You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
NNs for optimizing 4 variables
- Thread starter Runningbear
- Start date
Quote from Runningbear:
Yeah that's the problem. I know the relative extremes of each of the four variables in isolation, but not the optimal values in combination for best trade performance. It's pretty easy when dealing with two variable to use set logic, but once you add the third and fourth variable, it gets tricky. For example, three variables at 80% of their extreme, may give better trade signals than two variables at 100% of their extreme. This is what I need in a system to do - identify and weight the value of each variable in combination. Such a system may reveal that one of the four variables doesn't add anything at all, and allow it to be eliminated. In other words, I need something that can find hidden patterns in the darkness.
Runningbear
Do you really only have 4 variables, or are there more? For 4, least squares regression might work. I believe you can even do this with Excel Solver. This does not include the interaction terms. And, if the variables/interactions are not linear, this won't be accurate. However it would get you started and give you a rough idea of what's going on.
The analysis could potentially be broken up into several problems. (1) How many variables should be considered. (2) Should interactions be considered, or can the inputs be assumed to be independent. (3) Optimization/probability/likelihood calculation.
In line with what intradaybill is saying, you could use multivariate probability distributions, and not use neural networks.
One way to determine if interactions are present:
signal = w1*constant + w2*A + w3*B + w4*C + w5*D + w6*AB + w7*AC + w8*AD + w9*BC + w10*BD + w11*CD + w12*ABC + w13*ABD + w14*ACD + w15*BCD + w16*ABCD
It sounds to me you need to look into a Genetic Algorithm (GA's) and not NN's.
If you can come up with an objective/fitness function with your variables, the variables can be solved very easily.
Another method and my personal favorite method is the use of Particle Swarms (PSO's). They are easy to implement and their operation when attemting to converge on a solution can provide insight to your problem.
If you can come up with an objective/fitness function with your variables, the variables can be solved very easily.
Another method and my personal favorite method is the use of Particle Swarms (PSO's). They are easy to implement and their operation when attemting to converge on a solution can provide insight to your problem.
Quote from gtor514:
It sounds to me you need to look into a Genetic Algorithm (GA's) and not NN's.
If you can come up with an objective/fitness function with your variables, the variables can be solved very easily.
Another method and my personal favorite method is the use of Particle Swarms (PSO's). They are easy to implement and their operation when attemting to converge on a solution can provide insight to your problem.
Will have to look into particle swarm optimization!
How would the GA be implemented? Split each variable into pieces? In this case, only 4 variables, a brute force approach might be fine. It would take more time than a genetic algorithm, however it would still run in seconds, unless there is tons and tons of data to go through. How much data are we talking about here?
Quote from gtor514:
It sounds to me you need to look into a Genetic Algorithm (GA's) and not NN's.
If you can come up with an objective/fitness function with your variables, the variables can be solved very easily.
Another method and my personal favorite method is the use of Particle Swarms (PSO's). They are easy to implement and their operation when attemting to converge on a solution can provide insight to your problem.
gtor514.
I'm not really familiar with particle swarms. Do you have any references that can help bring me up to speed?
Runningbear
Quote from Stoxtrader:
How would the GA be implemented? Split each variable into pieces? In this case, only 4 variables, a brute force approach might be fine. It would take more time than a genetic algorithm, however it would still run in seconds, unless there is tons and tons of data to go through. How much data are we talking about here?
GA's and PSO's are very fast and efficient methods of optimization however converging on a solution will depend on the size and complexity of the objective/fitness function and the number of variables. With four variables, it should be no problem.
I code in c++ and have an app that can optimize on a time series of 20,000 points with a dozen variables with 100 runs in just a few minutes and get close to some sort of convergence. It just depends on what your trying to solve, how big a population you want to begin with, your error for convergence, your solution space, etc, etc, etc
There are plenty of implementations in Java, python, matlab, on the web. Most trading platforms have some sort of implementation of the GA to optimize trading strategies.
Quote from Runningbear:
gtor514.
I'm not really familiar with particle swarms. Do you have any references that can help bring me up to speed?
Runningbear
A seach on google will give a ton of stuff. It all may seem really complicated but it's not it's actually quite simple. A PSO can programmed in less than 50 lines of code.
This is essentially how it works. You create a population of particles whose position represent a variable you want to solve for. In your case 4 variables => 4 populations. Each population is limited to a solutions space that represents the min and max or variable limits say (-1.0 to 1.0) and each particle is randomly initialized within this search space.
You then let the particles "swarm" or move a random direction and velocity. You take each particles position (which are your variable values) and pump them thru your objective function. You pick the particle from each population that provided the best solution and set the direction and velocity of each other particle in its direction. You repeat until you have a set of particles that best fits your solution.
Overkill methods. Just use backtesting in optimization mode with the 4 variables incremented.