Trend Model August report card: An invaluable lesson on model design

While no one likes to see drawdowns, they happen. What's invaluable from drawdowns is the lessons learned about the investment process. In this case, I gained some worthy insights about the tradeoffs made in my model building process. The latest performance update on my long-short account based on my Trend Model signals (see An intriguing Trend Model interim report card) can either be encouraging or FUGLY or encouraging. The Trend Model based trading account lost -16.7% in August; the one-year return was -5.1%; and the return from inception of September 30, 2013 11.6%.

While this is no excuse, I suppose I can be comforted by good company as lots of big name hedge funds got carried out on their shields in August, starting with David Einhorn's Greenlight Capital with a loss of -5.3% in August and -14% YTD (via Dealbook):

Leon G. Cooperman, founder of the $9 billion Omega Advisors, told investors on Aug. 21 that he was down 11 percent for the month. Last week, William A. Ackman surprised investors by announcing that all the gains for his Pershing Square Capital Management hedge fund for the year had been wiped out by “significant volatility” in the global markets. Mr. Ackman’s multibillion-dollar hedge fund was down 7.3 percent for the quarter and 4.3 percent for the year as of Aug. 25.

The world’s largest hedge fund, Bridgewater Associates, led by Raymond Dalio, told investors that its Pure Alpha fund was down 4.77 percent as of Aug. 21. And the flagship fund of Third Point, led by Daniel S. Loeb, is up just 0.6 percent for the year, after losing 5.1 percent in August.

I reiterate my disclaimer that I have nothing to sell anyone right now. I am not currently in a position to manage anyone`s money based on the investment strategy that I am describing.


Trend Model description
For readers who are unfamiliar with my Trend Model, it is a market timing, or asset allocation, model which uses trend following techniques as applied to commodity and global stock market prices to generates a composite Risk-On/Risk-Off signal (risk-on, risk-off or neutral). I have begun updating readers on the Trend Model signals on a weekly basis and via Twitter @humblestudent as new developments occur.

The chart below shows the actual (not back-tested) changes in the direction of the signal, which are indicated by the arrows, overlaid on top of a chart of the SP 500. You can think of the blue up arrows, which occurred when the trend signal changed from negative to positive, as buy signals and the red down arrows, which occurred when the trend signal changed from positive to negative, as sell signals.

Trend Model Signal History

A proof of concept
While the signals from the above chart representing paper trading is interesting, there is no substitute for actual performance. As a proof of concept, I started to manage a small account that traded long, inverse and leveraged ETFs on the major US market averages and, on occasion, sector and industry ETFs. Trading decisions were based on Trend Model signals combined with some short-term sentiment indicators. The inception date of the account was September 30, 2013, For more details on how the Trend Model or how the account is managed, see my post Trend Model FAQ).

When evaluating the performance of this trading account, keep in mind that this is intended to be an absolute return vehicle. While I do show the SPY total return, which includes re-invested dividends, for illustrative purposes, the SP 500 is not an appropriate benchmark for measuring the performance of this modeling technique.


Trade-offs in model design
When I refer to the "Trend Model", it is actually a model with two components, The "Main" model, which is a trend following model applied to global commodity and equity prices. The Main model worked well, but returns were subject to wild swings. I had warned about a possible correction in early February (see A correction is brewing) but the stock market continued to grind upwards to reach new all-time-highs. I had received so much criticism that I wrote a post outlining my bear case (see Why I am bearish (and what would change my mind)). Another drawback of trend following models like the Main model is its slow moving nature and delayed response. As an example, Mebane Faber's trend following model just went to 100% cash, albeit a little late but in anticipation of further weakness.

One of the ways that I used to address the issue of short-term return volatility was to supplement the Main Model with a "Trading" model, which is composed of sentiment and overbought-oversold models to spot short-term counter-trend moves. The purpose of the combination is to prevent the Main model from getting into positions when the move is getting over-extended. Thus, the Trading model will prevent the overall model from buying when the market is overbought and shorting when the market is oversold.

That was the root of the trouble in this latest episode. The Main model had spotted the downtrend early, very early, but the Trading model prevented it from shorting the market because sentiment had moved to a crowded short and the market was mildly oversold (see You can`t hurry tops! and Groundhog Day). When the stock market downdraft began, the account was long, instead of short as dictated by the Main model. Unfortunately, oversold markets can get more oversold and readings became more and more extreme. Analysis from Bespoke showed that you would have to go all the way back to May 1940 to see go down this quickly.

When the market bounced last week, I had run out of ways to say "oversold" and pointed to analysis from Richard Chappell writing as Springheel Jack that indicated that the SPX had only seen these kinds of oversold readings in 2008, 2004, 2001, 1987, 1962 and 1940. Only in 1940, when the market reacted to the Fall of France, did stock prices go even lower. It became evident to me that when statistics get thrown around comparing current conditions to the Lehman Crisis, the Crash of 1987 and the Fall of France that the market is seeing an extreme tail-event.

In effect, the addition of the Trading model, with its sentiment and overbought-oversold models, raised daily and weekly returns at the price of taking on extreme tail-risk. 

The combination of the two models was a conscious decision which I take full responsibility. Is the drawdown, which I could have missed and turned into a very large (20%+) gain a bug or a feature?

Models all fail, we just have to know their vulnerabilities. Despite the negative returns, which I described as FUGLY earlier in an absolute sense, I nevertheless believe that I took the correct decision in making these trade-offs.


Model performance in context
Overall, the returns of this strategy remains promising in the context of what is asked of an absolute return strategy:

We can now quantify the extreme drawdown associated with this strategy. One of my key questions in my manager due diligence process is to ask under what circumstances does the investment approach fail. This strategy will truly blow up with a -30% to -80% month when extreme tail events like the Fall of France happens. Under more "normal" circumstances, maximum monthly drawdown should be limited to around 10%. How worried should investors be about blow-up risk events that last happened 75 years ago? The return distribution chart below tells the story. There are really two nevdistributions at work here. The more "typical" monthly returns show a bell-shaped Gaussian return distribution centered at about 2-4% a month, while the extreme tail-risk came in at -16%.

Would you rather have a return pattern like the one above, with a median monthly return of 1.4% and a distribution that's slightly skewed to the right and tail-risk of -16% that happens probably once every 10-20 years, compared to the SPX return pattern below? As the chart below shows, the stock market has a median return of 0.9% per month, which is lower, and a skew that tilts to the left, or negative.

Long term returns are acceptable, despite the recent volatility. Despite the difficult returns experienced this year, return from inception (September, 30, 2013) is still a respectable 11.6%..

The Calmar ratio, or reward-to-pain ratio of return to max drawdown, is 0.7, which is still respectable  By comparison, during the 2008-2009 period saw the drawdown for equities come in at about 50% and most balanced funds saw losses of about 20%. Assuming a Calmar ratio of 0.5 and working backwards, it would be hard to envisage long-term return expectations of 25% for stocks and 10% for balanced funds today. As time progresses, I would expect the long term returns to rise again and the Calmar ratio to rise again.

Returns continue to be consistently positive, with a 65% monthly batting average.

Returns are highly diversifying compared to major asset classes. They are uncorrelated with equities (correlation of 0.18 with SPY) and bonds (0.13 with AGG).

A bug or feature?

I have said before that the market environment in 2015 has been challenging for the Trend Model trading strategy. We now know in detail of the pros and cons of this trading strategy.

Is the tradeoff of better average monthly returns for tail-risk a bug or a feature? The recent market downdraft showed the "feature" of this strategy. Given sufficient diversification, trading strategies like these can add value to a portfolio over time. It works, on average, but don't put your eggs in this basket.

Readers who want to monitor the signals of the Trend Model to subscribe to my blog posts here, which include Trend Model updates, or follow me on Twitter @humblestudent.