Getting Too Attached to Your Model

I have recently decided to put a model aside after spending many hours trying to build one for a magazine article.  I wanted to create a ChaosHunter pairs trading model that would show good results in an out of sample period at the end of 2013.    I had built dozens of models of stocks compared to the S&P 500 index.  I had oil, copper, and energy futures models.  I had FOREX models for most of the major currency pairs.  What I finally realized was that I was trying to fit a square peg into a round hole for this current market.  With the overriding uptrend that was 2013, in general it was more profitable to just buy and hold both securities in the pair.

This led me to realization that I should have known better.  One of the main ideas we tried to present at our training seminars was to not get too attached to modeling your favorite stock.  If you build a neural network model, for example, it is better to model an instrument that has up and down patterns for the net to learn rather than simply a huge trend that swamps the model.  The same problem can occur when you get too attached to the type of model you want to use and just keep switching instruments.

The great thing about having excellent model building tools like NeuroShell Trader and ChaosHunter is that you can easily test a new concept.   That way you’ll always have the right model on hand for the current market conditions.

Check Out the New ChaosHunter.com

The ChaosHunter web site has a new look for 2014.  Check it out at www.chaoshunter.com.  Next up, NeuroShell.com.

Here’s a simple system that uses price action off of volatility bands in order to generate entries based on false breakouts. A lot of people like to incorporate different types of volatility bands into their trading so I thought it would be interesting to show the results of the same system applied with different types of bands. Two popular types of volatility bands are Bollinger Bands and Keltner Channels, so we’ll test those. In NeuroShell, the Bollinger Bands are included in the Time Series category; the Keltner Channels are included in the Advanced Indicator Set 1. We’ll also create and test variations of Bollinger Bands and Keltner Channels using indicators included in the BMJ Fractal Analysis Indicator Set. This will hopefully illustrate how our indicators can sometimes be used in order to improve already profitable systems.

In the chart with the BMJ indicators, I created bands similar to Bollinger Bands by adding and subtracting a multiple of the standard deviation to the BMJ FFilter 30 of the close.

I did a similar process in order to create bands that resemble Keltner Channels. In this case I added and subtracted a multiple of the BMJ FFilter ATR 30 to the BMJ FFilter 30 of the close.

For long and short entries, I want to find moments where the price briefly breaks outside the bands but still manages to close within them. These will represent a false breakout of the band so I’ll enter in the direction away from the band. This type of entry rule is really easy to accomplish using the Boolean And2 indicator included in NeuroShell with two simple rules as inputs. For longs I want to look for bars where the low crosses below the lower band, but the bar still manages to close above it. Here’s a simplified version of how it looks within the Indicator Wizard:

And2

      Operand #1 = CrossBelow(Low, Lower Band)

      Operand #2 = A>B (Close, Lower Band)

Shorts are the opposite, so I’m looking for bars where the high crosses above the upper band, but the close still manages to close below it. This time the simplified version of the rule looks like this:

And2

 Operand #1: CrossAbove(High, Upper Band)

 Operand #2: A<B (Close, Upper Band)

I chose to trade five of the more popular mini index futures contracts on a six hour timeframe using data provided by IQFeed. The contracts traded were the e-mini S&P 500 (ES), e-mini S&P Midcap 400 (EMD), e-mini Dow Jones (YM), e-mini Nasdaq 100 (NQ), and mini Russell 2000 (TFS). For parameter optimization, I fixed the standard deviation and ATR multiples at 2 and linked all the remaining parameters together with a range of 20-40. Match Chart Page was selected for the optimization period and three years were added for out-of-sample trading. Maximize Return On Account * Log Equity Curve Correlation was selected for the optimization objective and position size was fixed at one contract. Margin was set to $4,510 (typical margin for the ES), commissions of $2.50 per contract per side were added and the point values for each contract were entered. I also generally like to include slippage but this becomes difficult when trading multiple futures contracts because they often have different tick increments. I overcame this by simply adding an extra two ticks of commissions per contract per side in addition to the $2.50.

The results are very promising, with all four strategies making money out-of-sample on average. Below are the results for each strategy.

Chart 1:  Results using basic Keltner Bands and Bollinger Bands

Chart 2:  Results using the BMJ Fractal Filter Variations of Keltner Channels and Bollinger Bands

A few things stick out to me upon closer inspection of the results. First, the fact that this simple entry technique is able to generate strong returns regardless of the type of volatility band being used shows me that it has promise. In fact, the lowest out-of-sample average for annual return on account was 84.2% with standard Keltner Channels. Second, the BMJ Fractal Filter variations of Bollinger Bands and Keltner Channels were slightly more consistent in their results. Both types of Fractal Filter bands made money on all five securities both in and out-of-sample. Although the standard Bollinger Bands outperformed their FFilter counterparts on average, they lost a small amount of money out-of-sample on the e-mini Dow Jones. The Keltner Channels lost money on two of the five securities traded and only made a small amount on another, meaning almost all of their profits came from only two securities. The returns are still strong on average, but I’d like to see more consistency across the instruments traded.

The standout of the group appears to be the BMJ Fractal Filtered variation of Keltner channels (Strategy #3 in Chart 2). The system using these was profitable on all five securities both in and out-of-sample. It made the most profits, generated the highest average return on account (223.6% annually out-of-sample), and achieved a 71.9% average percent profitable trades out-of-sample. When using these bands, the lowest annual return of the five securities was 41.5% during the out-of-sample period and all the others achieved triple-digit returns.

It’s important to note that these strategies are true reversal strategies. They do not include stops or exits and they do result in reasonably sizeable drawdowns. Some traders might find this a perfectly acceptable cost of achieving the returns illustrated. But others might prefer a lower risk profile at the expense of a slightly lower return. Our next article will illustrate how we might add stops and/or exits to an already successful trading system in order provide the opportunity for a reasonable return while controlling risk.

I hope some of you found this article to be helpful. If you have any questions regarding this strategy or anything else related to our products, please don’t hesitate to contact either myself or my father, Dennis Voigt.

Graham Voigt

BMJ Software

Note:  BMJ Software no longer offers add-ons for NeuroShell Trader.

Click here to download both charts described in this article.  Chart 1 requires the Keltner Channels Indicator from the WSG Advanced Indicator Set 1.  Chart 2 requires the BMJ Fractal Analysis Indicator Set