Overview

Investors typically seek to exploit the power of diversification: It is possible to improve risk-adjusted returns simply by combining different diversifying strategies of similar risk and return. In this paper, we ask if this approach should be applied to trend following models. There are many different methods of creating trend following models, and we consider a wide range of common approaches to investigate whether the combination of several of these models can lead to improved performance, or whether there is a better way to construct a trend following system.

We show how different trend following models, when applied to the same portfolio of markets and operated at similar speeds, generally have high correlations with each other and thus offer limited diversification benefits. We also demonstrate that a better approach to trend following is to apply a holistic methodology which aims to capture the most effective features of many different techniques and to integrate them in a single high-calibre model. We demonstrate that Aspect’s own trend following model, which has evolved over many years through innovative research that essentially combines the benefits of multiple different approaches, is such an example.

Introduction to Different Trend Following Models

In this paper we consider a group of 13 different trend following ­models. All the models have been applied to the same portfolio of 
146 markets and have been parameterized such that they all capture medium-term trends of approximately two to three months in duration and generate the same levels of annualised return volatility. For the purposes of this analysis we do not include the impact of trading costs, as we are purely focused on the difference between various trend following approaches, and they will be broadly equivalent across the set of consistent time-scale models.

The origins of these 13 models are varied, but all are in the public domain. They include models popularized by the “TurtleTraders” of the 1980s,¹ models that have been regularly cited in recent years in academic literature, and a number of other well-known trend-capture techniques including look-back and look-forward windows, moving average approaches, technical indicators, and other statistical methods.

Figure 21.1 compares the performance of all these models over the period January 1999 to June 2016, and Figure 21.2 compares the models’ information ratios, a measure of return for a given level of risk. The performance of the different models is generally fairly consistent and the average information ratio achieved in simulation by the 13 trend following models is 0.95, with the maximum being 1.10. (Explanations of all abbreviations used in the figures can be found in the Appendix of this chapter.)

Diversification between Different Trend Following Models

When different diversifying strategies of similar risk and return are combined, better risk-adjusted returns might be expected through the lower volatility achieved as a result of the diversification. In this section we consider whether the combination of different trend following models can lead to improved performance.

While the 13 models we have introduced represent a broad range of different approaches to systematic medium-term trend capture, we find that they are highly correlated to each other. Figure 21.3 shows the correlations between the 13 models, again over the January 1999 to June 2016 period. The lowest correlation we see is 67% (between the MOP and Turtle-ATR models), while the average correlation between models is 89%.

As a first step in investigating whether a combination of these models is preferable, Figure 21.4 shows the performance of a strategy which is a simple average of all the 13 models, obtained by blending the output returns from each on a daily basis (and adjusting for the slight reduction in volatility that results).

We see from Figure 21.4 that the performance of the averaged strategy is comparable to the performance of some of the individual models. In addition, its information ratio is 1.00, which is only slightly better than the average of the individual model information ratios of 0.95. As a consequence of the high levels of correlation between the individual models, the diversification benefit that arises from combining them is only slight.

To investigate this further, we consider all possible equally weighted combinations of the 13 models under consideration and determine the average risk-adjusted return as the number of models combined is varied. We start with just one model (of which there are 13 to choose from), and then in turn consider combinations of two, three, four, etc. up to the final combination of 13 models. Significantly, Figure 21.5 shows that there is very little diversification benefit to be had from combining models, as the impact on average risk-adjusted performance is insignificant.

Aspect’s Approach to Trend Following

The results of the previous section suggest that, due to the high levels of correlation, diversification between different trend following models is illusory. Instead, Aspect takes the holistic view that if the goal is to maximize performance from trend following, it is better to build the best possible single trend following model that integrates distinguishing features of many different approaches.

Aspect’s trend following model has been developed over many years of rigorous, scientific, and hypothesis-based research, incorporating features of multiple different approaches in a carefully considered, coherent framework. (Aspect’s trend following model referenced in this paper is that which has an allocation of 80% within the Aspect Diversified ­Programme.^*) One of the major research innovations in Aspect’s history was its move in 2005 from a multi-model approach to trend following to a single, holistic approach.

As an example of an Aspect innovation capitalizing on a feature of one family of models, some breakout models use methods to calculate a high and low range for their channels, but do so with a coarse binary signal construction. This observation led us to research more thoroughly the usefulness of considering channel data as part of our “data processing” stage, and has enabled us to refine the way our trend following model deals with specific market moves. This approach has enabled us to make the most of the diversifying features of breakout models, while at the same time avoiding its less desirable limiting features (i.e., its coarse binary signal construction).

Our research has led to a number of key trend following innovations over the years, which fall into three main stages within our single holistic model.

• Data processing: This stage deals with the way in which market data is processed in order to create the most appropriate data series for the trend measurement stage.
• Trend measurement: This stage filters the processed data, in order to measure the strength and direction of trends.
• Position mapping: Having determined appropriate trend measurements, these are then mapped to appropriately sized positions.

Figure 21.6 shows the performance of a base trend following model built by Aspect which still benefits from many of Aspect’s portfolio construction, position-sizing, and risk management processes, but uses a simple trend following approach. Its performance is similar to some of the 13 strategies considered earlier. The chart also shows the simulated performance enhancements that arise from Aspect’s research improvements over this base trend following model. Each of these enhancements adds value to Aspect’s trend following model incrementally and consistently; the simulated information ratios also rise steadily from 0.76 for the base trend following model, to 1.41 for Aspect’s full trend following model.

Aspect’s Model Compared to Other Trend Following Models

Our principled approach, in which we observe different features of trend following, learn from them, and integrate innovations to our model based on this research, is a key part of Aspect’s systematic investment process. In our view, this approach is superior to combining multiple different trend following models.

In Figure 21.7 we compare the 13 models with Aspect’s trend following model, developed over almost 20 years of evolutionary research. We see that Aspect’s holistic trend following model outperforms all of the 13 models over the period under consideration. Aspect’s information ratio of 1.41 is also superior to those of the other models.

The results support the argument that if the goal is to maximize performance from trend following, it is better to build the best possible single trend following model that integrates features of various approaches rather than relying on diversification from different models.

Additionally, if we subdivide the period into five-year windows, we see that Aspect’s integrated approach outperforms in each individual window, as shown in Figure 21.8. By contrast, among the 13 other trend following models, there is almost no consistency in performance between the different five-year windows.

Finally we demonstrate that Aspect’s model cannot be improved by adding any of the 13 models considered earlier. Figure 21.9 shows the effect on the risk-adjusted return when Aspect’s trend following model is blended with any of the other 13 models.

Again we see that the best option is to choose only Aspect’s holistic trend following model, as any combination with other models will degrade its performance.

Conclusion

This paper considers the question of how to build the best trend following system: either to focus on building a single, high-quality model that combines the best features of other trend following approaches, or to adopt a multi-model approach to trend following, relying on diversification between models to improve the overall risk-return profile.

We have considered a wide range of different systematic models that all capture medium-term trend following, and have investigated whether any combinations of these models provide the best outcome. Given the high levels of correlation between the strategies, we actually find that there is very little diversification to be had.

In summary, the number of individual trend following models that comprise a trend following portfolio is not in itself a measure of its superiority. The best approach is to focus on building a single, well-researched trend following model that integrates key features of many different trend following approaches in a superior, coherent framework. Aspect applies an ongoing research effort in order to continue to enhance its trend following model in this way.

Chart Disclaimer

The 13 trend following models and aspect diversified’s results are based on simulated or hypothetical performance results that have certain limitations. Unlike the results shown in an actual performance record, these results do not represent actual trading. Past performance is not necessarily indicative of future results.