three factor goes up to 0.87 for NAESX and 0.92 for the other two. It “explains” more of the returns (since you’re adding parameters and capturing the styles), but that doesn’t necessarily mean the returns were always superior since the size and value premiums aren’t always positive.

2) If a factor gets left out of the one factor model and that factor has systematically higher returns, it should show up as alpha.

3) Small caps look to be somewhat high beta (using the classic beta definition) and do not seem to have systematically positive alpha, at least not that I can tell. So this would mean higher volatility with additional return roughly proportionate to CAPM.

4) In other words, small cap ought to be roughly comparable to a portfolio of large and mid caps with a beta in the 1.1-1.2 range. Unless those alternative strategies have systematically negative alpha (maybe that’s the case?). In practice though I’m not sure what should be used for such a comparison.

5). One possible comparison might be mid cap growth since it has that similar, modestly elevated beta (large cap growth only has a beta around 1.0). In portfolio visualizer, if I backtest the generic “Mid cap growth” and “small cap,” it looks like small cap has done better since 1972. But starting from the early 80s, for lump sum and DCA they have virtually identical returns and volatility.
User avatar
nisiprius
Advisory Board
Posts: 52105
Joined: Thu Jul 26, 2007 9:33 am
Location: The terrestrial, globular, planetary hunk of matter, flattened at the poles, is my abode.--O. Henry

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nisiprius »

lack_ey wrote: Sat Dec 16, 2017 10:44 am...Here's the problem:

The HmL series is a differential return and involves 0 net investment, unlike stocks or most other assets you mix using these efficient frontiers. You're penalizing allocations using HmL.

For illustration, consider the 100% HmL portfolio in our frictionless world. You get a 4% return from the long/short formulation, but you haven't actually invested in anything. The money to invest in the long side comes from selling things on the short side. So maybe this asset allocation is really +100% long stocks, -100% long stocks (i.e. 100% short stocks), 100% cash. So 100% - 100% + 100% = 100%. That checks out. Without that 100% cash at the end, this is only 0% allocated.

On the curve you present, an [80% market, 20% HmL] portfolio is 80% stocks, 20% stocks, and -20% stocks for 80% net investment, with somehow 20% being thrown into the void and counting for nothing. You need to give it a 20% cash allocation. Or better yet, in an actual portfolio, you could use 20% bonds and get some term premium that's also mostly uncorrelated with equities and on average over the period had excess returns above cash.

Normally when you're mixing assets on and off the efficient frontier, these are long investments that represent actual allocations of money. With 50% stocks, 40% bonds, 5% gold, 5% bitcoin (lol?), you have put money in different places and have nothing left over. A differential return like HmL is on the other hand "free" in terms of budgeting your 100% to dole out. In practice there are implementation costs to an actual HmL formulation, though you can get value loadings for free in a sense with a long-only allocation (to a certain extent). My 80% stocks with 20% HmL is realistic in the sense that you can easily get a stock allocation that's 1 market beta and 0.25 value.
So, suppose a hypothetical Vanguard in 1926 decided to launch an S&P-500-like index fund, and a hypothetic AQR in 1926 decided to launch a long-short mutual fund that was intended to help investors improve a baseline stock market portfolio by making use of the value factor. Say we use SBBI Large Company Stocks to represent the untilted stock fund.

What would you suggest, using Fama-French or other data going back to 1926 or thereabouts, as something that can be calculated to stand in for an idealized long-short value factor fund? 150% * "B/H" - 50% x "B/L," 200% x "B/H" - 100% x "B/L???"
Annual income twenty pounds, annual expenditure nineteen nineteen and six, result happiness; Annual income twenty pounds, annual expenditure twenty pounds ought and six, result misery.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

nisiprius wrote: Sat Dec 16, 2017 10:03 am The reason for wanting the long-short portfolios would be a belief that a value tilt is advantageous, and that it is better to have a larger tilt than you can achieve in a long-only portfolio.
An academic reason for using a long-short methodology is that the market return is subtracted out and so you reduce the mixing of factors in the extraction of the factor weights, making the basis vectors closer to orthogonal.
lack_ey
Posts: 6701
Joined: Wed Nov 19, 2014 10:55 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by lack_ey »

nisiprius wrote: Sat Dec 16, 2017 11:02 am So, suppose a hypothetical Vanguard in 1926 decided to launch an S&P-500-like index fund, and a hypothetic AQR in 1926 decided to launch a long-short mutual fund that was intended to help investors improve a baseline stock market portfolio by making use of the value factor. Say we use SBBI Large Company Stocks to represent the untilted stock fund.

What would you suggest, using Fama-French or other data going back to 1926 or thereabouts, as something that can be calculated to stand in for an idealized long-short value factor fund? 150% * "B/H" - 50% x "B/L," 200% x "B/H" - 100% x "B/L???"
By long-short here what are you intending for the beta to be?

The hypothetical frictionless market neutral value factor fund as a baseline I think would be 100% long the value side, 100% short growth stocks, and hold 100% cash in the account. So the return would be HmL + cash. Most actual market neutral funds hold a lot of cash in the account just like that, including AQR's.

If still dealing in hypotheticals, a frictionless fund long-short fund unconstrained on our nonexistent costs might let's say elect for a load of 1 Mkt-RF and 1 HmL all in one fund. In theory we can accomplish this through 200% long stocks and 100% short stocks, doing the same 100% HmL portfolio as before and adding a 100% market allocation to that.

In the real world with costs, that is a terrible construction as there's no reason you should be shorting stocks you also own long (the short side of HmL includes stocks owned long in the market slice)—just reduce the weighting to zero, and continue shorting as necessary. In our hypothetical frictionless scenario, there could be room for some optimization by voltatility-adjusting the ratios of HmL and market. For example even under risk parity we would invest more in HmL than market because HmL is less volatile than the market. Though that requires a forward estimate of relative risks between the two. For even more optimal you'd need to know the returns. Hence I suggest just 1 market and 1 value as a starting point from our 1926 ignorance. I suspect that even then you could gather that the market was more likely to be volatile than HmL and maybe you might pick 1 market and 1.5 value or something like that, but oh well.

And in the real world, AQR and many other managers would not want to keep strictly 100% long, 100% short (+100% cash) allocations for market neutral funds. They're frequently more leveraged than that and frequently scale leverage with estimated forward volatility (less leverage when the portfolio is estimated to be more volatile; if nothing else you could scale using VIX or something in the formula, as market vol is related).
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

nisiprius wrote: Sat Dec 16, 2017 10:03 am The reason for wanting the long-short portfolios would be a belief that a value tilt is advantageous, and that it is better to have a larger tilt than you can achieve in a long-only portfolio.

However, if adding the value factor (a idea long-short value portfolio) to large stocks doesn't improve it, then that raises some questions.
No one wants a pure value portfolio (long-short) or even a portion of their portfolio to be pure value, because beta_market is zero for a pure value allocation and the expected market return is much larger than the expected value premium. What you want, if you can get it cheap enough, is for beta_market and beta_hml to both be as large as possible, with beta_market the higher priority. For example, if I had a portfolio that consisted of a 50/50 mix of Total Stock Market and a pure value fund (long-short), then beta_market would be about 0.5 and beta_hml would be about 0.5. Whereas, Total Stock Market by itself has beta_market=1 and beta_hml=0. Because, the expected Market>>HmL, Total Stock Market by itself has higher expected return. For those that believe in a future value premium, a fund with beta_market=1 and beta_hml=0.5 would be better than Total Stock Market. DFA Large Value has beta_market=1.08 and beta_hml=0.59, which would be even better.

Adding a positive value factor to large balanced stocks would only improve the return if the value premium was positive.

edit: the term pure value above should really be pure HmL
Last edited by rkhusky on Sat Dec 16, 2017 6:59 pm, edited 1 time in total.
User avatar
Robert T
Posts: 2803
Joined: Tue Feb 27, 2007 8:40 pm
Location: 1, 0.2, 0.4, 0.5
Contact:

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by Robert T »

nisiprius wrote: Sat Dec 16, 2017 10:03 am Vanguard Value Index, VIVAX, has an 0.23 loading on value, according to PortfolioVisualizer, so I think that means that if you had a perfect pure value factor long-short large-cap fund, a portfolio of 100% VIVAX, and a portfolio of 87% VIVAX/23% pure value would actually be essentially the same.
Just to note - that's not what it means. Presume you mean VTSMX or VFINX for the second portfolio. Even so, still not what it means. The second portfolio is equivalent to an 87:23 stock:bond portfolio, with a value tilt. The first is a 100% stock portfolio with a value tilt.
.
User avatar
in_reality
Posts: 4529
Joined: Fri Jul 12, 2013 6:13 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by in_reality »

rkhusky wrote: Sat Dec 16, 2017 6:57 am Caveat: I'm not a practitioner of factor investing, so have not gotten into the nuts and bolts of actually trying to set up a portfolio.
in_reality wrote: Fri Dec 15, 2017 8:52 pm This leads to my questions:

1) How can I know which fund to choose if loadings don't correspond to factor premium returns?
Loadings do not correspond to premiums, they are completely separate calculations.
I understand.

But to calculate returns you can take the loadings and multiply them by the factor returns, add in the annual alpha for anything unexplained and you are very, very close to actual annual returns. It's not surprising that it's not a 100% match since alpha is actual in a range.

If I look at actual lifetime returns of QSMNX (AQR small multi factor targeting size, value, momentum and quality), it underperforms other small cap funds (VB, IJS ) and other small cap value funds (VBR, SFSNX Schwab fundamental small). That underperformance isn't explained by the market, size, value, mom, quality, or bet against beta loadings. It's explained by negative alpha.

That's the same with the AQR momentum funds I previously posted about.

So even if I believe small size, value, momentum and quality will pay premiums if held long enough, I'm not at all sure that selecting a fund to target those premiums will benefit returns.

Please note, all the p-values for the loadings on QSMNX were statistically significant. It benefitted from MOM and quality exposure. It got hurt by size and value. The real difference though was in the about -5% yearly annual alpha. IJS did better even though it had about the same size and mom exposure but more value (which should hurt) because of the alpha.

So I don't see the real factor loading and returns explaining performance.

rkhusky wrote: Sat Dec 16, 2017 6:57 am
in_reality wrote: Fri Dec 15, 2017 8:52 pm 4) If ASMOX and AMOMX did truly load on MOM, and MOM did actually return a premium but was dragged down by other exposure (i.e. negative QMJ or BAB), then is it really even possible to create a fund that targets your selected factor? I mean you can, but doing so might create other factor exposure that effects returns even more.
If the model has other factors, then you can't disregard them when looking at fund performance, unless the loadings for the other factors are zero
That's a problem with the long term evidence chart I think. It's looking at factors in isolation. If MOM looks juicy, but implementation in an actual fund results in changing other loadings (or negative alpha) which can offset and counter-act the MOM premium, then are you really setting yourself up for better returns by concentrating your portfolio on MOM?

That is what is see in actual AQR MOM (AMOMX - large cap, ASMOX -small cap) returns. MOM exposure is shown by portfolio visualizer to have paid a premium during the time of those funds(1) have been in existence but the funds did not (due to negative alpha and other loadings).


(1) Shown in their factor statistic page: https://www.portfoliovisualizer.com/factor-statistics
larryslocum1982
Posts: 45
Joined: Sat Dec 09, 2017 9:36 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by larryslocum1982 »

How do you know which factors are going to work better in the future.
Random Walker
Posts: 5561
Joined: Fri Feb 23, 2007 7:21 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by Random Walker »

Linnainmaa and Roberts [2016] new data) examine investment returns before and after the original researchers’ sample periods. The average OOS outperformance from these anomalies was, on average, 58% below the in-sample numbers. McLean and Pontiff [2016] examined 97 factors published in top scholarly journals, finding that post-publication returns were (coincidentally) also 58% lower than the published research reported, while correlations with other pub-lished return predictors rose, impairing the diversifica-tion benefits of multifactor investing.
YIKES!

This is why the investor really needs to have conviction about the intuitive risk and behavioral based rationale behind the factors. If I remember, Larry has reported studies showing about a 1/3 haircut on the factor premia.

Dave
paisano
Posts: 43
Joined: Sun Nov 30, 2014 8:44 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by paisano »

Factor Investing: Long-Only versus Long-Short by David Blitz, Joop Huij, Simon Lansdorp, and Pim van Vliet (March 2014) discusses that topic in detail, including analyzing several levels of leverage for the long-short approaches.
garlandwhizzer
Posts: 3562
Joined: Fri Aug 06, 2010 3:42 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by garlandwhizzer »

in_reality wrote:
That's a problem with the long term evidence chart I think. It's looking at factors in isolation. If MOM looks juicy, but implementation in an actual fund results in changing other loadings (or negative alpha) which can offset and counter-act the MOM premium, then are you really setting yourself up for better returns by concentrating your portfolio on MOM?

That is what is see in actual AQR MOM (AMOMX - large cap, ASMOX -small cap) returns. MOM exposure is shown by portfolio visualizer to have paid a premium during the time of those funds(1) have been in existence but the funds did not (due to negative alpha and other loadings).
1+

I think this is an important point. There is a wide (and many believe growing) gulf between the robust returns of factor premiums as reported in the academic literature and the real world results of funds that attempt to harvest those premiums. It's not just AQR's MOM funds. RZV, Gugenheim's S&P 600 Small Cap Pure Value Fund, is very small (micro-cap, average market cap less than 1 billion, lots of exposure to the small premium) and not just value but deep value according to the Morningstar style box. As a deep value micro-cap fund, we would expect for this fund to maximize the small and value premiums over a long time span. In point of fact, since its inception 12 years ago it has significantly underperformed plain old vanilla Vanguard Small Cap Index (larger cap, no value slant at all) as well Vanguard's Total Market Fund. Factor enthusiasts explain that SCV has recently suffered one of its long spells of underperformance and will prevail in the end. On the other hand, as in_reality points out, that explanation doesn't work for AQR's MOM funds. Factor backtesting analysis of the MOM premium has been robustly positive for the lifespan of AQR's 2 MOM funds and yet they too have underperformed comparable cap weighted indexes partly due to substantial negative alpha. Alpha appears to be a fudge factor and in this case a big one. It appears that when academic factor backtesting results don't add up to real world results of funds, we cal the difference alpha. MOM has done well on academic backtesting but AQR's 2 MOM funds have at the same time underperformed beta. AQR is run by very smart and knowledgeable students of the market who incorporate extensive factor research knowledge into stock selection of their factor funds. Apparently outperformance is not easy to come by in the real world even by the brightest among us. A fund with attractively positive factor loads in a 4 or 5 factor model apparently is not guaranteed to outperform. And as I have pointed out before, a fund with totally unimpressive even negative factor loads in any model, Vanguard's Primecap Fund (mostly a LCG fund) has massively outperformed both beta and factor approaches since its inception in 1984, 33 years ago, a fact that factor analysis alone cannot explain. The state of factor knowledge at present is not perfect. It suggests but does not define returns in the real world.

Garland Whizzer
Random Walker
Posts: 5561
Joined: Fri Feb 23, 2007 7:21 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by Random Walker »

I believe AQR’s funds are multi factor funds. As some have pointed out, sometimes exposure to one factor can sort of cancel out exposure to another factor. I think when AQR selects investments for its funds, it tends to use investments that score strong B’s in all the requisite categories rather than A’s in some and C or D in others.

Dave
lack_ey
Posts: 6701
Joined: Wed Nov 19, 2014 10:55 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by lack_ey »

Random Walker wrote: Mon Dec 18, 2017 10:03 am I believe AQR’s funds are multi factor funds. As some have pointed out, sometimes exposure to one factor can sort of cancel out exposure to another factor. I think when AQR selects investments for its funds, it tends to use investments that score strong B’s in all the requisite categories rather than A’s in some and C or D in others.

Dave
Their single factor funds are not multifactor funds.

Code: Select all

AUEIX Large Cap Defensive Style Fund
ANDIX International Defensive Style Fund
AZEIX Emerging Defensive Style Fund
AMOMX Large Cap Momentum Style Fund
ASMOX Small Cap Momentum Style Fund
AIMOX International Momentum Style Fund
QEMLX Emerging Momentum Style Fund
ATMOX TM Large Cap Momentum Style Fund
ATSMX TM Small Cap Momentum Style Fund
ATIMX TM International Momentum Style Fund
You can readily get negative alpha when loading on a non-market factor if returns in the small caps were better than returns in the large caps and you're using large caps (or vice versa), or factor returns were driven by the short side. These relative performance issues can come and go. You can't use a short performance history to conclude something about real-world returns and investability when all of this kind of data is noisy.

On a related note, it matters which definition of momentum you use. There are plenty of variations other than the traditional last-12-months-except-the-last-one sort (commonly referred to as "2-12" price momentum, looking at 2 months to 12 months back). The iShares ETF is 4.5 years old, running an MSCI index that uses both 2-6 and 2-12 relative return, scaled by volatility. Using multiple measures and then accounting for volatility are two changes that might enhance performance and definitely played out well the last few years. It also changes weights by the composite score, rather than using market cap weighting within the selected group. I'd guess that there's some slight edge to this methodology—in addition to the edge from the lower ER, 0.15% compared to 0.40% for AQR on I shares—but that it's also been additionally lucky. Sometimes simpler is better, but sometimes not, actually.

MTUM compared with AMOMX and other funds since inception:
Image

Frankly AQR charges way too much for such a plain, easy-to-manage strategy, using a standard methodology. 5.5% Apple, 4.1% Microsoft, 3.7% Facebook, 3.4% Alphabet, 2.6% JPMorgan, 2.4% Amazon, 2.1% Bank of America, etc., okay. I think some of their more expensive funds are probably better values.

Wes Gray had a somewhat related article on MTUM returns early this year, maybe useful if thinking about implementation issues and definitions, randomness of returns, etc.
rkhusky
Posts: 17654
Joined: Thu Aug 18, 2011 8:09 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by rkhusky »

garlandwhizzer wrote: Mon Dec 18, 2017 9:25 am And as I have pointed out before, a fund with totally unimpressive even negative factor loads in any model, Vanguard's Primecap Fund (mostly a LCG fund) has massively outperformed both beta and factor approaches since its inception in 1984, 33 years ago, a fact that factor analysis alone cannot explain.
According to Portfolio Visualizer, the CAPM model does a good job of explaining Primecap's returns over the last 10-15 years. It predicts the daily returns with an R^2 of 0.94 with just 2 parameters. Primecap appears to move with the market, but with an annual alpha of 2%. The 3-5 factor FF models do not improve the R^2 over the CAPM results. The model does a poorer job of explaining daily returns for the first 15 years of Primecap, with an R^2 of 0.82.
User avatar
in_reality
Posts: 4529
Joined: Fri Jul 12, 2013 6:13 am

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by in_reality »

lack_ey wrote: Mon Dec 18, 2017 11:08 am
You can readily get negative alpha when loading on a non-market factor if returns in the small caps were better than returns in the large caps and you're using large caps (or vice versa), or factor returns were driven by the short side. These relative performance issues can come and go. You can't use a short performance history to conclude something about real-world returns and investability when all of this kind of data is noisy.
Well, I would use a few years of returns to judge whether a fund is hitting its benchmark or not. Obviously we long only investors shouldn't be looking at long-short portfolios to be a benchmark, but anyway the long-term evidence in the article is using long-short and most investors are using long only and told that we might get 33-50% of the premium.

Anyway, I don't understand a few things here:

1) Why would one get negative alpha (in small cap mom fund) if large cap mom did better than small cap mom? Why wouldn't factor analysis show the small cap mom fund with a positive small loading and which detracts from performance due to the small premium being negative during the time?

2) How could a long only fund have such negative alpha in cases where the short side was driving returns? Why wouldn't factor analysis show the fund to have a weaker MOM loading which gets multiplied by the stellar mom premium (which overall is driven by the short side in this example period) when calculating returns. Why would a long only mom fund have the same loadings as a long-short fund and thus need negative alpha to bring returns back to reality. It doesn't make sense to me.

Let me give a real example. Let's look at the components of two funds returns. They are calculated by taking the loadings * the factor premium for the period.

-------------------------------------------
QSERX (AQR small, value, mom, quality) in it's three year life had a 9.3% CAGR:
Market 10.4%
small -2.75%
value -1%
mom 1.1%
quality 4.8%
BAB 2.3% (low vol)
alpha -5.51%

-----------------------------------------
IJS (iShares small value) over the same time had a 11.28% CAGR.
Market 10.7%
small -2.8%
value -2.7%
mom 1.3%
quality 6.5%
BAB 0.7% (low vol)
alpha -2.91%

So what you are suggesting is that QSERX has negative alpha because the returns are from the short side and it's a long only fund. Well IJS is long only too. It has a larger loading on mom that QSERX. Why would only QSERX generate negative alpha in the model due to there not being short positions. IJS doesn't have them either.

Also, QSERX and IJS have very similar size loadings. If it's the large mom that is doing well and driving the 6.36% mom premium for the period, why wouldn't IJS have an equal amount of negative alpha. After all, it isn't holding large cap mom either.

So maybe the -2.91% negative alpha in IJS is to reflect that short positions and large mom are what is creating the premium. But why is QSERX so much worse.

OK, OK, the data are fuzzy and this is a short period. I'm not drawing any conclusions.

I do have questions though about the applicability of "The Long-Term Evidence" of Factor-Based Investing when it comes to actual funds though.
lack_ey wrote: Mon Dec 18, 2017 11:08 am On a related note, it matters which definition of momentum you use. There are plenty of variations other than the traditional last-12-months-except-the-last-one sort (commonly referred to as "2-12" price momentum, looking at 2 months to 12 months back). The iShares ETF is 4.5 years old, running an MSCI index that uses both 2-6 and 2-12 relative return, scaled by volatility. Using multiple measures and then accounting for volatility are two changes that might enhance performance and definitely played out well the last few years. It also changes weights by the composite score, rather than using market cap weighting within the selected group. I'd guess that there's some slight edge to this methodology—in addition to the edge from the lower ER, 0.15% compared to 0.40% for AQR on I shares—but that it's also been additionally lucky. Sometimes simpler is better, but sometimes not, actually.
That was my concern when I sold it.

The implementation might work in this environment, but not in another. So even if "The Long-Term Evidence" shows that MOM produces results in long-short portfolios (that by the way don't have trading costs or shorting costs and don't reflect that sometimes you can't short a stock even if you wanted to), that having faith in the premium and having faith that a particular implementation will catch the premium in my investment horizon are for me two different things.
User avatar
nisiprius
Advisory Board
Posts: 52105
Joined: Thu Jul 26, 2007 9:33 am
Location: The terrestrial, globular, planetary hunk of matter, flattened at the poles, is my abode.--O. Henry

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by nisiprius »

I haven't done it yet, and I don't want to do many different analyses... above, I showed an analysis that people have convinced me is bogus, showing that the pure HmL factor does not improve an S&P 500 portfolio. It doesn't, but it's because the net total investment is zero in the factor is zero, therefore the you'd expect the return to be cut by whatever percentage of the base portfolio has been replaced by pure HmL.

So, here's what I plan to do, but I want to be sure it's been discussed and approved... the general idea is: supposing that a "pure Fama-French HmL value factor" implemented as a long-short portfolio had existed in 1927. I would like to explore whether and how much it would have improved a "total market" portfolio.

Now, HmL is defined thus:
"HML (High Minus Low) is the average return on two value portfolios minus the average return on two growth portfolios,

HML = 1/2 (Small Value + Big Value)
 - 1/2 (Small Growth + Big Growth).
Since that has 0% net investment, it seems to me that in order for the hypothetical "HmL fund" to have 100% invested in stocks, it should, then have, added to it, 1/2 Small Value + 1/2 Big Value, representing the same stock investment that's been offset by short positions. Notice, too, that this portfolio is tilted toward small stocks (I think!) and therefore presents a favorable case since factor mavens generally believe small stocks are superior to the total stock market.

So that's "the diversifier: 50% * (S/H + B/H) + 50% (S/H + B/H) - 50% (S/L + B/L) = 100% * (S/H + B/H) - 50% (S/L + B/L).

For "the base portfolio," to which it is to be added, I think I will use the "CRSP 1-10 index" values as obtained from the Dimensional Matrix Book 2017.

So I will explore the efficient frontier for that combination.

As my measure of how valuable the diversifier is, I will assume that the goal is to have the same standard deviation as the base portfolio, and will therefore take increase in return between the tangent line and the base portfolio... that is, I assume that after we mix in the diversifier, we either add in the riskless asset or use (cost-free!) leverage to equalize the standard deviation.

Before I make the test, do people accept it as reasonable?
Annual income twenty pounds, annual expenditure nineteen nineteen and six, result happiness; Annual income twenty pounds, annual expenditure twenty pounds ought and six, result misery.
lack_ey
Posts: 6701
Joined: Wed Nov 19, 2014 10:55 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by lack_ey »

in_reality wrote: Tue Dec 19, 2017 5:42 am 1) Why would one get negative alpha (in small cap mom fund) if large cap mom did better than small cap mom? Why wouldn't factor analysis show the small cap mom fund with a positive small loading and which detracts from performance due to the small premium being negative during the time?
There are large caps and small caps not included in the long or short side of the momentum factor that can partially determine the size premium.

If large momentum does well, it's either from long side large cap high momentum or short side large cap low momentum, or both. There are all kinds of possibilities for outcomes and potential interaction effects over periods of times between factors. But the factor models only provide loadings in individual factors and don't consider interactions. Those end up in alpha. So can be highly random, especially over shorter periods, and probably are not significant.
in_reality wrote: Tue Dec 19, 2017 5:42 am2) How could a long only fund have such negative alpha in cases where the short side was driving returns? Why wouldn't factor analysis show the fund to have a weaker MOM loading which gets multiplied by the stellar mom premium (which overall is driven by the short side in this example period) when calculating returns. Why would a long only mom fund have the same loadings as a long-short fund and thus need negative alpha to bring returns back to reality. It doesn't make sense to me.
Imagine this hypothetical scenario: short side does terribly, driving factor returns highly positive. The 30-70 percentile range included in neither short nor long sides does well, and the long side does okay overall but moves in a different pattern from the 30-70 range. In this case, if you are exposed to the long side, you will be correlated with the long-short factor return (after all, any movements in the long side stocks will show up in both long-short factor return and your long side holdings) and thus the factor return will regress significantly on your returns. However, you will have lower returns than the 30-70 group and maybe lower than the market. So you can have lower returns even than the market, despite a positive loading on the factor and the factor having a positive return.
in_reality wrote: Tue Dec 19, 2017 5:42 amSo what you are suggesting is that QSERX has negative alpha because the returns are from the short side and it's a long only fund. Well IJS is long only too. It has a larger loading on mom that QSERX. Why would only QSERX generate negative alpha in the model due to there not being short positions. IJS doesn't have them either.

Also, QSERX and IJS have very similar size loadings. If it's the large mom that is doing well and driving the 6.36% mom premium for the period, why wouldn't IJS have an equal amount of negative alpha. After all, it isn't holding large cap mom either.

So maybe the -2.91% negative alpha in IJS is to reflect that short positions and large mom are what is creating the premium. But why is QSERX so much worse.

OK, OK, the data are fuzzy and this is a short period. I'm not drawing any conclusions.
For your specific example with those funds, I don't know offhand and haven't checked. You'd have to check which groups of stocks did what, and how they influenced different underlying factor returns. There's a good deal of randomness in the returns, factors, etc. especially for shorter periods, using large number of regressors.

There's always a certain amount of variability in returns that are not going to be explained by any given model well.
in_reality wrote: Tue Dec 19, 2017 5:42 amThe implementation might work in this environment, but not in another. So even if "The Long-Term Evidence" shows that MOM produces results in long-short portfolios (that by the way don't have trading costs or shorting costs and don't reflect that sometimes you can't short a stock even if you wanted to), that having faith in the premium and having faith that a particular implementation will catch the premium in my investment horizon are for me two different things.
Well said. That is one of the concerns you should have in factor investing.

nisiprius wrote: Tue Dec 19, 2017 1:33 pm Before I make the test, do people accept it as reasonable?
Looks reasonable to me, with the diversifier at 100% (S/H + B/H) - 50% (S/L + B/L).

Looking at it, that should end up with a market loading in the ballpark of 1, maybe slightly more, with a size loading, and with over 1 value loading. It should be riskier than the market for sure.

That said, I suspect the optimal portfolio over the period is going to be 100% diversifier diluted by cash to get back to the vol of the market.
User avatar
Taylor Larimore
Posts: 32839
Joined: Tue Feb 27, 2007 7:09 pm
Location: Miami FL

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by Taylor Larimore »

Before I make the test, do people accept it as reasonable?
Nisiprius:

The test sounds very reasonable -- except for one thing.

"Past performance does not forecast future performance."
Bill Schultheis, adviser and author of The Coffeehouse Investor: "Using past performance numbers as a method for choosing mutual funds is such a lousy idea that mutual fund companies are required by law to tell you it is a lousy idea."
Best wishes
Taylor
"Simplicity is the master key to financial success." -- Jack Bogle
lack_ey
Posts: 6701
Joined: Wed Nov 19, 2014 10:55 pm

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by lack_ey »

nisiprius wrote: Tue Dec 19, 2017 1:33 pm I haven't done it yet, and I don't want to do many different analyses...
Just wondering if you were still going to do this. See above. I'm kind of mildly curious about the result but of course am not expecting anybody to do all this kind of work just for me.

Otherwise, I might just check on it myself if I ever get that bored one of these days. :twisted:
hdas
Posts: 1395
Joined: Thu Jun 11, 2015 8:24 am

[Deleted]

Post by hdas »

[Deleted]
Last edited by hdas on Wed Jan 29, 2020 1:55 pm, edited 1 time in total.
....
pkcrafter
Posts: 15461
Joined: Sun Mar 04, 2007 11:19 am
Location: CA
Contact:

Re: Factor-Based Investing: The Long-Term Evidence - Dimson & Marsh

Post by pkcrafter »

Here's a review of the paper by Larry Swedroe.

https://thebamalliance.com/blog/the-lon ... investing/
As Dimson, Marsh and Staunton note, we must recognize that all factors have experienced, and almost certainly will continue to experience, long periods of underperformance.

This fact isn’t a reason to avoid factors and miss out on their diversification benefits and expected premiums. However, it does mean investors must be prepared to endure those long periods and stay disciplined, which I have learned requires a strong belief system. If you are not convinced you have that strong belief system, a necessary condition for having the discipline and patience you will need to maintain your exposure to a factor through those inevitable bad times, you should not invest in that factor, and that includes market beta.
My thoughts are that you are exposed to these factors if you hold total market. Overweighting factors, like overweighting small caps, requires above average knowledge, patience, and discipline. The average investor, i.e. most investors, are probably better off simply staying with the classic 3-fund portfolio. If you can manage to hold factor tilting long term, then you have to deal with high expenses.

If overweighting factors really works, the market will incorporate them to a greater extent and you're done. :happy

Paul
When times are good, investors tend to forget about risk and focus on opportunity. When times are bad, investors tend to forget about opportunity and focus on risk.
Post Reply