The past two articles have detailed the current state of artificial intelligence in trading. We discussed the reemergence of interest in neural networks and outlined newer methods, such as kernel regression.
Having learned from our past mistakes, we’re ready to apply this technology to develop a new generation of trading systems. One framework is built around small, manageable logical components. Small components combined into larger blocks that are combined into systems combat the problem of asking one uber-system to do too much. Meta controls help the blocks work reliably.
We’ll start this discussion in familiar territory for most traders: trend following. This popular approach encompasses a family of methods that trade in the direction a given market is moving. There are many different trend-following techniques. Each performs using the same general concept, but they are not 100% correlated. Let’s now discuss three trend following approaches.
We’ll use a diversified portfolio in our examples, including the following 10 markets: cotton, the dollar index, orange juice, coffee, silver, Treasury bonds, crude oil, natural gas, the Japanese yen and the euro. The data company Pinnacle merged data that combines pit and electronic data will be used. This data reflects the highest activity, whether it is electronic or pit. This switch reflects reality for most traders. Tests will run from Jan. 3, 1980, through Nov. 26, 2009. We will charge all tests $25 per trade commissions and $75 for slippage.
Channel breakout is one of the simplest of trend-following systems. Richard Donchian expounded on the system in a 1970 booklet. He represented his four-week rule using price channels. The top boundary was the highest high for the previous four weeks, and the lower boundary was the lowest low for that period.
He advocated always being in the market, covering short positions and entering long ones when price exceeded the top channel, while liquidating long positions and entering short ones when price declined below the bottom channel. The default 20 setting on most Donchian channels reflects that four-week setting, with 20 days composing four trading weeks. The code for this system is shown in “Betting on a breakout.”
In this case, don’t confuse simplicity with ineffectiveness. Channel breakout is one of the most reliable trend-following methods. Today, almost 40 years after Donchian popularized the four-week rule, this strategy is still profitable over time even in markets that didn’t exist in the 1970s.
The method has been consistently profitable since 1980, making money 22 out of the 30 years since. The drawdown level is high, which is a problem of all trend-following methods. Drawdown comes from two sources: suspected breakout trades that fail, and equity giveback from large winning positions.