Automated Trading – Choosing Models Part 2
The first round of testing of testing of the FXCM Forex System Selector is now finished, with a end profit of $1200 after 5 weeks of trading, on a $100,000 investment. The trial account has been closed, so it is time to move to a new trial.
Of course, one month is too short a time period to measure the performance of these models – it can only give an idea of the risk profile of the system. But many investors will want to try out this new tool, and a real time test does give some idea of what can be expected in the short term.
In the last article in this series, I demonstrated how some simple filtering can be used to prepare a short list of models for further evaluation. I mainly looked for models with a good long term track record, and ended up with 9 models that looked promising. This article is about the processes I have used to further fine tune this selection to build a final trading system.
Firstly, here is a list of the models on my initial shortlist:
- Currency-Specialist (GBP/JPY)
- AUDCAD-Mover (AUD/CAD)
- FxFons (EUR/JPY)
- Tecnofinanzas (EUR/USD)
- PipboxerV2 (USD/JPY)
- FXSignaler (EUR/JPY)
- Quants-Carnival (EUR/JPY)
- Quants-VIP (EUR/JPY)
- EspritFX (GBP/USD)
The first step is to evaluate the equity curve that results from these models. Remember, the equity curve of the system is simply the sum of the equity curves of the individual models. Here is a chart of the historical equity curve (actual results) from end of June 2007 to present:
This equity curve is generated by the Forex System Selector software. You can see that the starting balance is $10,000 and the ending balance is around $26,000. This is from trading around 10 mini contracts for each model.
You can also see that the equity curve is upsloping (good) but there are a couple of big jumps, and some flat periods that extend for some months (bad). Our goal in selecting models is to have the curve sloping upwards (good) for almost all the time, and with no significant downturns. We are seeking a system that steadily produces profits, with small losses, so a very conservative, low stress system.
Alternatively, you could choose a high risk, high return model. In this case, you would be less concerned about the smoothness of the equity curve, and more concerned about the ending balance, and the slope of the curve.
To change the equity curve, look at the equity curves of the individual models. I looked through and found some that seemed quite bumpy, with big up and down movements. I progressively eliminated these, and went from 9 models down to 4. The models that are left are:
- AUDCAD-Mover (AUD/CAD)
- Tecnofinanzas (EUR/USD)
- Quants-VIP (EUR/JPY)
- EspritFX (GBP/USD)
Now the equity curve looks like this:
You can see that apart from the initial downturn, and a little dip about half way along, it is a fairly smooth equity curve. The overall ending equity is about $16,000, lower than the initial $26,000 but with much lower risk.
After eliminating more than half of the models, you would think that the curve might be bumpy due to lack of diversification across trading styles, but this is not true. The result actually diversifies across currencies more than before, and the smoother equity curve shows that you don’t need more models to end up with more predictable results.
The approach of reviewing individual equity curves, understanding their contribution to the whole and adding and removing models is time consuming and difficult. I would like to be able to access the underlying trading data so that I can extract my own custom statistics (such as Sharpe ratio, which measures risk adjusted return).
Ideally, I’d set a genetic algorithm on the job, to automatically develop an equity curve that is as smooth as possible by selecting the optimal models. A genetic algorithm would work by randomly selecting models, then retaining those that contributed the most towards the combination of overall equity curve smoothness and ending equity. This is a search function, and genetic algorithms are very good at this task. This would not eliminate the risk of models not performing as well in real life as initially expected, but would aid the selection process.
As mentioned, a trading system probably needs more than a one month trial to determine how profitable it will be, and the equity curve even in simulation shows this. This article, and the previous article demonstrate how you can choose a combination of models by checking the individual equity curves after building a focussed list and fine tune your system for your own risk tolerance.
Tags: Automated Trading, forex system selector, FXCM, model selection