Quantitative Approach to Backtesting

Here is the live statement from the same system but with order entry determined by a random number generator (coin toss) ... you can see it trades allot more than the previous system but has a lower profit factor... basically saying the the NN is providing some extra benifit by filtering the trades entered.

Well I don't want to comment specifically on your strategy in that I don't totally understand what your doing but I will say that your recent performance, while commendable, is relatively meaningless.

A basic Neural Net works by fitting a curve to the training set data. In a back-propagation model it takes the inputs and provides weighting based on optimizing, literally fitting a curve, to the training set. Think of an equation of the form f(x) = a*x^2+... You immediately introduce a huge bias based on your input data - basically you can feed it anything and it will spit out some answer! The next bias is that the data is only trained on a sample of the whole population. Since time series of financial securities are non-stationary, it is highly likely that any future data that matches the trained curve is due to luck.

Walk-Forward testing MAY work as it re-optimizes at set intervals and the hope is that the current dynamics of the time-series is taught to the NN and the TS remains stationary long enough to eek out some profits.

"The whole point being that speculating on a single pair is much less efficient and more risky than taking calculated positions on the whole market..."

There is your answer regarding correlation! So ask, how can knowing the correlation help me reduce the risk of my portfolio and make it more efficient?

while I appreciate your responses I find most of your comments to be pretty useless. And your use of an equation on your answer HUH... what does that have to do with anything, oh wait here you go a(i)=r*j^3.1312/g :nerd: why dont you try explaining something for once.

And yes thats why we use NN, to get the best solution for our trading programs based on a given set of data. My NN are set to define the trading parameters(solutions) for the next 30 min of trading based on 1. the past weeks worth of data. 2. an analysis of the fundamentals (IMF data correlation to long term pricing) 3. the expected volitility for the next 30 min based on the volitility of that same time period for the previous 30 days.

And I guess they do walk forward. optimized every min for the next 30.

Still the question I am posing is how do you manage a basket of trades together ? I am trying to get away from managing each indivigul trade and manage the entire portfolio as a single position.

I write vaguely to allow you to think more about your problem. You've answered most of your own questions by yourself, including how to manage a basket, aka PORTFOLIO more efficiently and with less risk (HINT: CORRELATION). If you want me to spoon feed it to you can hire me as a consultant.

I think it's a standard result that if you add a large number of strategies to you portfolio whose returns are uncorrelated then your equity PNL will smoothen out.