40% returns seems like a lot. Your quote, "FX.Machine builds profitable trading strategies by learning from historical data", is true for most of the strategies. What makes your strategy different? I would think that many people have tried both statistical and machine learning methods. I do not hear about any 40% returns from them. Your claim of 40% returns makes me doubt your testing strategy. Your forward test shows only 7.9% increase till now. There are a whole set of backward tests missing and a lot of information missing to consider your strategy seriously. For example, how is the risk being measured? Is it drawdown, Sharpe ratio etc?
Thanks for your questions! Feedback like this really helps us.
To your first point about our development strategy: we used genetic programming, and tried to apply novel strategies for which there wasn't much of a precedent in financial applications.
To your second point, re: our testing strategy:
I will concede that we only have 2 months of live trading data, but so far it corresponds well to the behavior we were seeing in our forward tests (what we call the data that the system was not trained on).
We trained our system on 10 years of data (2001 - 2011) with two years of forward tests (2012 and 2013) for evaluating its effectiveness (and to verify that our algorithms have not been overtrained to fit the testing data set). Our forward testing involved taking the same data we would have received from our financial institution and feeding it through the algorithm to see what trades it makes. Having proven our algorithm worked with forward tests, we hooked it up to live data to get real results. So far, so good!
To you third point--risk--we need to explain how our algorithm trades.
We've trained our system to recognize patterns such that it will open a trade with a stop-loss and a take-profit. Because the take-profit and stop-loss are preconfigured, our system doesn't need to monitor data feeds with a time resolution of “real time” to make decisions; the onus is on the brokerage to close winning and losing trades.
The lot size of the trade is calculated such that a losing trade (the price of the currency pair reaching the stop-loss) will incur a loss equal to the risk. So at a risk of 2%, a losing trade on a $1000 account will bring the account value down to $980. Winning trades earn double the risk, so a winning trade on an account of $1000 at 2% risk will increase the account value to $1040.
With all this in mind, we claim 40% annual profit because our live trading exhibits patterns similar to our forward tests, and while our forward tests exceed 40% annual profit, we're extrapolating our two months' worth of live testing performance out to the rest of 2014. There's a good chance that this number will change in the coming months. We will revise it up or down accordingly.
It's possible that our system will need to be used "in the wild" for more than just two months in order for people to take us seriously. If that's the case, I am happy to continue dogfooding FX Machine; so far, it’s been a profitable experience :)
Thanks again for your great questions. We’ll use questions like these to build an FAQ section on the site!
If your stops are limit orders, then they can get blown through without execution. If your stops are market orders then your limit trigger price can be quite distant from your fill price. You need to do some serious thinking about your risk management and the long tails associated with extraordinary market moves.
