As reported recently in a highly acclaimed blog, Dual Momentum GEM with a lookback period of 12 months lost more than the benchmark last year and forced in some cases overriding the systematic model. We show in this article that these failures should be expected since this and related simple momentum models lack intelligence to adapt.
For more information about the dual momentum strategy please see the references below. Our implementation uses the following ETFs: SPY, VEU, BND and BIL. The logic is very simple as follows:
If in the last n (usually 12) months
SPY outperforms BIL and also VEU, then buy SPY
SPY outperforms BIL but not VEU, then buy VEU
SPY does not outperform BIL, then buy BND
Let us stop and think about it for a moment: the GEM trading strategy switches between US and World ex-US equities based on whether US equities outperform T-Bills. Otherwise it invests in bonds. There are several assumptions and maybe wishful thinking:
- No market impact assumption
- Minimum impact of parameter choice (lookback)
- Strategy has ability to adapt to changing market conditions
The above assumptions apply to most trading strategies and in many cases they are all violated. In some cases the violation is so strong that guarantees failure. An example of failure due to assumption 1 is in the case of large funds. Due to the market impact historical analysis of strategies fails to be a guide of future performance.
In other words, if a strategy like dual momentum GEM is used by some large funds, then there will be impact that may essentially negate expectations because the market has practical constraints in delivering alpha. It appears that some do not realize the severity of this assumption when they recommend strategies to the public: if many participants use these strategies, their effectiveness could diminish and in many cases it does. One problem is that it is hard to measure the impact on market popular strategies have.
On the other hand, impact due to parameter variations can be measured although it is rarely done. For example, below is a chart that shows how the maximum drawdown of the dual momentum GEM strategy with the implementation mentioned above has varied since 2008 if we vary the lookback period from 1 to 24 months.
It may be seen that for lookback periods below 5 months the maximum drawdown is larger than 25% and even gets as large as 55%. However, although the maximum drawdown remains below 20% for lookback periods between 10 and 12 months, it then increases for values between 13 to 15 to around 25% and then decreases again in the range 15% – 20%.
Now, there is a large psychological difference between a 19% drawdown and a 26% drawdown and also potential impact on decisions to override the strategy. However, the most important conclusion from the above chart is that there is certain randomness in the drawdown/lookback period relationship that is due to particular paths that may not repeat in the future. The variation shows that the strategy is in the mercy of the market not the other way around. Therefore, it appears that the strategy does not have the ability to adapt to changing market conditions, which is a failure of the third assumption.
The results of the above failure were obvious last year. The chart below shows the return of the strategy net of $0.01/share commission as a function of the lookback period.
As first noted in Reference 1, the return is all over the place depending on lookback period choice. In fact, although the strategy returned about -8% for the popular choice of a lookback period of 12 months, for 17 months the return was around -14% but for 10 months is was around +1% !
The above observations credited to Reference 1 show that the intelligence of this strategy, or its ability to adapt to changing market conditions, is based on the intelligence of choosing the proper lookback period in advance, which is impossible in general because no one knows the future path of the market. Therefore, the strategy lacks intelligence.
Below is the equity curve of the strategy since 2008 and a comparison to benchmark SPY total return for 12 month lookback period:
The strategy did well during the 2008 crash but that was due to anti-correlation of stocks and bonds. But this may not be the case if this anti-correlation is violated in the future.
Although the strategy underperforms the benchmark in absolute return basis, it outperforms on risk-adjusted basis but only in the limited time period considered, which represents a specific market path. Maximum drawdown is around 17% versus 55% for benchmark. What is the probability of a drawdown larger than 25%, for example?
This is a difficult question to answer because we have limited sample. We can try a Monte Carlo simulation but results will be ambiguous due to possible violation of certain key assumption such as normality and no serial correlation. However, this is the best we can do and the results for maximum drawdown are shown below:
According to this analysis, the probability of a drawdown larger than 25% is about 30%. The result assumes more or less the same market conditions will prevail, something that may not be the case. In fact, given enough time a 50% drawdown may be the certain event.
Should investors and fund be using these simple strategies?
It appears that these strategies were developed based on hindsight, some optimization and also wishful thinking, which are all normal by the way. Using them properly assumes knowledge of limitations and assumptions. In Reference 1 a way of minimizing the fluctuations is proposed based on averaging different implementations. This increases complexity and makes the transition from strategic to tactical even more complex. In essence, the investor turns into a trader and this can lead to loss of discipline. For most people the best choice is going passive and this is also what the numbers show.
If you found this article interesting, I invite you follow this blog via any of these methods below:
If you have any questions or comments, happy to connect on Twitter: @mikeharrisNY