Systematic Trading research and development, with a flavour of Trend Following

## A Look back at Trend Following in 2010

#### January 26th, 2011 · 22 Comments · the State of Trend Following, Trend Following, Trend Following Wizards

2010 was a good year for Trend Following. The year closed on a high for both Trend Following Wizards and the Trend Following index of the simple strategies I track in the State of Trend Following report.

Now that all results are in, it is a good time to look back at the results, crunch some numbers and run some analysis. Let’s see if simple strategies managed to track the Wizards performance (hint: it looks promising)…

### How did the Wizards perform as whole in 2010?

In order to get an overview for 2010, I constructed a Wizard index aggregating individual fund performances. The index monthly return was simply calculated as the average of all Wizard returns.

Looking through this lens, performance in 2010 came at +18.91%, with an annualized standard deviation of 14.42% (based on monthly standard deviation of 4.16%). Below is the following monthly breakdown:

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
-4.93%
0.45%
4.33%
2.07%
-4.47%
1.47%
-1.11%
4.81%
5.32%
5.65%
-2.74%
7.54%

### How did the Wizards compare?

Results from Trend Following Wizards are often thought to correlate with each other. Empirically, this has appeared true throughout the year, with monthly results often being “in sync”.

In order to quantify this, I calculated a correlation matrix of Wizards tracked on the blog. Of course, correlation is far from being perfect, but with this data set, it should give us a good idea of similarity in the Trend Following Wizards performance. The matrix displays the correlation (using Pearson) between the Wizards monthly returns in 2010:

Wizard Correlation Matrix: click on table to zoom in

To reduce down the information, I have also calculated an alternate correlation measure between each Wizard and the Wizard index, based on monthly returns. The table below contains correlation of monthly returns over 1, 3 and 6 years:

Correlation 1-yr 3-yr 6-yr
Abraham 0.84 0.83 0.66
Altis 0.86 0.83 0.8
Bluetrend 0.83 0.86 0.85
Campbell 0.94 0.67 0.6
Chesapeake 0.91 0.79 0.77
Clarke.Cap 0.72 0.76 0.72
Drury 0.97 0.79 0.73
Dunn 0.82 0.87 0.83
Eckhardt 0.75 0.75 0.72
Emc 0.86 0.91 0.89
Hawksbill 0.64 0.62 0.64
Hyman.Beck 0.74 0.69 0.73
Jwh 0.04 0.61 0.64
Man 0.82 0.78 0.76
Millburn 0.87 0.84 0.83
Rabar 0.9 0.8 0.84
Saxon 0.79 0.64 0.67
Superfund 0.87 0.83 0.81
Tactical 0.56 0.79 0.82
Transtrend 0.87 0.83 0.8
Winton 0.82 0.84 0.8

And in visual form for the 1-year correlation on 2010:

Apart from a few exceptions, correlation numbers run fairly high across the board.

### How about the State of Trend Following report?

The State of Trend Following, which aggregates the performance of simple Trend-Following based-strategies, similarly had a good year in 2010.

One the drivers behind creating this report and index is the idea that a big part of Trend Follower returns are made up of style-beta, which are easier to replicate than pure alpha (see the Betafication of Alpha: towards a Commoditization of Trend Following? for more on this). This appears to be a plausible proposition when looking at the results from the correlation calculations for the Trend Following Wizards, above.

The headline return for the report index was +54.08%, which is not directly comparable to the Wizard index return: the volatility was much higher. As volatility is a function of leverage, applying a volatility-normalization to the returns will give us a better comparison. I adjusted the report index with a standard deviation-normalization (ie leveraged down the monthly returns so that their standard deviation match those of the Wizard index), which produced a return of 21.19%.

The three curves are plotted below:

Striking resemblance, don’t you find? Monthly return correlation is 0.84, which seems to support the idea of style-beta embedded in the Wizards performance.

### What Next?

In the light of these results, one could come to the quick conclusion that Trend Following Wizards do not add much value compared to fairly simple Trend Following strategies. I believe this would be over-simplifying the issue.

let’s not forget that the index tracked is a theoretical model, not a real-life trading program. It does not include the impact of slippage and commissions and all other constraints associated with running a real operation. Moreover, Wizards results are net-of-fees, meaning that their gross performance is superior.

However, there is probably some truth in the concept of some style-beta associated to Trend Following, which should be encouraging to all traders/investors wanting to join the “Trend Following party”, but not able to do it through CTA offerings.

Another evolution based on this concept could be the emergence of “Trend Following tracker funds”, aiming to capture the style-beta associated with Trend Following (with tracker-like, much cheaper fees), similarly to the recently launched GTAA Trend Following ETF by Mebane Faber. It will be interesting to see how the industry looks like in 5-10 years and if changes at all.

### Note on R  (for geeks)

I would have normally used Excel to do most of the correlations calculation (before I saw the “R” light). For the correlation matrix for instance, on top the fact that my version of Excel (2003) does not support color scales (to apply a specific color a the cell based on its value to generate a heat map), this would have required a fair bit of manipulation, formulas, etc. Instead I used R, and wanted to share how quick the whole operation can take. Here is the code that imports the monthly returns data, calculates the correlation table and plots it as a heat map:

```x<- read.csv("2010-returns.csv", header = TRUE, as.is = TRUE, sep = ",", dec="."); x_cor<- cor(x,,,"pearson") x_heatmpap <- heatmap.2(x_cor, Rowv=FALSE, Colv=FALSE, dendrogram="none", col=cm.colors(256), trace="none", cellnote=x_matrix, notecol="black")```

Three operations = three lines… Fanbloodytastic!

Note that the function heatmap.2 (improves on standard heatmap by, amongst other things, allowing values to appear on cells of heat map) requires an additional R package: gplots. A sample returns file can be found here for you to test this out. Go on and download R (Windows, Mac, or Linux) to give it a try if you do not use it already. You will not regret it.

Update as per Josh’s comment below: to use the gplots package you need to install it and load it (as it is not standard). For this, you can either use the Packages menu “Install Packages…” option and pick your CRAN mirror and select the gplots package (if you’re using RGui). Or you can do it from the console with the “install.packages” command as per code below, which installs and then loads the package (required before running the previous code):

```#install package from CRAN repository mirror install.packages("gplots")   #load package library(gplots)```
Picture credits: digitalcodi via flickr (CC)

Tags: ··

### 22 Comments so far ↓

• Jez: Do you know roughly what the impact of commissions and slippage would be on your back-tested systems?

Collective2 recently changed the default view for equity curves of the trading systems on its site to show the impact of commissions which I strongly agree with. As for the slippage, C2 has a “Keep after worst-case slippage” statistic. Some of the systems are negligibly affected by commissions whereas some of the systems trade much too frequently and are adversely affected.

• Fred,
I do not have much real trading data on slippage for these systems so we can only speculate on what sort of fills you would get and how it would impact the system based on the average trade duration/R-multiple average profits.

I did write a couple of posts on the subject earlier on however:
These look at the impact of varying slippage on two systems of varying timeframes.

How does Collective2 calculate worst-case slippage by the way? In terms of the tests run in the posts above, I would imagine a slippage of 100% (see definition of that figure in post), which would probably “destroy” any system, except the super long-term ones.

• From C2:

Keep After Worst-Case Slippage

This statistic describes the percentage of the system’s profits you are likely to keep if you are unlucky and receive the worst-case slippage on every trade. Thus, the higher the percentage, the more profit you are likely to keep, and the more favorable this statistic.

This statistic penalizes systems that attempt to scalp for profits that are small relative to the typical bid/ask spread of the instrument being traded. Keep in mind this is only one statistic among many, and it should be weighed in the context of other numbers you see here.

• soso

Thanks Jez for the great article.

Your conclusion, as I see it, simple trend following systems performance is very close to those of wizards, meaning the wizards performance is mostly extracted from beta rather than alpha?

Do you envision a continuous degradation of performance of trend following systems much like what happened to the turtles system – the 70s compared with the 90s and afterwards.

Thanks

• Hi Jez,

There is the tutorial on the R website itself (http://www.r-project.org/) under “Manuals” called “An Introduction to R”.

An excellent tutorial, not nearly as over-whelming as its 100+ page count might suggest:
http://heather.cs.ucdavis.edu/~matloff/R/RProg.pdf

A beautifully written literary conceit:
http://www.burns-stat.com/pages/Tutor/R_inferno.pdf

Useful stuff here too:
http://www.r-tutor.com/

The breakthrough I had in R was when I finally started thinking in terms of variables as vectors, matrices and arrays rather atomic variables. This is where R becomes incredibly fast (e.g. close.price:= a vector of the daily closing prices for the last 15 years can be treated as a single variable). I remember re-visiting an early prime-number generator I had written, and “vectorising” it – it took 1/20 of the run time to generate the same number of primes.

Having said that, a lot of work in system development involves loops – each day is dependent on the previous day. This does not lend itself to the vector format. Do you have a preference for a scripting language that does handle loops well?

• @Josh – yep, updated the post. Thanks for the pointer.

@soso
Well, I do not think this is clear-cut. For a start, we could say that Trend Following Wizards do generate alpha (over-performance to the style-beta benchmark) when considering gross returns, but keep it for themselves via their fees. I could also well have been “lucky” in my market selection, which I think is part of the “skill” of a Trend Follower. Another portfolio might have resulted in a different performance. I think this is more subjective than other clear alpha/beta categorisations, but I do believe in the idea of style-beta driving a big chunk of Trend Following.

Whereas the Turtle Trading system’s performance might have degraded (for several reasons), Trend Following as a whole did not show much sign of erosion. Check this Trend Following Wizards historical chart.

In terms of predicting what Trend Following will do in the future, this is a tough call, but – at a very simplified level – I believe that Trend Following captures inefficiencies from market irrationality and/or opportunities from market hedgers. This sum of inefficiencies/opportunities could be considered as the strategy’s source of returns (whether we call it alpha or style-beta). Is this sum of potential profit limited? What would happen if the space becomes more crowded? Could more participants sharing the same amount of alpha result in lesser returns and “performance degradation”? Could this be driving away some Trend Follower participants and allowing the remaining participants to then grab a bigger “share of the pie” (and start attracting newcomers again)?
Maybe such a cycle is bound to repeat in a sort of adaptive market evolution (check the Adaptive Market Hypothesis for further discussion in this area).
I can only speculate but human nature (with all the market impact that it has, including on Trend Following) is not likely to change much over the coming centuries, I shall think…

@Eventhorizon
Thanks for sharing these links and your experience. I’m still very new to R but really impressed with it.
I’m more of a compiled language developer so my experience using generic scripting languages is very limited.. Josh, who does a lot of his work on financial stuff and is an expert in R might be able to expand on how he handles loops..

• Hi Jez,

Another excellent post with conclusions backed by numerical results, something uncommon nowdays.

My comments were long and I have posted them in my blog:

—Commoditization of Trend following will not affect trends—

These are my views as related to alpha, beta and the commoditization of trend following:

I started 23 years ago as a trend follower. My system was based on a moving average envelope and we, my partner and I, traded commodity and currency futures. It returned over 80% on a 200K starting capital in the first year. I was excited but I have missed something. I was lucky because the year I started was a good one for trend followers. The following year it was a choppy market. Consecutive losers started accumulating and drawdown increased. I remember I could not sleep at nights. I had a monitor next to my bed and I would wake up every 5 minutes to see how currencies were trading overnight. The stress was too high, especially when you have profits more than when you have losses. When my system gave back half of its profits I stopped it and I abandoned trend following. I decided to look for a method that would replicate trend following results through short term trading.

Trend following is all about discipline. It is not a very smart job (alpha), it is a job that requires a good nervous system, no distractions and sticking to a plan. You pay a manager for having those rare qualities to deal with this risk (beta). Actually, in this respect, the fee should be higher than what you would normally pay a short term trader, because there are many more smart people in the world than there are disciplined ones for this particular job.

Moreover, trend following ETFs have the same problem of exercising discipline when trading them. Because there is a trend following ETF that does not mean it will impact trend following since it is the same people that can do trend-following that can also trend follow it in the first place. Thus, trend following can only be impacted if the number of disciplined traders increases and even then that may affect only the volatility of trends but not their development since the latter depends on macroeconomic factors rather than on traders themselves.”

http://www.priceactionlab.com/Blog/2011/01/commoditization-of-trend-following-will-not-affect-trends/

• Michael,
The post was not too long, I included it here so that the discussion can be continued here and shared with other readers as well.

As per my earlier comment, I am not sure I agree with the fact that commoditization of Trend Following would not affect trends (although I do not have anything else than my opinions to back this up… I might be wrong!). From a high-level point of view, I see alpha more likely to be a finite amount across the whole market, meaning Trend Following (or any other strategy) has a max capacity at any one time (but which varies over time based on multiple factors, ie. market ecology). This goes against the common wisdom that the more people follow trends, the stronger the trends will be. But if you think about it, more traders/Trend Followers might also mean more slippage at entry, more volatility/noise, meaning less profitability potential (from a risk-volatility adjustment point of view), or potentially more whipsaws from trend reversals – all of which would contribute to decreased returns.
I believe this is a likely reasion why the Turtle Trading system has fared much worse in the last decades compared to the 70’s (ie that specific strategy might have become overcrowded)

With regards to the LTCM comment made by D. Hom, I believe that story did also highlight a max capacity limitation for their specific strategy.

• soso

Turtle Strategy failing – same here, I can’t see anything else other than becoming overcrowded.

But this raises another question, since it performed so poorly in the past 10-20 years (if I remember correctly, I don’t have the source at hand) then why people KEEP BETTING (large amounts of) money on it?

I’m thinking that if people wouldn’t bet so much money on the turtle strategy then the superior edge will return (at least partially).

Because if overcrowding was the main reason for performance degradation then crowd fleeing would put the edge back in.

• Jez,

Very interesting discussion triggered by your very interesting articles.

IMO, slippage is not an issue in trend-following. Actually more participants in the market translates to higher liquidity and tighter spreads. Volatility is an issue because it causes intermediate trend exists and thus limits trend following potential. Actually this is what happens in my view: bad trend followers, in terms of exercising discipline to follow the simple techniques involved, create volatility due to their fear that makes the task of those with discipline more difficult.

As a matter of fact, I believe that trend following methods will work forever but as long as more individual traders and investors try to implement them, and these participants are emotive, the volatility they create will continue to adversely affect professional trend-followers that trade in automatic mode.

I have written in one of my books that since the opening of the markets and the fall of communism, many joined trading from eastern European countries with small accounts and tried to exploit mathematical methods for making quick fortunes. Because they cannot exercise prudent risk and money management and because they are ruled by fear and greed, they created volatility that affects the performance of some professional trend followers. Actually, some of the professional moved to HFT to take advantage of those retail traders. Now, I think most of those newcomers have been washed out and markets will slowly come back to pre 2000 mode.

Of course, I am talking qualitatively based on my experience trading the markets. In another post in my blog I will try to explain in a quantitatively why some trend following methods fail.

• Motomoto

Nice articles Jez, (I am slowly settling back into the sunshine)
As food for thought, regards the emotive aspects of new players being quasi trend followers, and merely adding to the volatility – while I have not done much reading on this – I wonder if the new ETFs – which seem to be getting killed on the contango, have had much or more of an effect?

• Jing

I don’t think overcrowd is the only reason for turtle system’s degradation. I don’t think there is a single system which can work forever because the market is indeed always changing. One reason of market changing is the overcrowd of particular strategy, but there are many other reasons, including the popularity of short-term trading, day-trading, high-frequency algorithm trading, the massive participance of future, commodity trading compared with a few years ago. All these contributes to the noisier, choppier markets. I don’t think the increasing number of trend followers will lead to stronger trends. On the other hand, more trend followers will make originally smooth trend much choppier, thus increasing the entry risk (higher ATR at entry) and increasing the possibility of premature stopping out in the middle of a trend (the trailing stop needs to be wider in order to hold position until trend reversal, which also negatively affect the return when trends end). For example, suppose there is an originally smooth trend in which the price develops from 10 to 20 in 3 months very smoothly. When the trend followers increase, the price may increase from 10 to 15 fiercely and quickly in the first week because all traders buy at new high, and then quickly fall back to 12 because of profit taking and many tight stop losses; and then price increases from 12 to 17 in the second run because of solid fundamental reasons and then drop back to 15, and then finally reach 20. I know that is an extreme case, but I think more trend followers will make originally smooth trends much choppier. These choppy trends will substantially adversely affect many trend systems. Even there are less and less turtle system followers, the above changes and market price behavior changes may lead to the permanent failure of the original turtle system.

• Hey Motomoto,
Good to “see” you back on the blog! Hope everything is going alright in the sun..
The little reading I have done on ETF impact seemed to indicate that there could well be an impact, if only at “roll-over” time for example..

@Michael, Ying, re: slippage, whereas I agree that _in general_ more trading activity/participants increases liquidity; more Trend Followers, with their usually common entry/exit points, would increase pressure in the “bad/same” direction, which should logically result in more slippage.
I think it is Linda Raschke, who designed a system, cheekily named Turtle Soup, which aimed to take advantage of this (and mean reversion) by fading Trend Following breakouts.

But same as for you this is only a qualitative point of view so I’ll watch out for your “quantitative” post, Michael…

• @Jez, re: slippage, whereas I agree that _in general_ more trading activity/participants increases liquidity; more Trend Followers, with their usually common entry/exit points, would increase pressure in the “bad/same” direction, which should logically result in more slippage.

Excellent answer! It can be viewed as slippage due to the volatility that these massive actions cause. We saw some of that action back in late 2007 and during 2008. We may see some now. I think yesterday was maybe the start.

Here is the link to my quantitative analysis about trend following difficulties:

P.S. My apologies for the posting of links. I have added you to my blogroll to make up for the outflow.

• Marco

Hi Jez, have you done a backtest of the strategies on the same contracts going back a few years?
if so, how do you account for roll-overs? how do you choose which contract to trade? and the same for the current test, could you explain how you choose the contract month?
Do you have a list of the trades somewhere so it can be seen?

• There is no accounting for slippage/commissions in the report and the same applies for roll-over. The actual roll-over trigger is based on Open Interest, so that the contract traded is the one with the highest OI. To switch from one contract to the next, OI has to shift first (the actual setting I use requires two consecutive shift/triggers off OI to switch contract).
The report is more intended to be a high-level look at Trend Following and I am not really planning on releasing the list of trades.

• Marco

thanks for the info. why wouldnt you post a chart with a backtest going back as many years as possible? Performing well in the short term can be interesting, but if there are massive drawdowns in the past, it makes a big difference.
thank you.

• The main goal of this report is to build an “index” of Trend Following to track its live performance. As such the index is not intended directly as a tradeable strategy but more as a monitoring tool.

Moreover the historical back-test would suffer from hindsight bias since the results from 1990 to 2009 have been used to normalize the results from 2010 onwards, using past MaxDD figures. You can read about it and how it was done here:
The table in that section also shows the long-term performance of all individual systems (from 1990 to 2009), which you might be interested in, as per your comment.
~Jez

• Marco

thanks, that is very helpful Jez

• Hi Jez,

I think I figured out an answer re: colors in heatmap.2

Set breaks=seq(-1, 1, length=21) to fix the color range into 20 buckets from -1 to 1. This avoids the problem of Red = -1 on one chart while Red = -.5 on another.

Also set symkey=T (it supposedly defaults to TRUE, but doesn’t actually seem to) to force the color key to display the entire palette from -1 to 1. This avoids the problem of the histogram’s appearing centered on a chart, when it is not centered on the range [-1, +1]

• Thanks for the tips Eventhorizon.
I was looking at doing comparisons over different time periods and was having to resort to hacks to “normalize” the values…