Time-Series Momentum: Is It There?

Discuss all general (i.e. non-personal) investing questions and issues, investing news, and theory.
Post Reply
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Time-Series Momentum: Is It There?

Post by Uncorrelated »

A few months ago I spent some time discussing momentum strategies. One thing that always bothered me was the overall lack of strong evidence. Proponents for momentum strategies would show a backtest or academic paper of some kind, then I would dismiss the evidence saying that it's low quality evidence, overfitting is likely, and that there was no evidence that the strategies success was the result of momentum. A lack of quality research hindered the debate.

But recently, I found this paper: Time-Series Momentum: Is It There? which contains the tests I was looking for, summarized below.

A quick primer on momentum, there are two types of momentum, cross sectional momentum and time series momentum. Time series momentum looks at different assets in the same time period. It is claimed that if Apple has outperformed Microsoft in the preceding 12 months, Apple is also likely to outperform Microsoft in the future. This type of momentum is commonly known as the momentum factor or UMD (up minus down).

Time series momentum looks at a single asset class over time, the claim is that if stocks went up in the last 12 months, stocks have slightly higher expected returns in the future. This can be summarized as "the trend of the market is likely to continue.". This is the type of momentum the paper investigates.

Back to the paper. First, they time-series momentum evidence across different asset classes. They find that at the 10% significance level, only 8 out of 55 assets exhibit statistically significant momentum, indicating the basic evidence is weak. They also note that only 3 out of 55 assets have out-of-sample R^2 that is significant at the 10% significance level.

Secondly, they perform a pooled regression. A pooled regression combines all asset returns and computes a single t-statistic for all different asset classes. They confirm the result of previous literature with a t-statistic of 4.34 (strongly statistically significant), this can be interpreted as strong evidence for time series momentum. The authors then show that this computations contains three statistical biases that inflate the t-statistic upwards. The authors then show that after correcting for these biases, a pooled regression does not improve the predictive power out-of-sample.

Finally, the authors investigate a proposed time series momentum trading strategy. They show various ways in the performance of a momentum trading strategy does not necessarily indicate time series momentum. Specifically, they show a strategy that does not require time series momentum, yet performs as well as the proposed momentum trading strategy. This indicates that the positive results from backtests were never the result of time series momentum in the first place.


So there you have it: time series momentum does not predict future returns.
User avatar
JoMoney
Posts: 9765
Joined: Tue Jul 23, 2013 5:31 am

Re: Time-Series Momentum: Is It There?

Post by JoMoney »

"Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
"To achieve satisfactory investment results is easier than most people realize; to achieve superior results is harder than it looks." - Benjamin Graham
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
rkhusky
Posts: 10179
Joined: Thu Aug 18, 2011 8:09 pm

Re: Time-Series Momentum: Is It There?

Post by rkhusky »

Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
They only (correctly) predict the future if you can (correctly) predict future factor premiums.
Last edited by rkhusky on Fri Oct 16, 2020 7:49 am, edited 2 times in total.
User avatar
JoMoney
Posts: 9765
Joined: Tue Jul 23, 2013 5:31 am

Re: Time-Series Momentum: Is It There?

Post by JoMoney »

Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
"To achieve satisfactory investment results is easier than most people realize; to achieve superior results is harder than it looks." - Benjamin Graham
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

JoMoney wrote: Fri Oct 16, 2020 7:43 am
Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
Please read the academic literature, such as The value premium or The Five-Factor Fama-French Model: International Evidence for the relevant statistical tests.

It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
skeptic42
Posts: 96
Joined: Mon Feb 11, 2019 5:27 pm

Re: Time-Series Momentum: Is It There?

Post by skeptic42 »

Uncorrelated wrote: Fri Oct 16, 2020 8:47 am
JoMoney wrote: Fri Oct 16, 2020 7:43 am
Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
Please read the academic literature, such as The value premium or The Five-Factor Fama-French Model: International Evidence for the relevant statistical tests.

It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Could you provide the source for the bold part, please?

And does "does not predict future returns" mean that trend-following, CTAs, and managed-futures are a waste of time and money? That was my feeling all the time. :twisted:
ScubaHogg
Posts: 714
Joined: Sun Nov 06, 2011 3:02 pm

Re: Time-Series Momentum: Is It There?

Post by ScubaHogg »

Uncorrelated wrote: Fri Oct 16, 2020 8:47 am It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Is it not relevant though how actual on the ground execution works out? If theory and reality keep disagreeing, is it reality we should dismiss?
“Unexpected Returns dominate the Expected Returns” - Ken French
User avatar
Forester
Posts: 1551
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Time-Series Momentum: Is It There?

Post by Forester »

Time series momentum is risk management, 7 of 8 exits are false signals, but the 8th is November 2007. The evidence is probably stronger on stock indices than commodities; a very long lookback period = buy&hold and we know that stocks always rise in the long run.
User avatar
imak
Posts: 58
Joined: Fri Mar 22, 2019 5:18 pm
Location: Austin, TX

Re: Time-Series Momentum: Is It There?

Post by imak »

On time-series momentum, I had come across this interesting paper in 2019 and it approaches this problem at asset class level (i.e. trend following S&P 500) by using monte-carlo simulation, parameterizing various look-back and portfolio holding periods. Both long-short (LS) and long-cash (LC) time series momentum are simulated.

Time-Series Momentum : A Monte-Carlo Approach
https://www.econstor.eu/bitstream/10419 ... 019_06.pdf

Quotes from paper:
"We create 10,000 paths of different TSM strategies based on the S&P 500 and across-asset class futures portfolio. The simulations reveal a probability distribution which shows that strategies that outperform Buy-and-Hold in-sample using historical backtests may out-of-sample i) exhibit sizable tail risks ii) underperform or outperform. Our results are robust to using different time-series models, time periods, asset classes, and risk measures."

"What is clear is that performance can vary widely between lookback windows and assets, and the reasons for a lookback window/asset combination to outperform or underperform, may not be the same reasons for another. However, by looking at their empirical distributions, we see that some strategies exhibit outcomes that have characteristics of having a risk-premium - paying to give up crash risks (12M LC-TSM vol.-unadjusted); and some strategies seem to have outcomes that are statistical artifacts (9M LS-TSM vol.-unadjusted). This uncertainty creates real risks in the choice of which window/asset combination to use."

"We find that strategies that outperform in-sample using historical backtests may, out-of-sample, underperform or outperform consistent with the predictions of our simulations."

No surprises here, at asset class level time-series momentum or trend following is not very useful and prone to whipsaws.

Also, note that this approach is very different from the approach used by momentum ETFs such as MTUM where securities are selected/ranked based on their individual time-series or relative momentum within an equity portfolio.
AA: 30% FNDX, 30% FNDA, 10% FNDF, 10% FNDC, 10% REET+VWO+DGS, 10% TMF; EF = VTEB; "Discipline matters more than allocation" ~ W Bernstein
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

skeptic42 wrote: Fri Oct 16, 2020 10:33 am
Uncorrelated wrote: Fri Oct 16, 2020 8:47 am
JoMoney wrote: Fri Oct 16, 2020 7:43 am
Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
Please read the academic literature, such as The value premium or The Five-Factor Fama-French Model: International Evidence for the relevant statistical tests.

It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Could you provide the source for the bold part, please?

And does "does not predict future returns" mean that trend-following, CTAs, and managed-futures are a waste of time and money? That was my feeling all the time. :twisted:
The paper I summarized in the opening post indicates there is no evidence time series momentum works. Any products that depend on that theory are likely a waste of time. Products that depend on cross-sectional momentum are not necessary a waste of time (they might be, but this paper simply doesn't discuss that subject).


See The Five-Factor Fama-French Model: International Evidence, table 2, Global HmL has a higher t-stat than global MKT. MKT is statistically significant in 1 out of 4 regions, HmL in 3 out of 4. (time period 1992-2014)

Image

ScubaHogg wrote: Fri Oct 16, 2020 10:47 am
Uncorrelated wrote: Fri Oct 16, 2020 8:47 am It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Is it not relevant though how actual on the ground execution works out? If theory and reality keep disagreeing, is it reality we should dismiss?
The theory says that HmL has a positive expected return and SmB has a positive expected return. The practice says HmL has a positive realized return and SmB has a positive realized return.

The theory says that funds with positive exposure to HmL and SmB and no negative alpha should post higher arithmetic returns than other funds. Since 1992 DFSVX has posted higher arithmetic returns than total stock market, after expenses(!), despite this being one of the worst periods for value, and despite me cherry picking the geographic region with the worst value returns.


It looks like theory and reality match up just fine.
Forester wrote: Fri Oct 16, 2020 11:07 am Time series momentum is risk management, 7 of 8 exits are false signals, but the 8th is November 2007. The evidence is probably stronger on stock indices than commodities; a very long lookback period = buy&hold and we know that stocks always rise in the long run.
Citation needed.
User avatar
nisiprius
Advisory Board
Posts: 41957
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: Time-Series Momentum: Is It There?

Post by nisiprius »

Forester wrote: Fri Oct 16, 2020 11:07 am...we know that stocks always rise in the long run...
Except in two of the 23 national stock markets studied by Dimson & al. in Triumph of the Optimists.
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
Forester
Posts: 1551
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Time-Series Momentum: Is It There?

Post by Forester »

Swedroe summarising a recent paper on trend following;

https://alphaarchitect.com/2020/10/15/e ... the-globe/
The trend strategy generates significant returns in the international stock market. At the individual market level, in 39 out of 49 markets, the trend portfolio generated economically meaningful and statistically significant returns.....

..... The trend premium cannot be explained by either the Fama-French three-factor or Fama-French-Carhart four-factor models.
Trend is likely more profitable on indices (1) it's less volatile so the most recent month is not skipped (2) better liquidity & lower costs vs trading individual stocks. Anyway if a FF model can't explain something, perhaps the problem is the model.
Northern Flicker
Posts: 6491
Joined: Fri Apr 10, 2015 12:29 am

Re: Time-Series Momentum: Is It There?

Post by Northern Flicker »

Uncorrelated wrote: Back to the paper. First, they time-series momentum evidence across different asset classes. They find that at the 10% significance level, only 8 out of 55 assets exhibit statistically significant momentum, indicating the basic evidence is weak. They also note that only 3 out of 55 assets have out-of-sample R^2 that is significant at the 10% significance level.
Even the statistical significance of the 3 or 8 there is a dubious claim. If you are trying to distinguish momentum as an effect from just a random outcome, you have to pick what you will look at a priori, and not just look at a large number to see where the effect shows up. Looking at 55 different assets significantly increases the likelihood of any arbitrary one of them exhibiting the symptom of momentum as an artifact of chance.
Last edited by Northern Flicker on Fri Oct 16, 2020 2:37 pm, edited 1 time in total.
Risk is not a guarantor of return.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

Forester wrote: Fri Oct 16, 2020 2:18 pm Swedroe summarising a recent paper on trend following;
The evidence linked by Swedroe is also discussed in the paper I summarized in the opening post. The authors argue that the regressions performed by prior research are faulty due to multiple statistical biases. With corrected statistical tests, they find no evidence of time series momentum.
ScubaHogg
Posts: 714
Joined: Sun Nov 06, 2011 3:02 pm

Re: Time-Series Momentum: Is It There?

Post by ScubaHogg »

Uncorrelated wrote: Fri Oct 16, 2020 12:06 pm
The theory says that funds with positive exposure to HmL and SmB and no negative alpha should post higher arithmetic returns than other funds. Since 1992 DFSVX has posted higher arithmetic returns than total stock market, after expenses(!), despite this being one of the worst periods for value, and despite me cherry picking the geographic region with the worst value returns.


It looks like theory and reality match up just fine.
:sharebeer

Personally I’m agnostic on factors, but find it confusing when other posters (not you here) will casually dismiss actual real world performance because of “academic evidence” and “journals”, as if fundamental models can’t be wrong just because they were peer reviewed.
“Unexpected Returns dominate the Expected Returns” - Ken French
skeptic42
Posts: 96
Joined: Mon Feb 11, 2019 5:27 pm

Re: Time-Series Momentum: Is It There?

Post by skeptic42 »

Uncorrelated wrote: Fri Oct 16, 2020 12:06 pm
skeptic42 wrote: Fri Oct 16, 2020 10:33 am
Uncorrelated wrote: Fri Oct 16, 2020 8:47 am
JoMoney wrote: Fri Oct 16, 2020 7:43 am
Uncorrelated wrote: Fri Oct 16, 2020 7:13 am

The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
Please read the academic literature, such as The value premium or The Five-Factor Fama-French Model: International Evidence for the relevant statistical tests.

It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Could you provide the source for the bold part, please?

And does "does not predict future returns" mean that trend-following, CTAs, and managed-futures are a waste of time and money? That was my feeling all the time. :twisted:
The paper I summarized in the opening post indicates there is no evidence time series momentum works. Any products that depend on that theory are likely a waste of time. Products that depend on cross-sectional momentum are not necessary a waste of time (they might be, but this paper simply doesn't discuss that subject).


See The Five-Factor Fama-French Model: International Evidence, table 2, Global HmL has a higher t-stat than global MKT. MKT is statistically significant in 1 out of 4 regions, HmL in 3 out of 4. (time period 1992-2014)

Image

ScubaHogg wrote: Fri Oct 16, 2020 10:47 am
Uncorrelated wrote: Fri Oct 16, 2020 8:47 am It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Is it not relevant though how actual on the ground execution works out? If theory and reality keep disagreeing, is it reality we should dismiss?
The theory says that HmL has a positive expected return and SmB has a positive expected return. The practice says HmL has a positive realized return and SmB has a positive realized return.

The theory says that funds with positive exposure to HmL and SmB and no negative alpha should post higher arithmetic returns than other funds. Since 1992 DFSVX has posted higher arithmetic returns than total stock market, after expenses(!), despite this being one of the worst periods for value, and despite me cherry picking the geographic region with the worst value returns.


It looks like theory and reality match up just fine.
Forester wrote: Fri Oct 16, 2020 11:07 am Time series momentum is risk management, 7 of 8 exits are false signals, but the 8th is November 2007. The evidence is probably stronger on stock indices than commodities; a very long lookback period = buy&hold and we know that stocks always rise in the long run.
Citation needed.
Thanks a lot for pointing me to this table. I misunderstood and thought you were saying that the market premium is statistically weak out-of-sample since total stock market investing is feasible and HmL is statistically strong out-of-sample since Fama-French introduced it.

According to the table the HmL premium was comparable to the market premium, but is there an international value fund which performed that much better compared to an international total market fund?
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

Northern Flicker wrote: Fri Oct 16, 2020 2:32 pm
Uncorrelated wrote: Back to the paper. First, they time-series momentum evidence across different asset classes. They find that at the 10% significance level, only 8 out of 55 assets exhibit statistically significant momentum, indicating the basic evidence is weak. They also note that only 3 out of 55 assets have out-of-sample R^2 that is significant at the 10% significance level.
Even the statistical significance of the 3 or 8 there is a dubious claim. If you are trying to distinguish momentum as an effect from just a random outcome, you have to pick what you will look at a priori, and not just look at a large number to see where the effect shows up. Looking at 55 different assets significantly increases the likelihood of any arbitrary one of them exhibiting the symptom of momentum as an artifact of chance.
That surprised me as well. I checked a few more papers, and it looks like prior research on time series momentum depends on pooled regressions across many different asset classes in order to produce "statistically significant" numbers.

I don't really understand the details, but these pooled regressions appear to be highly flawed. "Time series momentum: Is it there?" notes various potential biases, but I suspect there are more.
skeptic42 wrote: Fri Oct 16, 2020 5:13 pm Thanks a lot for pointing me to this table. I misunderstood and thought you were saying that the market premium is statistically weak out-of-sample since total stock market investing is feasible and HmL is statistically strong out-of-sample since Fama-French introduced it.

According to the table the HmL premium was comparable to the market premium, but is there an international value fund which performed that much better compared to an international total market fund?
You cannot invest directly in HmL. HmL is a theoretical investment that is defined as the difference between all value stocks and all growth stocks.

You can invest in funds that attempt to obtain exposure to this theoretical investment. The longest running international small cap fund I can find is DISVX (DFA International Small Cap Value I), which realized a return of 0.78% per month since sept 1999 (after expenses). In the same time period, VTMGX (Vanguard Developed Markets Index Admiral) realized 0.37% per month.

You likely don't have access to DISVX, but the point is that A) factor realized returns are within the expectations set by Fama/French in 1992-ish, and B) at least some real-money funds succeeded in capturing this premium.
skeptic42
Posts: 96
Joined: Mon Feb 11, 2019 5:27 pm

Re: Time-Series Momentum: Is It There?

Post by skeptic42 »

Uncorrelated wrote: Fri Oct 16, 2020 6:21 pm
You cannot invest directly in HmL. HmL is a theoretical investment that is defined as the difference between all value stocks and all growth stocks.

You can invest in funds that attempt to obtain exposure to this theoretical investment. The longest running international small cap fund I can find is DISVX (DFA International Small Cap Value I), which realized a return of 0.78% per month since sept 1999 (after expenses). In the same time period, VTMGX (Vanguard Developed Markets Index Admiral) realized 0.37% per month.

You likely don't have access to DISVX, but the point is that A) factor realized returns are within the expectations set by Fama/French in 1992-ish, and B) at least some real-money funds succeeded in capturing this premium.
Thanks, but that is comparing a small cap fund to a large cap fund. The performance difference to an international small cap blend or growth fund was much smaller, e.g. DFISX (DFA International Small Company) or VINEX (Vanguard International Explorer) and far away from the stated HmL premium, why?

And what is the explanation for the difference between VWIGX (Vanguard International Growth) and VTRIX (Vanguard International Value) in favor of the growth fund?

Sorry for derailing this thread, but I would like to understand what statistically significant really means in the context of finance research.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

skeptic42 wrote: Sat Oct 17, 2020 3:44 pm
Uncorrelated wrote: Fri Oct 16, 2020 6:21 pm
You cannot invest directly in HmL. HmL is a theoretical investment that is defined as the difference between all value stocks and all growth stocks.

You can invest in funds that attempt to obtain exposure to this theoretical investment. The longest running international small cap fund I can find is DISVX (DFA International Small Cap Value I), which realized a return of 0.78% per month since sept 1999 (after expenses). In the same time period, VTMGX (Vanguard Developed Markets Index Admiral) realized 0.37% per month.

You likely don't have access to DISVX, but the point is that A) factor realized returns are within the expectations set by Fama/French in 1992-ish, and B) at least some real-money funds succeeded in capturing this premium.
Thanks, but that is comparing a small cap fund to a large cap fund. The performance difference to an international small cap blend or growth fund was much smaller, e.g. DFISX (DFA International Small Company) or VINEX (Vanguard International Explorer) and far away from the stated HmL premium, why?

And what is the explanation for the difference between VWIGX (Vanguard International Growth) and VTRIX (Vanguard International Value) in favor of the growth fund?

Sorry for derailing this thread, but I would like to understand what statistically significant really means in the context of finance research.
Statistically significant means that we are 95% confident that the return (of HmL) in some time period was greater than zero. Within expectations means that the realized return is within a 95% confidence interval of the expected return.

I compared DFISX (int. small cal) with DISVX (int. small cap value). The difference between these funds is 0.3% annually. The breakdown is:
The difference in value exposure (0.12 vs 0.34 * 3.12% annually) explains 0.67% of the performance difference.
The difference in momentum exposure (-0.08 vs -0.10 * 9.58% annually) explains -0.25% of the performance difference.
The difference in expense ratio explains -0.12% of the performance difference.
The remainder (small difference in other factor exposure and unexplained alpha) amounted to 0.00%.

DFISX is not really a normal small cap fund. It has a value exposure of 0.12, an RmW exposure of 0.10 and CmA exposure of 0.20. It's more like a value fund.

I just realized that vanguard's international fund follows a different index than DFA's fund. Specifically, the two funds disagree on which countries are developed, this also means the comparison in my previous post is invalid. I don't have access to a factor set that matches vanguard's country allocation, which makes explaining the returns of VINEX impossible. Unfortunately I can't find a good comparison candidate for these funds.

Although DFA and vanguard disagree on the country allocation for international developed, they agree on the country composition of international + emerging markets. We can abuse this to run a comparison between VGTSX (Vanguard Total Intl Stock Index Inv, which includes both emerging markets and international) and DISVX (DFA international small cap value) + DFEVX (DFA emerging markets value. There is no small cap value fund). This backtest goes back to 1996. The monthly annual return for vanguard's funds is 0.48% and for DFA's funds it's 0.91%. portfolio visualizer link

If you prefer to compare value with non-value, we can use DFIVX (DFA International Value I) instead of small cap value. In this case we can go back to 1998. We find 0.48% for vanguard's index and 0.64% for DFA's index. portfolio visualizer link.

The weights in the portfolio's above were chosen to maximize the correlation between both portfolio's. Hopefully this means the country allocation was most comparable. But I don't know.


Be careful interpreting these results. The literature says HmL is significant, but there is no consensus why it exists in the first place.
skeptic42
Posts: 96
Joined: Mon Feb 11, 2019 5:27 pm

Re: Time-Series Momentum: Is It There?

Post by skeptic42 »

Uncorrelated wrote: Sun Oct 18, 2020 9:05 am
Statistically significant means that we are 95% confident that the return (of HmL) in some time period was greater than zero. Within expectations means that the realized return is within a 95% confidence interval of the expected return.

I compared DFISX (int. small cal) with DISVX (int. small cap value). The difference between these funds is 0.3% annually. The breakdown is:
The difference in value exposure (0.12 vs 0.34 * 3.12% annually) explains 0.67% of the performance difference.
The difference in momentum exposure (-0.08 vs -0.10 * 9.58% annually) explains -0.25% of the performance difference.
The difference in expense ratio explains -0.12% of the performance difference.
The remainder (small difference in other factor exposure and unexplained alpha) amounted to 0.00%.

DFISX is not really a normal small cap fund. It has a value exposure of 0.12, an RmW exposure of 0.10 and CmA exposure of 0.20. It's more like a value fund.

I just realized that vanguard's international fund follows a different index than DFA's fund. Specifically, the two funds disagree on which countries are developed, this also means the comparison in my previous post is invalid. I don't have access to a factor set that matches vanguard's country allocation, which makes explaining the returns of VINEX impossible. Unfortunately I can't find a good comparison candidate for these funds.

Although DFA and vanguard disagree on the country allocation for international developed, they agree on the country composition of international + emerging markets. We can abuse this to run a comparison between VGTSX (Vanguard Total Intl Stock Index Inv, which includes both emerging markets and international) and DISVX (DFA international small cap value) + DFEVX (DFA emerging markets value. There is no small cap value fund). This backtest goes back to 1996. The monthly annual return for vanguard's funds is 0.48% and for DFA's funds it's 0.91%. portfolio visualizer link

If you prefer to compare value with non-value, we can use DFIVX (DFA International Value I) instead of small cap value. In this case we can go back to 1998. We find 0.48% for vanguard's index and 0.64% for DFA's index. portfolio visualizer link.

The weights in the portfolio's above were chosen to maximize the correlation between both portfolio's. Hopefully this means the country allocation was most comparable. But I don't know.


Be careful interpreting these results. The literature says HmL is significant, but there is no consensus why it exists in the first place.
Thanks a lot for the breakdown. Indeed I prefer to compare value with non-value. By the way, the backtest shows a weak performance of the international value portfolio for the last 10 years.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

skeptic42 wrote: Sun Oct 18, 2020 4:56 pm
Thanks a lot for the breakdown. Indeed I prefer to compare value with non-value. By the way, the backtest shows a weak performance of the international value portfolio for the last 10 years.
10 years of data is so little it's effectively useless. Even for the equity risk premium, the probability of under-performing a savings account over 20 to 30 years is nontrivial.
Northern Flicker
Posts: 6491
Joined: Fri Apr 10, 2015 12:29 am

Re: Time-Series Momentum: Is It There?

Post by Northern Flicker »

Uncorrelated wrote: Fri Oct 16, 2020 8:47 am
JoMoney wrote: Fri Oct 16, 2020 7:43 am
Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
Please read the academic literature, such as The value premium or The Five-Factor Fama-French Model: International Evidence for the relevant statistical tests.

It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Some would argue that the 27+ years since HmL was published is a more valid out-of-sample test than one based on historical data, the aggregate behavior of which surely was known to the developers of the HmL measure, creating a bias in the out-of-sample test based on historical data.
Risk is not a guarantor of return.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

Northern Flicker wrote: Sun Oct 18, 2020 5:40 pm
Uncorrelated wrote: Fri Oct 16, 2020 8:47 am
JoMoney wrote: Fri Oct 16, 2020 7:43 am
Uncorrelated wrote: Fri Oct 16, 2020 7:13 am
JoMoney wrote: Fri Oct 16, 2020 6:44 am "Does not predict future returns"
The people selling all of these "factor" funds and other back-tested data-mined strategies are even required to tell investors that's the case, but that doesn't stop people from chasing them.
The difference is that factors (at least HmL, and possibly SmB5, MOM, BAB, QmJ as well) actually do predict future returns in a way that is statistically significant and works out-of-sample.
HmL has not worked "out of sample" to predict or deliver higher returns. Real world returns have under-performed as much as they've outperformed. Funds like DFA's Small-Value DFSVX returns since inception, 27+ years, are at par with the broad market (but a lot more volatile).
Please read the academic literature, such as The value premium or The Five-Factor Fama-French Model: International Evidence for the relevant statistical tests.

It sounds nice that DFSVX' performance looks meager in backtests, but that is not a valid way to approach the claims made by the academics. So far, the out-of-sample evidence for HmL is quite overwhelming and even stronger than the market premium.
Some would argue that the 27+ years since HmL was published is a more valid out-of-sample test than one based on historical data, the aggregate behavior of which surely was known to the developers of the HmL measure, creating a bias in the out-of-sample test based on historical data.
I agree with that. But my point is that the evidence should still be evaluated based on statistical tests and not based on backtests. Evaluating evidence with actual funds introduces many confounding variables. For instance, DFSVX did not have a stable management strategy over it's lifetime.
User avatar
nisiprius
Advisory Board
Posts: 41957
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: Time-Series Momentum: Is It There?

Post by nisiprius »

Uncorrelated wrote: Sun Oct 18, 2020 6:18 pm...I agree with that. But my point is that the evidence should still be evaluated based on statistical tests and not based on backtests. Evaluating evidence with actual funds introduces many confounding variables. For instance, DFSVX did not have a stable management strategy over it's lifetime...
If an effect isn't robust enough to show up in funds managed by a firm with Fama and French on the board of directors, it isn't robust enough for me to care about. Indisputable theory says that putting helium balloons in the trunk of my car must increase its fuel efficiency, whether you can measure it or not--but I'm not going to do it.

The EPA says removing 100 pounds increases fuel efficiency by 1-2%, say 0.5 mpg. A party balloon has a lift of about 15 gms. if you fit twenty of them into the trunk that would be 300 gm = 0.14 pounds. So that should increase fuel efficiency by 0.0014 * 0.5 = 0.0007 mpg. You could do an experiment and prove to me that you really had increased fuel efficiency by 0.0007 mpg, with a statistical significance of p = 5%, and I could believe it and still not bother to do it.
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.
garlandwhizzer
Posts: 2987
Joined: Fri Aug 06, 2010 3:42 pm

Re: Time-Series Momentum: Is It There?

Post by garlandwhizzer »

Presumably if you're investing to make money which most of us are, the results of real funds run by experts in pursuing a strategy such as SCV is much more important than what has happened with academic models. No one gets rich from academic model results in the absence of funds that exploit that research effectively. I really don't care in the least about the details of the models if they can't produce favorable results relative to simple cheap cap weighted index funds. When DFSVX, the SCV fund that has long been considered by many factor mavens to be best, fails to outperform TSM, a one factor fund (beta) for 27 years running, it seems reasonable to question the basic model assumptions as oracles for investing performance. SCV is theoretically supposed to produce long term outperformance from those 2 premiums, otherwise why put up with their increased risk and volatility? 27 years is long term. The risk and volatility showed up consistently. Where is the outperformance?

Personally if I have to choose which is more reliable between factor model research and 27 years of real investment results, I lean toward the latter. Actually I hold both mostly TSM and some SCV because like everyone else, I don't know with certainty how the future of these 2 assets will unfold and I have become a bit concerned about very optimistic valuations in LCG.

Granted we're in a very bad multi-year period for SCV during which whatever performance advantage it had over simple cheap beta has disappeared rather quickly. Granted, it is possible that SCV may take off and outperform again going forward at some point. Time will tell, but it's been a tough slog for value, especially SCV, and for INTL for more than a decade. This in spite of many experts who have been predicting outperformance from them for a long time. These erroneous predictions were based on rational analysis of investing theory and factual backtesting history. Backtesting history and factor theories that sound compelling have ironically added up to seriously bad results over this time span. At some point these ugly ducklings will cease to underperform. When is an open question? It could be soon or much delayed. Another open question: whether it will catch up with or even surpass investors who instead of SCV have been holding TSM for the last 13 years? There is a lot of ground to make up.

Garland Whizzer
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

nisiprius wrote: Sun Oct 18, 2020 6:40 pm
Uncorrelated wrote: Sun Oct 18, 2020 6:18 pm...I agree with that. But my point is that the evidence should still be evaluated based on statistical tests and not based on backtests. Evaluating evidence with actual funds introduces many confounding variables. For instance, DFSVX did not have a stable management strategy over it's lifetime...
If an effect isn't robust enough to show up in funds managed by a firm with Fama and French on the board of directors, it isn't robust enough for me to care about. Indisputable theory says that putting helium balloons in the trunk of my car must increase its fuel efficiency, whether you can measure it or not--but I'm not going to do it.

The EPA says removing 100 pounds increases fuel efficiency by 1-2%, say 0.5 mpg. A party balloon has a lift of about 15 gms. if you fit twenty of them into the trunk that would be 300 gm = 0.14 pounds. So that should increase fuel efficiency by 0.0014 * 0.5 = 0.0007 mpg. You could do an experiment and prove to me that you really had increased fuel efficiency by 0.0007 mpg, with a statistical significance of p = 5%, and I could believe it and still not bother to do it.
But the effect did show up. Since 1993 DFSVX outperformed total stock market by 0.8% annually, after expenses. A factor regression shows no statistically significant negative alpha. A non-small cap value fund such as DFFVX (DFA US Targeted Value I) has outperformed total stock market by 3% annually since the inception of that fund (2000). And this is before we move over to international.

Imo it's a five part question:
  1. Are factors statistically significant?
  2. It is possible to create a fund that has consistent factor exposure without negative alpha?
  3. assume 1) and 2), for a mean-variance investor, using value funds results in higher expected utility?
  4. Does the individual investor believe that HmL performance will persist in the future?
  5. Is it worth management risk?
As for 1), affirmative, the academic literature indicates this. 2) close enough. 3) can be answered with my mean variance framework with some limited accuracy. 4) and 5) I don't know the answer.

Regarding the magnitude of the gains, my mean-variance analyzer says investors who use 70/30 can increase their certainty equivalent return from 3% to 4% if they use value funds. This calculation assumes only half of the historical factor premia can be captured. If that isn't significant enough, I don't know what is.
User avatar
nisiprius
Advisory Board
Posts: 41957
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: Time-Series Momentum: Is It There?

Post by nisiprius »

Uncorrelated wrote: Mon Oct 19, 2020 3:47 am ...But the effect did show up. Since 1993 DFSVX outperformed total stock market by 0.8% annually, after expenses....
I don't think it did. The important thing is that it didn't on a risk-adjusted basis: Up to 12/31/2019:

Source

Image

If we call that a "tie" (although DVSVX was a smidge lower), then the extra return was just reward for extra risk taken, nothing more.

Why did I cut off at 12/31/2019? Because from inception to date, we don't even have outperformance on raw return.

Source

Image

Now, how to regard bursty phenomena is a puzzle. But it is what it is, and in my opinion a phenomenon is not robust if twenty-six years of it can be wiped out so easily. And this is not particularly unusual. We've had at least three "bursts" of this general magnitude in the last twenty years. The history is what it is. Maybe the recent drop, which looks a little like 1998, will be followed by three years of going up while the whole stock market goes down, as happened after 1998. How convinced are you that that will happen?

Bill Miller was able to buy not just a yacht, but a superyacht, thanks on his beating the S&P 500 in 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, and 2005. I think he was following a martingale-like gambling system, getting very visible seemingly steady small wins, in exchange for rare catastrophes. Who could possibly have anticipated that a roulette wheel would come up red ten times in a row? The global financial crisis? The COVID-19 pandemic?

Miller beat the S&P 500 fifteen years in a row and then lost back all the accumulated outperformance. So I see a bit of the same thing in, at least, the kinds of factor-based portfolios that were published and investable in the late 1990s. Consistent (if bursty) outperformance for two decades, though not literally outperformance every year--and all lost in less than a year.

Based on a model portfolio published by Larry Swedroe in 1998, which I implemented by using the specific DFA funds whose names were an obvious fit for his asset class names.

The time period was constrained by the available data for DFA Enhanced US Large Company I (DFELX) [Aug 1996 - Sep 2020]. The comparison is the three Vanguard "Total" funds, matching percentage allocations for US stocks, international stocks, and bonds.

Source

Image

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
Forester
Posts: 1551
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Time-Series Momentum: Is It There?

Post by Forester »

If the value phenomenon is finished why not apply the same logic to bonds vs stocks? Last 30 years bond outperformance vs stocks yet the relative opportunity for bonds forward from here is poor. Same is possibly true for TSM/growth vs value given the valuation spread.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

nisiprius wrote: Mon Oct 19, 2020 6:41 am
Uncorrelated wrote: Mon Oct 19, 2020 3:47 am ...But the effect did show up. Since 1993 DFSVX outperformed total stock market by 0.8% annually, after expenses....
I don't think it did. The important thing is that it didn't on a risk-adjusted basis: Up to 12/31/2019:

Source

Image

If we call that a "tie" (although DVSVX was a smidge lower), then the extra return was just reward for extra risk taken, nothing more.

Why did I cut off at 12/31/2019? Because from inception to date, we don't even have outperformance on raw return.
I was referring to the expected return, which is the average of the distribution of outcomes. Supposing you are an investor that can receive either an investment with the same (past) probability distribution of DFSVX or total stock market, then DFSVX will have higher average returns, and more risk. Therefore, we can say the investments have a different distribution of possible outcomes.

The extent to which investment is preferred depends on the risk aversion of the individuals. There is at least one class of investors that will prefer DFSVX based on this specific period of returns, and those are leverage constrained investors that follow a lifecycle investing strategy. These investors have large human capital relative to their stock allocation, for these investors the arithmetic return determines the expected utility.

CAGR is a measure of risk adjusted return. Specifically, it is the median outcome assuming returns are lognormally distributed, and it is also equal to the expected utility for investors with a logaritmic utility function. For some investors, DFSVX had a more favorable probability distribution of expected returns. For other investors, total stock market had a more favorable distribution. Sharpe ratio is not an useful measure of risk-adjusted performance because it does not correspond to any particular utility function.

If we want to be totally fair, we should add 0.40% to the returns of DFSVX to account for it's high expense ratio compared to factor funds you can buy today.

Anyway, I don't really see how this talk about risk adjusted performance is related to my claim. Markets are random. The purpose of DFSVX is to obtain positive exposure to HmL and SmB. In the last 27 years HmL and SmB were low, but DFSVX did succeed in capturing them. In international HmL returns were less low, and international funds succeeded in capturing them. Of course they are "risk" premium, so this did result in additional risk. Nobody disputes that SCV has additional (multi-dimensional) risks.
Source

Image

Now, how to regard bursty phenomena is a puzzle. But it is what it is, and in my opinion a phenomenon is not robust if twenty-six years of it can be wiped out so easily. And this is not particularly unusual. We've had at least three "bursts" of this general magnitude in the last twenty years. The history is what it is. Maybe the recent drop, which looks a little like 1998, will be followed by three years of going up while the whole stock market goes down, as happened after 1998. How convinced are you that that will happen?

Bill Miller was able to buy not just a yacht, but a superyacht, thanks on his beating the S&P 500 in 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, and 2005. I think he was following a martingale-like gambling system, getting very visible seemingly steady small wins, in exchange for rare catastrophes. Who could possibly have anticipated that a roulette wheel would come up red ten times in a row? The global financial crisis? The COVID-19 pandemic?

Miller beat the S&P 500 fifteen years in a row and then lost back all the accumulated outperformance. So I see a bit of the same thing in, at least, the kinds of factor-based portfolios that were published and investable in the late 1990s. Consistent (if bursty) outperformance for two decades, though not literally outperformance every year--and all lost in less than a year.
The bursty performance is not something that is unique to HmL. In the paper "volatility lessons", Fama/French investigated the performance of various factor portfolio's and concluded that the shape of the outcome distribution over long horizons is similar for mkt and HmL. The bursty performance is just random market gyrations.

I ran a 3-factor regression based on Bill Miller's fund in 1991-2005, the statistical significance can be described as weak:

Image

It's nice to have a model that allows us to accurately estimate the confidence in a particular decision. Bill Millers fund is an in-sample 1.5 standard deviation event. US HmL performance (1963-1991 or 1963-2020) is a 3 standard deviation event, and global HmL performance in 1991-2015 is a 4 standard deviation event. The latter of which is out of sample. We can safely say that the evidence that we use here to support an HmL investment would not support an investment in Bill Miller's fund in 2005.

I recognize what you're trying to say, but it's not fair to dismiss a 4-standard deviation out-of-sample anomaly because a previous 1.5 standard deviation in-sample anomaly didn't play out. For reference, the market factor itself is a 2.5 standard deviation anomaly. The market routinely under performs t-bills over multiple decades. That's not a reason to avoid the market, all evidence should be evaluated to determine whether the market is worth investing.

Fama/French investigated the expected return of HmL (i.e. the underlying random variable that generated the sequence of returns) in 1963-1991 and 1991-2019, they state that there is a probability of 87% that the expected return is equal in both periods. That is not even a one standard deviation event.

Based on a model portfolio published by Larry Swedroe in 1998, which I implemented by using the specific DFA funds whose names were an obvious fit for his asset class names.

The time period was constrained by the available data for DFA Enhanced US Large Company I (DFELX) [Aug 1996 - Sep 2020]. The comparison is the three Vanguard "Total" funds, matching percentage allocations for US stocks, international stocks, and bonds.
That's some serious home bias. I ran a similar comparison while writing my previous post, and reached the exact opposite conclusion:

Image
https://www.portfoliovisualizer.com/bac ... tion6_2=30

I opted not to post it previously because I don't think backtest are not a good way to evaluate past performance. The confidence interval on such a backtest is enormous, even if the backtests showed a return differential of 4% per year, I would still consider it weak evidence. The reason I opened this topic in the first place was that some academics use backtest to show time series momentum exists, but it turns out that these backtests can generate an positive alpha for many different reasons, only some are related to time series momentum. Likewise, a backtest between value and non-value can decide in either favor for many reasons which are not related to the expected return of either portfolio, and that's what ultimately matters in portfolio construction.

I appreciate your input but backtests are terribly unsuited for this purpose. The same methods that you use can be used to argue for an overweight allocation in US tech, and we all know that's a terrible idea. If we want to have confidence in our investments, then we need to use a mathematical method that rejects the outperform of US tech as random noise, a backtest is not that method.
Northern Flicker
Posts: 6491
Joined: Fri Apr 10, 2015 12:29 am

Re: Time-Series Momentum: Is It There?

Post by Northern Flicker »

Uncorrelated wrote: If we want to be totally fair, we should add 0.40% to the returns of DFSVX to account for it's high expense ratio compared to factor funds you can buy today.
If we want to be fair, we also would note that VTSAX did not exist 27 years ago, and VTSMX (investor shares) had a significantly higher ER at that time, probably close to .40%.

But to be totally fair, we should look at cost (ER and transaction cost) of investments you actually could hold. And you would have had to pay an AUM fee to an advisor on top of ER to get DFSVX. It is unlikely that an individual investor could have implemented a SCV strategy themselves in a cost-effective manner. I don't believe Fama & French considered the higher trading costs of a SCV strategy.

The difficulty with dismissing the last 27 years as an anomalous outcome is that Fama & French analyzed the same length period to establish thrir result. The last 27 years is an out-of-sample test of the same length as the original sample used in the analysis.
Risk is not a guarantor of return.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

Northern Flicker wrote: Mon Oct 19, 2020 1:15 pm
Uncorrelated wrote: If we want to be totally fair, we should add 0.40% to the returns of DFSVX to account for it's high expense ratio compared to factor funds you can buy today.
If we want to be fair, we also would note that VTSAX did not exist 27 years ago, and VTSMX (investor shares) had a significantly higher ER at that time, probably close to .40%.

But to be totally fair, we should look at cost (ER and transaction cost) of investments you actually could hold. And you would have had to pay an AUM fee to an advisor on top of ER to get DFSVX. It is unlikely that an individual investor could have implemented a SCV strategy themselves in a cost-effective manner. I don't believe Fama & French considered the higher trading costs of a SCV strategy.

The difficulty with dismissing the last 27 years as an anomalous outcome is that Fama & French analyzed the same length period to establish thrir result. The last 27 years is an out-of-sample test of the same length as the original sample used in the analysis.
I didn't knew VTSMX as an higher ER at the time.

I don't plan to hold DFSVX, nor would I recommend it. The purpose of this exercise to see if it is reasonable to assume SVC funds actually capture the premium. If that is true, then you should invest in another fund such as S&P600 value ETF (0.15% at vanguard) or iShares Core S&P U.S. Value ETF (0.04%), I choose 0.4% as a conservative estimate of the difference in ER between DFSVX and modern value funds. The only reason I use DFSVX at all is because people insist on using backtest with live funds, even though the management strategy of DFSVX wasn't stable over time, the ER of the fund is significantly higher than anything that can be considered reasonable today, country weights are seriously different than in 1993 and 27 years is far too short to draw conclusions based on backtests.

I'm not discarding 27 years as an anomalous outcome. I'm arguing it's not an anomalous outcome at all based on the evidence available in 1991. A frequent anti value argument is that HmL did not deliver post-publication, but I don't agree with that: HmL delivered an outcome well within the 95% confidence interval of FF 1991 estimate of the HmL premium.
User avatar
nisiprius
Advisory Board
Posts: 41957
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: Time-Series Momentum: Is It There?

Post by nisiprius »

Uncorrelated wrote: Mon Oct 19, 2020 10:05 am...That's some serious home bias. I ran a similar comparison while writing my previous post, and reached the exact opposite conclusion...
I place a great deal of weight on the fact that this was a published, specific, on-the-record recommendation. It was an out-of-sample test of a portfolio recommended in 1998. Was your comparison to a portfolio published at the start of your test period?
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
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

nisiprius wrote: Mon Oct 19, 2020 6:06 pm
Uncorrelated wrote: Mon Oct 19, 2020 10:05 am...That's some serious home bias. I ran a similar comparison while writing my previous post, and reached the exact opposite conclusion...
I place a great deal of weight on the fact that this was a published, specific, on-the-record recommendation. It was an out-of-sample test of a portfolio recommended in 1998. Was your comparison to a portfolio published at the start of your test period?
There were hundreds of portfolio's published in 1998. Many by Larry.


Let's take a look at the actual differences between the larry portfolio and the vanguard portfolio, first, the bond portion:

Image
Image
https://www.portfoliovisualizer.com/bac ... ion4_2=100
That's a 0.9% difference in the bond portion, let's make it 1% after expenses. Was showing that short term bonds underperformed long term bonds part of your argument? Do you know any papers (before 1998) that show strong statistical evidence of active bond management? Me neither.

Then, the international portion:
Image
Image
https://www.portfoliovisualizer.com/bac ... ion4_2=100
Here DFA wins, that's great! Until you investigate what's actually in those funds. DFA's allocation is in large cap fund with statistically insignificant value exposure. In this time period SmB returned 1.11% in international markets and -0.37% in emerging markets. Hmm, that's odd? There must be something else going on... And that's the country allocation. Around 2000, the closest country allocation is VGTSX = 10% emerging markets + 90% developed. In Around 2008, it was 15 / 85. Then in the early 2010's, 22 / 78, and then finally from 2015 to now, also 22 / 77.

Larry massively overweighted emerging markets, and emerging markets happened to outperform. I'm sure this can be attributed to random country gyrations instead of Larry's foresight. Was showing that different country allocations perform different part of your argument?

Now for the actual meat of the argument, the US portion.
Image
Image
https://www.portfoliovisualizer.com/bac ... ion6_2=100
There is obvious problems here: 22% of the portfolio is allocated to REIT's. As of today, there is absolutely zero evidence that you should over-weight REIT's, there definitely wasn't any in 1998. Classic Larry, recommending asset classes without strong evidence. So we throw that out, this also takes care of a hidden 10% allocation to long term corporate bonds.

Image
Image
https://www.portfoliovisualizer.com/bac ... ion6_2=100
...and conclude that DFA US value has outperformed vanguard US by 0.51% annually. Portfolio visualizer claims there is a 0.2% difference in ER, so make that 0.71%.

The funny thing is that HmL was actually negative during that period, and almost all of the outperformance is attributed to SmB5. Factor diversification at work.

The morale of the story: you attempted to use a backtest to make an argument about value, but actually made an argument that was firstly about bonds, secondly about country allocations, thirdly about SmB, and then finally a little bit around value. If this doesn't show that backtests are completely unfit for this purpose, I don't know what will.
User avatar
nisiprius
Advisory Board
Posts: 41957
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: Time-Series Momentum: Is It There?

Post by nisiprius »

1) Thanks for taking the time to dig into those details.

2) I think we're talking past each other and that you probably have no interest in continuing, but you seem to be missing my point, which is all about "the real world." That is, rather than an analysis of why you don't think the portfolio Larry Swedroe published in 1998 correctly incorporates factor investing theory, I would be much more interested in your exhibiting a factor-informed portfolio, published in the 1990s or earlier, that did--something in print, verifiable, in a book or something similar. Published as a specific, on-the-record recommendation, either with ticker symbols or with an obvious implementation in mutual funds or ETFs that were available at the time of publication.

I don't think backtests of portfolios designed and constructed today are terribly interesting. But I think "backtests" of implementable portfolios that were were published and on the record at the beginning of the time period tested are interesting--to me, anyway--unless they are, of course, just the best of a flock of dozens published at the same time by the same person.

3) I think it's at least an interesting coincidence that both a direct comparison of 100% stocks (Total Stock versus DFA US Small-Cap Value Portfolio and their use in an expert-recommended portfolio yield the same verbal description. Inception to end of 2019, the factor-based strategy outperformed in return but not in risk-adjusted return; inception to 10/19/2020, neither outperformed in return nor risk-adjusted return.
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
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

nisiprius wrote: Tue Oct 20, 2020 6:53 am 2)I think we're talking past each other and that you probably have no interest in continuing, but you seem to be missing my point, which is all about "the real world." That is, rather than an analysis of why you don't think the portfolio Larry Swedroe published in 1998 correctly incorporates factor investing theory, I would be much more interested in your exhibiting a portfolio, published in the late 1990s that does--something in print, verifiable, in a book or something similarly accessible... as a specific recommendation, and either with ticker symbols or with an obvious implementation in mutual funds or ETFs available at the time of publication.
Investment decisions should be based on the expected return and risk. If the Larry portfolio has higher expected return than the vanguard portfolio but similar risk, it's a better choice. But that leaves us with a problem: if we go to a casino, bet on black, and we win, we haven't actually shown that betting on black has higher expected returns. Likewise, Larry bet on short term bonds, emerging markets and small cap. Some of these bets worked, some didn't. Larry selected this portfolio because he believed that his portfolio would have higher expected return, can we disprove this belief? No, not at all, the statistical evidence is too weak.

If you believed in 1998 that the Larry portfolio had a higher expected return than the vanguard portfolio, then the events in the last 20 years were well within expectations and there is no reason to change your mind. If you were unconvinced in 1998 that the Larry portfolio was superior, then the events of the last 20 years were also within expectations and there is no reason to change your mind. This impasse can only be broken if we collect sufficient evidence (which, for this kind of thing, will take over a hundred years), or if we are willing to escape the anchoring bias that is called CAPM and step into a multi-factor world. In Fama/French' world we don't have to directly show that portfolio 1 is better than portfolio 2, but rather we have to show that the individual factors are strong and that the individual funds have different factor exposures.

The academic world is well-versed in producing statistical tests that allow us to accurately asses the confidence we have in a particular investment, both in and out of sample. Those are the methods we should use. I shouldn't have to perform bogleheads archeology to find someone that recommended a particular collection of funds that just happened to outperform. That's not evidence, that's anecdote. I understand your point quite well, I just disagree with it.
3) I think it's at least an interesting coincidence that both a direct comparison of 100% stocks (Total Stock versus DFA US Small-Cap Value Portfolio and their use in an expert-recommended portfolio yield the same verbal description. Inception to end of 2019, the factor-based strategy outperformed in return but not in risk-adjusted return; inception to 10/19/2020, it outperformed neither in return nor risk-adjusted return.
Of course it's not surprise that both comparisons yield the same result, they were both based on the same academic papers. If they disagreed with one another, I would be very concerned about the ability of factor funds to provide stable factor exposure.

A direct comparison between stocks and US SVC does result in out performance (inception to date) :
Image
Arithmetic average is a measure of return, CAGR is a measure of risk adjusted return (I know this sounds strange, but it makes sense from a mathematical perspective). In this case, CAGR is not a valid measure because it depends on the total portfolio volatility, but the image above only contains a small part of a globally diversified portfolio.

This problem also makes the US portion of the Larry portfolio look much worse than it actually is. Here is the US portion of the larry portfolio without REIT's, monthly rebalanced, vs VTSMX:
Image
Now we are going to do something really stupid: we simulate adding international and bonds by adding a 70% CASHX allocation to this portfolio.
Image
Now suddenly, the difference in CAGR has increased from 0.06% to 0.11%.

This is the reason why I'm always placing the arithmetic average clearly in view. Unless you are at the final step of evaluating your complete portfolio (including human capital), CAGR is a very misleadingmeasurement.

There are a few methods to calculate better measures of risk adjusted return, but they are not easy. A valid measure can only be computed based on the complete portfolio and depends on the risk aversion of the individual.
Last edited by Uncorrelated on Wed Oct 21, 2020 2:53 am, edited 1 time in total.
rkhusky
Posts: 10179
Joined: Thu Aug 18, 2011 8:09 pm

Re: Time-Series Momentum: Is It There?

Post by rkhusky »

Not sure how applying statistics to the stock market even makes sense, since no one knows the distribution or how to describe its non-stationarity.
Northern Flicker
Posts: 6491
Joined: Fri Apr 10, 2015 12:29 am

Re: Time-Series Momentum: Is It There?

Post by Northern Flicker »

Uncorrelated wrote: Investment decisions should be based on the expected return and risk. If the Larry portfolio has higher expected return than the vanguard portfolio but similar risk, it's a better choice.
It is hard to use the expected return of a stock and bond portfolio as an input to choosing such a portfolio when we do not know the expected return of any such portfolio.
Risk is not a guarantor of return.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

rkhusky wrote: Tue Oct 20, 2020 6:07 pm Not sure how applying statistics to the stock market even makes sense, since no one knows the distribution or how to describe its non-stationarity.
The same argument applies to the market premium. When can we be confident the equity risk premium exists? 100 years? 200 years? Is it really worth betting on the equity risk premium and it's large variance for uncertain expected return?

At some point you'll have to say: I'm confident enough to wage a gamble. How confident? Well, maybe p < 0.05? p < 0.01?
Northern Flicker wrote: Tue Oct 20, 2020 8:10 pm
Uncorrelated wrote: Investment decisions should be based on the expected return and risk. If the Larry portfolio has higher expected return than the vanguard portfolio but similar risk, it's a better choice.
It is hard to use the expected return of a stock and bond portfolio as an input to choosing such a portfolio when we do not know the expected return of any such portfolio.
If you can't know the expected return of any portfolio, there is no reason to hold anything other than the market portfolio (70% bonds / 30% stocks?). Without estimates of expected return, things like vanguard target date funds or tilting towards equities are highly irrational.
rkhusky
Posts: 10179
Joined: Thu Aug 18, 2011 8:09 pm

Re: Time-Series Momentum: Is It There?

Post by rkhusky »

Uncorrelated wrote: Wed Oct 21, 2020 2:52 am The same argument applies to the market premium. When can we be confident the equity risk premium exists? 100 years? 200 years? Is it really worth betting on the equity risk premium and it's large variance for uncertain expected return?
You can never be sure that the equity premium will exist over any particular future time period. It is a bet and one that is made without knowing the odds.
Uncorrelated wrote: Wed Oct 21, 2020 2:52 am At some point you'll have to say: I'm confident enough to wage a gamble. How confident? Well, maybe p < 0.05? p < 0.01?
Past performance is no guarantee of future returns, so p values are of no use. What are the p values for a war in the Middle East in 10 years? Or a pandemic 10x worse than Covid-19 in 5 years? Or the US president assassinated in 20 years? Or a cure for most cancers in 15 years? You can't predict the future, so you can't predict the stock market.

As Taylor has said, don't invest money in stocks that you can't afford to lose.

Personally, I invest with the hope that things will work out. I started with about a 80/20 AA and gradually moved to a 40/60 AA, keeping the money that I could not afford to lose in bonds.
User avatar
Topic Author
Uncorrelated
Posts: 1039
Joined: Sun Oct 13, 2019 3:16 pm

Re: Time-Series Momentum: Is It There?

Post by Uncorrelated »

rkhusky wrote: Wed Oct 21, 2020 7:26 am As Taylor has said, don't invest money in stocks that you can't afford to lose.
In other words, don't invest anything. A position that has demonstrably led to worse certainty equivalent returns in most of history, and demonstrably results in portfolio selection that is not mean-variance efficient.

I can always count on Taylor for making statements that make no financial sense at all.
Personally, I invest with the hope that things will work out. I started with about a 80/20 AA and gradually moved to a 40/60 AA, keeping the money that I could not afford to lose in bonds.
That approach implicitly assumes some expected return on stocks. My approach is to make that implicit assumption explicit so you at least know what your underlying assumptions are.
rkhusky
Posts: 10179
Joined: Thu Aug 18, 2011 8:09 pm

Re: Time-Series Momentum: Is It There?

Post by rkhusky »

Uncorrelated wrote: Wed Oct 21, 2020 7:45 am
rkhusky wrote: Wed Oct 21, 2020 7:26 am As Taylor has said, don't invest money in stocks that you can't afford to lose.
In other words, don't invest anything. A position that has demonstrably led to worse certainty equivalent returns in most of history, and demonstrably results in portfolio selection that is not mean-variance efficient.

I can always count on Taylor for making statements that make no financial sense at all.
When you are young, you have time to make up for losses in your smaller portfolio. When you are in retirement, it is very difficult to make up for losses in your larger portfolio. Invest accordingly.
Uncorrelated wrote: Wed Oct 21, 2020 7:45 am
Personally, I invest with the hope that things will work out. I started with about a 80/20 AA and gradually moved to a 40/60 AA, keeping the money that I could not afford to lose in bonds.
That approach implicitly assumes some expected return on stocks. My approach is to make that implicit assumption explicit so you at least know what your underlying assumptions are.
I have no problem looking at past performance to get a general idea of what has happened in the past in order to guide future expectations. But once the analysis gets too precise, it stops making sense.
Post Reply