I definitely agree. I now realize I grossly overestimated my own comfort level with this part of it.lazyday wrote: ↑Thu Jul 11, 2019 12:39 pmInvesting without following an index seems to offer significant benefits for a fund like this. Of course, the downside is that when the fund does poorly, it can be hard to know why. And if the fund isn’t index hugging, then there should be times where it underperforms.
VFMF (Vanguard Multifactor) == Closet Value Index Fund

 Posts: 112
 Joined: Wed Sep 26, 2018 7:03 pm
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
VTWAX and chill
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
Check this out:vineviz wrote: ↑Wed Jul 03, 2019 7:54 pmUsually this boils down to a systematic mismatch between the factor model the fund is using and the factor model used in the regression. In effect the regression is moving returns from beta to alpha.
However, it’s also possible that this is partly an intentional trade off: maintain higher factor loads at the expense of transaction costs.
Vanguard’s QEG is taking a bottoms up approach to their factor model, for better or worse, so we’ll likely always be a little uncertain about their process.
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
Since this is an equityonly fund, it would be more appropriate to only use the equity factors (and exclude Term and Credit)
Rerunning the regression there and using daily returns, one finds that USMF has statistically significant loadings on market, size, and BAB. It does result in positive alpha, but it's also not statistically significant
Rerunning the regression there and using daily returns, one finds that USMF has statistically significant loadings on market, size, and BAB. It does result in positive alpha, but it's also not statistically significant
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
People keep repeating this but I haven't seen a good reason for it. Do you have a reference from a credible source. Thanks
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
No resource at my fingertips now, but I believe credit is to measure the premium of lowercredit bonds over highercredit bonds, and term is to measure the premium of longerterm bonds over shorterterm bonds. In the case of a portfolio without bonds I don't think it would make sense to include those variables. But again I'm not an expert here, maybe someone else would chime in.
As an aside  of course you can always include as many variables in any type regression as you want, but you'll start to lose interpretability, robustness against correlated features, and the ability to make good inferences about the meaning of the coefficients.
As an aside  of course you can always include as many variables in any type regression as you want, but you'll start to lose interpretability, robustness against correlated features, and the ability to make good inferences about the meaning of the coefficients.
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
Thanks. I don't see how anybody thinks that the yield curve and credit spreads are not extremely relevant factors. I do agree with you regarding interpretability but perhaps you take one of the other factors out. If you run a PCA analysis you get about 2 independent sources of risk in stocks, and my personal experience running an actively managed portfolio of stocks sort of confirms this.muffins14 wrote: ↑Wed Jul 17, 2019 11:28 amNo resource at my fingertips now, but I believe credit is to measure the premium of lowercredit bonds over highercredit bonds, and term is to measure the premium of longerterm bonds over shorterterm bonds. In the case of a portfolio without bonds I don't think it would make sense to include those variables. But again I'm not an expert here, maybe someone else would chime in.
As an aside  of course you can always include as many variables in any type regression as you want, but you'll start to lose interpretability, robustness against correlated features, and the ability to make good inferences about the meaning of the coefficients.
Term and credit risk based 2factor model where the term risk premium is calculated as the difference between longterm treasuries and treasury bills and the credit risk premium is calculated as the difference between longterm corporates and longterm treasuries. This model is not adjusted to account for the differences in the interest rate sensitivities of longterm treasuries and corporate bonds (refer to the Hallerbach and Houweling, and Asvanunt and Richardson papers listed below).
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
The reasons don't have anything specifically do with whether the variable is an "equity factor" or a "bond factor", but anytime you see a FamaFrench regression with high values for unadjusted R^{2} or regression F statistic combined with a large number of statistically insignificant variables then you have a big red flag for overfitting.
Overfitting results from using more (correlated) independent variables than are necessary to explain the variance in your data.
There are plenty of formal tests for this (many econometrics books should cover it) but I don't have a solid reference handy. Look for a source that describes overfitting, collinearity, and/or variance inflation.
What is happening is usually that the correlated variables are insignificant individually but significant in combination (because they are correlated), increasing the raw coefficient of determination (i.e. R^{2}). That explanatory power has to go somewhere, so it (falsely) goes to alpha.
A quick and dirty check is to drop one or two of the least significant variables and see if the adjusted R^{2} and/or regression Fstat go up. If the adjusted R^{2} goes up while the raw R^{2} goes down, for example, chances are good that you have compromised your DoF somehow.
In this case with USMF reducing the factor set to RMF, SMB, HML, & MOM increases the adjusted R^{2} from 92.4% to 93.8% and the Fstat from 34.29 to 84.31. As a broad rule of thumb, that's a good indication that the smaller variable set is maximizing both explanatory power and degrees of freedom.
"Far more money has been lost by investors preparing for corrections than has been lost in corrections themselves." ~~ Peter Lynch
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
Thanks for the comment. A couple of observations/questions:vineviz wrote: ↑Wed Jul 17, 2019 12:10 pmThe reasons don't have anything specifically do with whether the variable is an "equity factor" or a "bond factor", but anytime you see a FamaFrench regression with high values for unadjusted R^{2} or regression F statistic combined with a large number of statistically insignificant variables then you have a big red flag for overfitting.
Overfitting results from using more (correlated) independent variables than are necessary to explain the variance in your data.
There are plenty of formal tests for this (many econometrics books should cover it) but I don't have a solid reference handy. Look for a source that describes overfitting, collinearity, and/or variance inflation.
What is happening is usually that the correlated variables are insignificant individually but significant in combination (because they are correlated), increasing the raw coefficient of determination (i.e. R^{2}). That explanatory power has to go somewhere, so it (falsely) goes to alpha.
A quick and dirty check is to drop one or two of the least significant variables and see if the adjusted R^{2} and/or regression Fstat go up. If the adjusted R^{2} goes up while the raw R^{2} goes down, for example, chances are good that you have compromised your DoF somehow.
In this case with USMF reducing the factor set to RMF, SMB, HML, & MOM increases the adjusted R^{2} from 92.4% to 93.8% and the Fstat from 34.29 to 84.31. As a broad rule of thumb, that's a good indication that the smaller variable set is maximizing both explanatory power and degrees of freedom.
1. In my portfolio of individual stocks, the situation is the reverse, I need to add more factors to increase the R^2, only Term and Alpha are significant.
2. A priori it seems to me that yield curve and credit spreads are more important (macro) factors than the style equity factors.
3. The Machine Learning tools seem a bit more useful than pure econometrics dealing with this stuff. Have you tried updating your toolkit?.
Cheers
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
It's important to distinguish between unadjusted R^2 (which will ALWAYS go up with the addition of more independent variables) and adjusted R^2 (which will increase only so long as the added variable adds enough explanatory power to compensate for the smaller number of degrees of freedom).hdas wrote: ↑Thu Jul 18, 2019 3:49 pm1. In my portfolio of individual stocks, the situation is the reverse, I need to add more factors to increase the R^2, only Term and Alpha are significant.
For instance, using a random portfolio of 50 stocks I plotted the adjusted R^2 and unadjusted R^2 with various numbers of factors.
You can see that unadjusted R^2 increases no matter how many variables are added (within the rounding error reported by PortfolioVisualizer, anyway) whereas the adjusted R^2 peaks with a fourfactor model.
That wouldn't be my prior belief, but it does raise a hypothesis that can be tested. As it turns out, term and credit factor are NOT more powerful at explaining the crosssection of stock returns than socalled equity factors (e.g. market, size, value, quality, etc.)
"Far more money has been lost by investors preparing for corrections than has been lost in corrections themselves." ~~ Peter Lynch
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
If you look at the link I sent, you'll notice that both the raw and adjusted R^2 go up as I add more factors. I start with 3 factors and keep adding one by one sequentially. Cheersvineviz wrote: ↑Thu Jul 18, 2019 4:19 pm
It's important to distinguish between unadjusted R^2 (which will ALWAYS go up with the addition of more independent variables) and adjusted R^2 (which will increase only so long as the added variable adds enough explanatory power to compensate for the smaller number of degrees of freedom).
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
Without access to the holdings and regression settings I can't help you further diagnose the causes, but I can tell you that's a problematic regression.hdas wrote: ↑Thu Jul 18, 2019 5:03 pmIf you look at the link I sent, you'll notice that both the raw and adjusted R^2 go up as I add more factors. I start with 3 factors and keep adding one by one sequentially. Cheersvineviz wrote: ↑Thu Jul 18, 2019 4:19 pm
It's important to distinguish between unadjusted R^2 (which will ALWAYS go up with the addition of more independent variables) and adjusted R^2 (which will increase only so long as the added variable adds enough explanatory power to compensate for the smaller number of degrees of freedom).
"Far more money has been lost by investors preparing for corrections than has been lost in corrections themselves." ~~ Peter Lynch
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
Vineviz,
I get the same effect when running the factor regression on my beta portfolio, here are the specs:
> Last 3 years
> VUG 30%, USMV 30%, IJS 20%, XSLV 20%
> You can start adding factors sequentially, start with AQR 3 factor, then 4 factor, then add quality, then BAB, finally Term + Credit.
> Notice how at every step you get higher adjusted R^2
Cheers
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
For whatever reason, the regression on this portfolio (VUG/USMV/IJS/XSLV) is much more robust than whatever you were regressing in the other thread: Rsqured in the 98% range instead of high 60s and low 70s, fstat in the hundreds instead of the lowteens. Thdas wrote: ↑Mon Jul 22, 2019 1:23 pmI get the same effect when running the factor regression on my beta portfolio, here are the specs:
> Last 3 years
> VUG 30%, USMV 30%, IJS 20%, XSLV 20%
> You can start adding factors sequentially, start with AQR 3 factor, then 4 factor, then add quality, then BAB, finally Term + Credit.
> Notice how at every step you get higher adjusted R^2
Even so, the regression quality is marginally higher without including CDT (the adjusted Rsquared stays at 98.2% but the fstat is higher without than with). What makes the difference here, versus some other earlier examples, is that all 7 factors (once credit is excluded) are statistically significant.
With such a significant allocation to USMV and XSLV, I'm not surprised that TERM is statistically significant since the regression doesn't explicitly contain a low volatility or minimum variance factor: HML, MOM, and BAB are left to pick up the weight and they are too imperfect as proxies.
"Far more money has been lost by investors preparing for corrections than has been lost in corrections themselves." ~~ Peter Lynch
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
What does adjusted r2 mean?vineviz wrote: ↑Thu Jul 18, 2019 4:19 pmIt's important to distinguish between unadjusted R^2 (which will ALWAYS go up with the addition of more independent variables) and adjusted R^2 (which will increase only so long as the added variable adds enough explanatory power to compensate for the smaller number of degrees of freedom).
Re: VFMF (Vanguard Multifactor) == Closet Value Index Fund
CheersThe adjusted Rsquared is a modified version of Rsquared that has been adjusted for the number of predictors in the model. The adjusted Rsquared increases only if the new term improves the model more than would be expected by chance. It decreases when a predictor improves the model by less than expected by chance. The adjusted Rsquared can be negative, but it’s usually not. It is always lower than the Rsquared.
"whenever there is a randomized way of doing something, then there is a nonrandomized way that delivers better performance but requires more thought" ET Jaynes