Follow-Up: Percent Exposure Donchian Channel Method
One of our more astute and diligent readers- Corey Rittenhouse- actually backtested the Percent Exposure Donchian Channel Method mentioned in my earlier Quick Take article on the subject. I will make the more detailed and nicely-presented PDF created by Corey available shortly. For now I will present a brief summary. As a point for new readers–Quick Take articles are simply ideas to help inspire system developers that I have not had a chance to test. Thus, I had no idea whether the system would actually work in practice. Nonetheless, the system concept was designed based on sound logic and theory–which is a key ingredient in system design. Corey created a summary of the results for the last 25 years tested on the following futures markets: ES, NQ, TF, GC, SI, HG, CL, HO, NG, RB, ZC, CC,CT, KC,ZW, LE, HE, ZS, SB, GBP, AUD, JPY, EUR, CHF. This is essentially a diversified basket of stock indices, commodities, currencies and fixed income futures which helps to verify robustness.
Summary Stats |
||||
CAGR | StDev | Sharpe | Equity Curve R-squared | DVR (R-square x sharpe) |
12.70% |
14.10% |
0.83 |
88% |
0.73 |
% Profitable | Average winner | |||
trades | /average loser | Worst month | Worst year | Max Drawdown |
31% |
3.22 to 1 |
-12% |
-13% |
31% |
Here is a year by year return history separated into two 10-year periods- 1999-2009, and 1988-1998:
2009 |
2008 |
2007 |
2006 |
2005 |
2004 |
2003 |
2002 |
2001 |
2000 |
1999 |
-1% |
50% |
-2% |
24% |
-13% |
6% |
29% |
11% |
-1% |
20% |
0% |
1998 |
1997 |
1996 |
1995 |
1994 |
1993 |
1992 |
1991 |
1990 |
1989 |
1988 |
11% |
5% |
28% |
5% |
13% |
10% |
13% |
-3% |
23% |
13% |
14% |
Here is a breakdown of the percentage of markets showing a profit for the system. This helps to verify robustness:
Long Only | # of Markets | % of Markets |
showing a profit | Profitable | |
Small Channels (lowest bet size) |
16 |
67% |
Medium Channels (medium bet size) |
18 |
75% |
Larger Channels (full bet size) |
20 |
83% |
Short Only | # of Markets | % of Markets |
showing a profit | Profitable | |
Small Channels (lowest bet size) |
14 |
58% |
Medium Channels (medium bet size) |
14 |
58% |
Larger Channels (full bet size) |
16 |
67% |
The annual returns of nearly 13% are fairly solid and likely superior to any buy and hold benchmark. The sharpe ratio of .83 is very solid, especially for a trend following strategy. Returns soared to 50% in 2008 while the world was melting down—displaying the positive kurtosis of trend-following strategies. Looking at the system from a risk-adjusted perspective, the Percent Exposure model shines. This is to be expected, as bet-sizing is adjusted for the channel breakout length and whether a trade is going with or against the long-term trend. A standard deviation of 14% is lower than both the S&P 500 and the commodity index. Furthermore, the max drawdown of 31% is lower than typical trend following strategies that trade both long and short. The worst month of -12% was tolerable even as a hedge fund strategy. The worst year of -13% was also very low for a trend-following strategy. The equity curve was very smooth with an R-squared of 88%, but becoming a little choppier in recent years as mean-reversion has become the dominant strategy. Finally, the system seems to be effective across a very wide range markets–especially on the long side.
Without going too deep into scrutinizing the results, I would have to conclude that the Percent Exposure Donchian Channel system is 1) fairly robust 2) a good risk-adjusted trend following system. There are several ways to improve upon it, while still keeeping robustness intact. Here are some suggestions:
1) Combine the Percent Exposure system with two momentum systems: RS Momentum: Buy only the top 3 markets ranked by a combination of 12-month price return and 1-month price return (12month rank + 1month rank)/2 . Contrarian Momentum: For a contrarian bent, you could buy the bottom 3 markets ranked by 3-year price return. Note that for short signals you would reverse this criteria for both momentum and contrarian. This should raise the returns on an absolute and risk-adjusted basis.
2) Use a trend and volatility filter: Only take signals if the ADX is greater than 25 or 30. Or you can use the R-squared of closing prices over the last 20 days as a proxy with a minimum value of 50%. As a volatility filter, only take signals if the 10-day ATR/ 10-day average price is below the 50th percentile over the last year. A signal is triggered only if it meets both of these criteria. These numbers have been chosen arbitrarily, but parameter testing should help you find more suitable thresholds.
Do you know what his exit parameters were? I think exiting when you break the opposite 50% channel (e.g. long on breakthrough 20d high and then close and breakdown of 10d low) may be too loose. This isn’t based on testing, just looking at charts for all sorts of combinations (sectors vs. SPY, country vs. country, commodity vs. commodity, etc.) What I noticed is that there were some huge winners (long-trending moves) coupled with many more reversion-type neutral or losing moves. For the big winners, I noticed that the trend was strong enough that the price line tended to hug the upper bands pretty closely for extended periods. For the neutral/losing trades, however, I noticed that the priceline would revert fairly quickly away from the upper bands, where it might stay neutral for quite some time, and the exist signal (breakdown) would be triggered only after a significant losing move had already occured.
So it seems that a better strategy might exist when the priceline breaksdown across the center line… This might do a better job of isolated and keeping the big winners (strong trends) while quickly exiting with minimal loss the potential losers (weak trends or flukes that will revert).
Nice job Corey.
DV, in an upcoming post, would you mind a little explanation of the Equity Curve R-squared calculation?
Thanks!
hi anon, i simply take an r-square which is a regression fit using excel of the equity curve. A perfectly straight line would be an r-squared of 1 and no relationship would be zero. I find this to be just as informative as the sharpe ratio, as it gives you a sense of how consistent things have been over time. A sharpe ratio can be really high, while the r-squared is mediocre or low– this can be due to a few really big years- or perhaps strong performance earlier in the dataset. The DVR combines these two to give a superior indicator.
dv