[B]Background[/B]
Before we jump into impacts, I want to cover the model just a little bit and how it came to be. I’ve been working with low level linear genetic programming for about 15 years now. At one time you could evolve trading models in a GP application and trade them with no tweaking. But like most approaches, implementation leads to nullification and they eventually become less efficient and more volatile. So adjustments have to be made.
I knew that going in, so I was able to get more life out of my original model than most but by 2007 it had rolled over and died a “volatile” death. I spent several months trying to build a new model, and had some success but I also knew that the difficulty I had in doing it had risen exponentially and eventually this approach would not produce results.
The initial model created in 2003, required over 10,000,000 models to be sampled before finding a good fit and it lasted for five years. In 2007 the system had to sample over 100,000,000 models and it became too volatile a year and a half later to trade. That’s a story for a different day though.
Knowing the 2007 trade model was going to be junk in short order I began to think about “efficiency” a lot more and how to measure it. I was also trying to get back into hedge-fund allocations and wanted to be able to show “capacity” was built into the approach. It’s important if you’re trading for institutional clients that you can prove you can handle really large allocations and have the ability to move monies in and out without suffering huge Market and Timing impacts. Despite a pretty large grid of computers I was never really able to get a good model going. I needed more computational power and more time.
[B]Enter 2012[/B]
In 2009 I took a sabbatical from all areas of retail forex so I could get back to research and modeling. By the early portion of 2012 I was ready to dig back in. I had just bought a pallet of new Dell T7500 Dual 6 core CPU workstations and had promised my first born to the electric company (these things consume a lot of power when they are at full tilt and I run them about 95% when doing a regression/classification.)
[B]Algo Details[/B]
The algorithm that we’ll use for our sample was created in 2012 using several years of EUR/USD 15m data. The two largest portions of the data were used for training and validation and the balance was left for out-of-sample testing. The out-of-sample data runs from June of 2011 to January of 2012. We’ll refer to it now simply as “2012”.
The algorithm was applied to each 15 minute period across the sample data and a forecast of where the market would be in 24 hours was recorded. A trade was assumed at the time with a 15 minute delay to duplicate real world market impact as was a 3 pip spread to imply transaction impacts and slippage.
Standard lots are assumed with a value of $10 per pip. Single trade leverage at the time of the trade is 1:1 and no overlapping trades are allowed - i.e. a trade is taken at 8:00 AM in the morning and held for @ 24 hours where (for this test) it is closed and a new trade is opened, even if it is in the same direction of the previous trade.
[B]The Apology[/B]
If all of that bored you to tears I apologize, just wanted to give you some background on what the model was based on and how it was measured. I’ll try to be brief from here on out 
- JP
