(4.88 out of 5)Artificial Neural Nets (Part 2 – Bogie-NN)
(2 votes, average: 4.00 out of 5)
Loading ...In my previous article I introduced the topic of neural nets in trading. In this article I want to review a specific neural net: Bogie-NN as a way of getting more in depth around neural nets.
Bogie-NN
Bogie-NN was the second place winner in the 2007 automated trading championships. Between the 1st of October and the 21st of December 2007, Bogie-NN managed to trade $30,000 up to $55,033.10, representing a stellar return of 83% over 3 months.
Since placing 2nd in the competition, William Boatright, the creator of Bogie-NN, has made it available on his website for $49.50 per month. You can also download a demo version of Bogie here, but the demo version expires in September so be quick if you want to have a look at it. If you do decide to sign up and buy bogie-NN, CK on his blog and mailing list has a 20% off deal.
How does it work?
Bogie-NN is a Metatrader 4 expert advisor. Like all commercial expert advisors, bogie-NN is a grey box trading system, where you can control and optimize some of the parameters of the system, but the internals of the system are proprietary.
Out of the box, Bogie-NN, comes with installation and optimization instructions and an installer. The installer drops two key files into your metatrader installation: an expert advisor (the system which does the trades) and a custom indicator (used by the expert advisor to trigger trades).
The expert advisor implements a trend trading system and allows you to control a number of properties around trade management including:
- Risk (%) – as the name indicates this variable determines the percentage of the account to risk per trade. An alternative to Risk is the lots variable which allows the trader to set a fixed set of lots to trade instead of a percentage to risk
- Take Profit, Stop Loss and Trailing Stop let you manage when trades are exited
- Buy Trigger and Sell Trigger determine when you enter (see a discussion of the Bogie-NN indicator below for how this is used)
- You can set which days not to trade
- You can set wether to use mini-sized lots or not and you can set if you want to use pipettes
Unlike other commercial EAs, Bogie-NN does not come with a forum for support. All support is conducted through email.
The real black box part of bogie-NN seems to be the bogie-NN indicator. The indicator is the part of the system which has been trained by a neural net and is used to identify suitable entry points in a trend. The details of how it was trained remains a mystery. In an audio interview with the William by CK and some of his trading buddies, all we know is Bogie-NN is a 3 layer neural net. William did not disclose more than that (however keep reading – and we will try and get to the bottom of how the indicator actually works later in this article).
The indicator seems to be an oscillator, like a Stochastic %K or an RSI, which ranges between 0 and 100. Like all oscillators, I suspect 100 represents the possible over-bought top of the market and 0 represents the possible over-sold bottom. The indicator is used to determine entries and exits. If the indicator rises above the buy trigger level Bogie-NN will do a long entry. If the indicator sinks below the sell trigger level Bogie-NN will do a short entry.
The Bogie-NN indicator looks like it is a stochastic that is dampened somehow. For example, if you overlay it with a stochastic %K (12/5/5 period), you will see the oscillations are similar, but the stochastic swings more widely.
Bogie-NN comes in at least 2 different flavors:
- Bogie-ATC2007-1 (the system used in the automated trading championship); and
- Bogie-NN-v8 (an enhancement of the trading championship Bogie)
Performance
As mentioned above, in the 2007 trading championship Bogie-NN produced an 82% return over a 3 month period (i.e. around 22% per month compounding). CK on his blog has forward tested Bogie-NN-V8 and managed to replicate a similar result. He managed to produce an 87% return over about a 4 month period between late-May to mid-August (i.e. around 17% per month compounding) with a maximum drawdown of around 19 – 20%.
It is not clear what settings William used in the 2007 ATC he used to achieve his outcome, but if you run a back test over the same period of the competition using his default settings, you need to commit to about 10% risk to achieve the same outcome (5% is what William normally recommends).
If you back test the system, it seems that the system is sensitive to wether or not the market is trending. Even if you use the newer Bogie-NN-v8 and trade it between March and April of this year it would return a loss as the market was ranging in that period. So it seems the new trending filter is not necessarily every thing it could be, but at least it stops you from draining your account too heavily.
In my case I have been forward testing Bogie-NN-v8 for the last month and Bogie has recorded a loss. During this period, the Euro has declined substantially. It seems that Bogie seems to have an edge in a bullish market, but not in a bearish market. If you back test Bogie during bearish periods it doesn’t seem to perform well at all. Why this is the case is not clear to me. I guess more will be revealed when we open the black box.
Breaking Open the Black Box
- The first thing to remember is that while neural nets cannot think, but they can be trained to do pattern recognition really well. Furthermore, you cannot train a neural net to recognize arbitrary patterns. You need to train it to do a specific task. In the tutorial the author trained the network on a strong bull trend during 2003/2004 on the EUR/USD. This means the bogie-NN is highly efficient at trading bull trends on the EUR/USD, but will suck in other conditions or on other pairs; and
- The second thing to note is around the kinds of inputs you provide it. Many neophytes will take a neural net and feed it every indicator under the sun and pray that the net will find patterns. This is just wishful thinking. The selection of indicators you choose to train a neural net is just as important as the selection of indicators that you will use in every day trading. You need to pick appropriate indicators. For example, the stochastic is very susceptible to whipsaws. If the market is caught in an untradeable tight range the stochastic will swing around just as widely as it will in a tradeable wide range and is not very helpful for mechanical systems. Therefore to accommodate this author presents his own “Normalize on Condition (NOC)” indicator. The indicator is similar to a stochastic, but when the historical high / low is quite tight, a reference range is substituted. This means the NOC behaves like a dampened stochastic which is less susceptible to whipsaws. This in turn means large amounts of noise is filtered out of the input data used to train the neural network.