A Better Look at Dual Momentum Fragility

Several weeks ago Corey Hoffstein of Newfound Research raised some important questions about the fragility of 12-month Dual Momentum (GEM) shared by Gary Antonacci. Gary responded in kind and it sparked a debate about 12M Dual Momentum and some related issues.

One of Corey's main arguments is that Dual Momentum, when measured only with a 12M lookback period, carries inherent and avoidable fragility. He notes that the model results vary meaningfully just by changing a single metric such as the number of months in the lookback period.

Corey suggests that investors can reduce risk by effectively tranching their Dual Momentum portfolio into multiple portfolios (he tests a model with 7 tranches) where each tranche invests in Dual Momentum following a different lookback period (6M through 12M). The tranches would be re-balanced each month to ensure equal impact of each lookback period.

This then led the debate to the next stage: setting aside Dual Momentum's recent signals, is there some reason to believe that the 12M lookback period is better than other periods (8M, 11M, etc.) to the extent that we should use it exclusively?

Is a 12-Month Lookback Better?

Gary argues essentially that 12M is better. It shows the best overall performance results for stocks, and this seems to be confirmed by academic evaluations going back a century or more. However, despite all the research, could this outperformance be a fluke?

I encourage you to read Gary's post which I linked above to understand some of the academic support behind the 12M lookback. This idea warrants a logical evaluation, especially how it relates to a self-directed investor implementing a Dual Momentum strategy for their retirement savings.

For the purposes of this argument, I will focus on the longer term lookback periods: 6M to 12M. While the shorter periods may have their own merits, they do require a lot more trading activity which I believe many self-directed investors would prefer to avoid.

First, although the 12M period does look really effective across the stock universe, I think it is somewhat speculative to say that 12M is the best. Lets look at some measures to test this.

Drawdowns

Sources: TheRichMoose.com, MSCI Inc., FRED

Looking at the drawdowns since 1970, it is hard to say that one lookback period is better or safer than the other. The 12M model has the largest absolute drawdown; however, the 6M spends more time in drawdown and comes very close the 12M model's biggest drawdown.

Both models share quite similar drawdown characteristics. They are highly correlated with the periods in drawdown showing significant overlap. When the 6M is in a drawdown, the 12M is also in a drawdown 86 percent of the time. Likewise, when the 12M is in drawdown, the 6M is also in drawdown 90 percent of the time.

However, although strongly correlated, they are certainly not precisely aligned. If I had to lend a preference to a lookback period based on drawdowns only, I would probably give a slight edge to 6M.

Returns

Sources: TheRichMoose.com, MSCI Inc., FRED

Looking at returns since 1970, we see the 12M system doing better overall. However, this is primarily due to lackluster performance in the 6M system in the late-1980s into the mid-1990s. Prior to, and following this period, 6M ran very close to 12M on an annualized rolling return basis.

Sources: TheRichMoose.com, MSCI Inc., FRED

Looking at the annualized 5-year rolling returns of each lookback period tested, we can clearly see the drop in the 1980s and 1990s for the 6M system. However, the chart also shows the systems performing quite comparably in the other market periods.

In fact, the two systems are 87.78 percent correlated on a monthly return basis over the period of 1970 through 2018. Despite the differences in end return, the monthly returns are very similar.

I think it is reasonable to assume that there may be some point in the future where 12M will underperform 6M by a reasonable margin for some time. After all, we do have limited data to work with and the returns observed for each system are within the historical ranges for the other system.

Sharpe Ratios

The Sharpe ratio is a common measure of portfolio performance factoring in risk. In this measurement, risk is measured via standard deviation which has its drawbacks—particularly when Dual Momentum has downside guards in place.

By this I mean, standard deviation measures return variance to the upside and downside as equals. However, in 6M and 12M Dual Momentum, downside is observed to a peak of -25 percent while upside goes over 60 percent in similar time frames. As investors, we like large upside movements, but would like to avoid large downside movements—a return stream Dual Momentum seems to provide.

These quirks aside, Sharpe ratios are still a useful rough measure to test risk-adjusted portfolio performance. In the 6M system, the Sharpe ratio is 0.733; in the 12M system, the Sharpe ratio is 0.874. The 12M system has a higher Sharpe ratio, so 12M is better lookback period by this measure.

What's more, the Sharpe ratio advantage for the 12M system seems to be pervasive throughout our lookback period. Even when we strip out 6M's diverging years (1988-1993), the 12M system still has a better Sharpe ratio at 0.841 compared with 6M's Sharpe of 0.815. (Although this difference is much smaller and I would consider them to be statistically "about equal".)

Newfound's 7 Period System

Next, we will take a closer look at Newfound's suggested system where the investor uses seven lookback periods to scale into and out of positions. Each system is re-balanced monthly as needed when trades are made; the impact of each system on the overall portfolio stays the same.

Setting aside trading frequency and cost for the moment, I have to say that Newfound's system does make logical sense. If the seven lookback periods have comparable expected long-term risk-adjusted returns, and if there are moderate differences between lookback periods, it should make a lot of sense for investors to spread their bets across a number of lookback periods.

In this way, each lookback period will have just 1/7th the impact on their total portfolio, but combined will still produce results very in line with a typical Dual Momentum investor's expectations. It would also help an investor avoid the effects of one underperforming lookback period in a given time frame.

Drawdowns

Sources: TheRichMoose.com, MSCI Inc., FRED

In this backtest, the Newfound 7 system is clearly better than the traditional 12M system in avoiding large drawdowns. Newfound 7's total maximum drawdown is -19 percent while 12M plunges lower at -25 percent.

The drawdowns are also more correlated. When Newfound 7's system is in drawdown, 12M is also in drawdown 92.4 percent of the time. Likewise, when 12M is in drawdown, Newfound 7 is in drawdown 94 percent of the time.

Over the entire testing period, Newfound 7 is in a drawdown slightly more often at 58.2 percent of months compared with 12M's 57.1 percent of months. However, this difference is clearly negligible.

Returns

Sources: TheRichMoose.com, MSCI Inc., FRED

When comparing the two models, again the 12M system does slightly better overall. 12M returns +16.1 percent compounded annually while the Newfound 7 system returns +15.5 percent compounded annually. Once again, the divergence comes primarily from the period around 1990.

Sources: TheRichMoose.com, MSCI Inc., FRED

The rolling annualized returns shown in this chart are indicative of how closely correlated Dual Momentum is—especially in the longer lookback periods. These two systems' monthly returns are 96.0 percent correlated across this test period.

While we see the familiar divergence in returns in the late-1980s and early-1990s, for the rest of the test period the returns are very tight with few meaningful divergences. It is a fair statement to say the overall performance for the two systems are extremely comparable.

However, the 12M system line (green) is noticeably above the Newfound 7 system line (blue) almost all the time.

Sharpe Ratios

When we compare the two system's average annual returns, the 12M system is better than the Newfound 7. This is naturally reflected in the higher overall returns we see with the 12M system. What surprised me is that the Newfound 7 system had a higher standard deviation.

This ultimately translated to the 12M system boasting the Sharpe ratio of 0.874 as shared previously. The Newfound 7 system was somewhat lower at 0.818.

Conclusion

What stood out most for me in doing this backtest is the robustness of Dual Momentum as a system, regardless of the precise lookback period used. The performance is truly outstanding with very admirable downside protection. Critique is absolutely fair, but Gary Antonacci deserves a lot of credit for sharing this with the public.

While I could have made an extremely long post comparing each lookback period from 1-month through 12-months with cluttered charts and graphs, I think the tests shared here are an adequate representation of how Dual Momentum performs on the longer side of the academically reviewed time-series momentum tests. I could quickly show more if need be.

On lookback periods, I think the evidence is pretty clear. For some reason, the 12-month lookback does perform better than shorter lookback periods including the 6-month. A 12-month lookback might not be better on every measure in every time period, but it's certainly an excellent option over full market cycles for self-directed investors.

When comparing 12-month DM to Newfound's 7 Period DM, I think Corey Hoffstein raised some important points. Scaling into and out of positions would have significantly reduced the maximum drawdown and also reduced the severity of most other drawdowns.

However, the important question is: Should a typical investor follow Newfound's suggested system to reduce risk in their portfolio?

Overall, I would say there are limited benefits to choosing Newfound's more complex system and there are some drawbacks.

First, we can't ignore that the two systems are extremely correlated at 96.0 percent. The difference in return stream offered by Newfound's system is very minimal. This is not a great way to diversify your portfolio. For perspective, U.S. stocks and International stocks only exhibit 70 percent correlation.

Second, the Newfound system would require significantly more trades. While I didn't count the trades for each system, a 6-month lookback trades twice as much as a 12-month lookback. Although trading costs are low, this is not insignificant. The true costs get higher for U.S. investors when considering short-term capital gains taxes in a non-registered investment account.

Third, the Newfound system does underperform the 12-month system on an absolute and risk-adjusted basis. While we may not be able to explain exactly why, a 12-month lookback period for broad stock indices has done extremely well. The higher Sharpe ratio, driven both by higher absolute returns and a lower standard deviation does not appear mere coincidence. The 12-month system stands above other lookback periods almost all the time. When it does underperform, it is not by much.

For a self-directed investor who has moderately sized retirement accounts, the best way to deal with the fragility of Dual Momentum is by adding a completely different non-correlated investment style as a significant component of your portfolio.

For example, a long-short strategy, managed futures, a tail-risk strategy, or an options trading strategy may nicely complement Dual Momentum. (I should point out this is precisely what I do in my personal portfolio, so take this with an ounce of bias.)

If a self-directed investor grows their personal accounts to a substantial size, there are some benefits to introducing additional lookback periods into the Dual Momentum portfolio of their portfolio. As this backtest demonstrates, additional lookback periods are likely to reduce overall drawdowns.

However, seven periods might be a bit extreme for any investor. For example, a quick test shows comparable downside mitigation with just four periods (6-month, 8-month, 10-month, and 12-month).

There are other considerations. An investor with large accounts is likely more interested in preserving capital (avoiding big drawdowns) than growing wealth. Multiple lookback periods as part of a trading system can be quite beneficial when this is the investor's focus.

A professional fund manager offering a Dual Momentum style fund to wealthy clients is likely to benefit the most from introducing multiple lookback periods. Professional managers get penalized heavily for drawdowns (fee revenue loss and client withdrawals), their clients are likely to have other funds in their portfolio for adequate diversification, and they have to make regular small trades in their course of business to account for distributions and fund flows.

In conclusion, no investor with even moderate wealth should put all of their money into a single investment model. Models can look great in backtests and may even perform very well going forward; however, we cannot predict that 12M Dual Momentum will provide 5-year rolling annualized returns between 3 percent and 45 percent going forward. It could be better, it could be worse.

While momentum-based strategies like Dual Momentum can provide a reasonable guideline for investing that is preferable to blindly holding the index, it does not guarantee protection from sudden drawdowns that could significantly exceed past events.

The best we can do is follow a sound premise, diversify across assets and strategies as needed based on our portfolio size and risk tolerance, and be prepared to invest counter to our emotional desires in the moment. The moment we override our strategy is the moment we no longer have a strategy. That is the real danger.

Comments & Questions

All comments are moderated before being posted for public viewing. Please don't send in multiple comments if yours doesn't appear right away. It can take up to 24 hours before comments are posted.

Comments containing links or "trolling" will not be posted. Comments with profane language or those which reveal personal information will be edited by moderator.

10 Replies to “A Better Look at Dual Momentum Fragility”

  1. Very nice analysis! I’ve also been pondering all this since these new articles came out.

    I believe Newfound’s argument is that 12M has historically been the best choice, but that could definitely change at any point in time (especially since we don’t really know what is special with that period, if anything). So, according to this premise, wouldn’t we EXPECT the “7 system” to underperform the 12M in backtests (even in terms of volatility)? However, if the conditions change in the future, the “7 system” would still work pretty much as well, whereas the 12M could begin to exhibit less desirable characteristics (i.e. comparable to the worst of the 7 periods in backtests). Maybe it could be interesting to add some curves to your last chart for the other 6 time periods, and see if the “7 system” could indeed be a good compromise, as a protection against unknown future changes in model behaviour, i.e. if the 12M becomes like the worst of these periods going forward?

    Still, like you, I am not convinced that this proposed approach really makes sense in practice for a retail investor due to all the trading involved. I wonder if aggregating the 7 indicators together would have a comparable effect to tranching the portfolio. I guess that would be similar to your previous DM approach. Thanks!

  2. Daren (Editor) says:

    Thanks for your comments. You’re correct that Corey acknowledges 12M has backtested well and Gary Antonacci used that historical information to select 12M as the preferred lookback period for DM. However, Corey has also argued that the other lookback periods perform equally well (within a statistically acceptable range) on a risk-adjusted basis. So he declines to say that 12M is “better”.

    This distinction is important in my view. If all the lookback periods are effectively equal on a risk-adjusted basis, then a Newfound 7 system should be basically equal to 12M (risk-adjusted) without the single period risks. We know the Newfound 7 system should underperform in pure returns, but the Newfound 7 system should theoretically be less volatile as it is often partially in stocks and partially in bonds which normally tampers volatility, especially when 12M is in stocks. This isn’t the case and for some reason, 12M does seem to be better for both returns and volatility looking back a long time (way past 1970), even if we don’t understand exactly why.

    However, I think the risk for the average investor is not necessarily that 12M somewhat underperforms say 8M or 9M going forward. There’s nothing wrong with these other long lookback periods and I would easily take 8M or 9M’s historical return profile over a 60/40 portfolio. In my view the biggest risk is being entirely invested in stocks and having the market go through a major 1987-on-steroids event. Maybe the Newfound 7 system would have you only 60 percent in stocks in that event (example). That’s why I say you should never put all your money in one strategy. This risk is also why aggregation is not as effective at reducing system risk, even if you find a certain combo that shows good backtested results.

    I could show a chart with all the lookback periods, but it basically ends up looking like a single fat, indistinguishable line. There are really no major deviations or terrible underperformance.

    There is actually a more real risk in the traditional DM system that I think is better to focus on fixing. I will post on this in the coming days.

  3. Fait enough. I still think there is an aspect to this that can’t really be backtested. In fact, the only potential justification to this process diversification that I can think of is if there were to be another flash crash in the vein of 1987 that you mention. It could happen that 12M would get whipsawed by such an event whereas other time periods not. Of course, with the “Newfound 7” approach, you would be sure to be affected by the event, but to a lesser extent if some periods avoid it. Whereas with a single time period, it’s all or nothing. That’s why I would definitely expect Newfound7 to have lower volatility on a hypothetical period in the future. (Keep in mind that in our real timeline, 12M DID avoid black Monday, so that’s why I was saying that backtested volatility may not be representative of the worst case). However, despite this, I’m also really not convinced that Newfound7 is a worthwhile strategy to implement.

    I’m very much looking forward to your next post. Thank you very much for these in-depth analyses, candy for the mind!

  4. Yanniel says:

    Another edge to explore is how these systems would perform when leverage is part of the equation.

    Gary covered in his book a version of DM using 1.30x leverage. My gut feeling is that Newfound 7 system‘s is more desirable in this case; if only to limit the drawdowns.

    Unrelated: Would’t the repetition of a Black Monday type of event be impossible given the adoption of “circuit breakers”?

  5. Daren (Editor) says:

    I’ll look into the leverage effects. I would expect the results to be similar to the results of 12M DM with leverage.
    I don’t know a whole lot about the market effects of circuit breakers and whether or not they actually work to reduce drawdowns. My thoughts are that circuit breakers could trigger several days in a row almost immediately after market open because of crowd behaviour pushing sell orders into the market as fast as possible at 930AM to beat the breaker. I don’t see a whole lot of people buying under those conditions if they have no clue what their fill prices are going to be. There should be some examples of their impact on single stocks.
    Under the current system a daily loss is capped at 20 percent for the index; that would limit the daily loss to 60 percent on a 3x leveraged ETF. However, a few days of 60 percent DD would have a very nasty impact.

  6. Yanniel says:

    Thanks for the feedback Daren.

  7. Yanniel says:

    “The tranches would be re-balanced each month to ensure equal impact of each lookback period.” Although empirically I sympathize with this approach I don’t think that’s the way NewFound approached rebalancing in that article:

    “But consider what happens if we try to neutralize the role of model specification risk and luck by diversifying across the seven different models equally (rebalanced annually). ” — NewFound

    Notice the “rebalanced annually” part.

  8. Daren (Editor) says:

    You’re right, I completely missed that detail when researching this post. (It wouldn’t significantly change the result though.)

  9. Given NewFound’s obsession with continuos rebalancing I am surprised that they chose to rebalance annually. It feels out of character.

    From an implementation point of view the monthly rebalancing is easier to accomplish in my opinion. One could approach the problem as if there is just one portfolio that gets rebalanced every month. The weighting of US, Int or bonds would be calculated as a sum of all the individual contributions as dictated with each lookback.

    Now, with an annual rebalancing one needs to keep track of 7 virtual portfolios. This is tricky because in reality these “virtual portfolios” would be held in the same trading account.

  10. Daren (Editor) says:

    I don’t think Newfound would promote a system like they suggested in the DM-GEM fragility post for any client. Rather, I think Corey somewhat hastily threw together a short backtest to make a point about a very specific form of risk observed in the DM-GEM system (and many other trading approaches). The point was a valid one, but not necessarily DIY investor friendly or reasonable given what Gary Antonacci was originally attempting to provide with DM-GEM.

Comments are closed