Fast Brute-Force Optimization Software
- Thread starter sideburn
- Start date
I'd be interested to know what's considered "fast" these days. I use my own optimization software, capable of processing about 500 million bars per second (on a 6-core Intel i7 processor machine).
OK, tell us more...
$60K purchase price, nice. I looked up their manual to see the speed, here is what it says:
"TSL makes use of a high speed Genetic Programming Engine and will produce Trading Systems at rates faster than 16 Million System-DataPoints/second (on fast processors such as the Intel Core i7 990x)"
Good to know that my own software is about 30 times faster.
Business opportunity for you.
You don't need to pay $60k for a genetic programming engine. Here's a good software library that contains a lot more search algorithms than genetic programming:
http://cs.gmu.edu/~eclab/projects/ecj/
There has been a lot of people who have tried the shotgun approach of "I'll just throw every technical indicator programmed in TA-Lib into a genetic algorithm and if I just have it optimize weights and maybe add in some filtering, I'll have a mega-indicator that will predict the future." Given enough free variables, I can generate some non-linear combination that will curve-fit the past perfectly. But the application of interest is in predicting the future. My point is that it's so easy to curve-fit and so many people have tried that approach that I would be very surprised if it has any predictive power today. Now if the inputs are something not common (say something not found in NinjaTrader and would be a real PITA or practically impossible to code up in any popular charting software) and not based on a moving average of price, then yes, maybe metaheuristic optimization algorithms can help. But if common computing hardware can process 16 million steps per second, my bet is that the indicators are not much more than moving averages or some simple arithmetic derivative of price.
http://cs.gmu.edu/~eclab/projects/ecj/
There has been a lot of people who have tried the shotgun approach of "I'll just throw every technical indicator programmed in TA-Lib into a genetic algorithm and if I just have it optimize weights and maybe add in some filtering, I'll have a mega-indicator that will predict the future." Given enough free variables, I can generate some non-linear combination that will curve-fit the past perfectly. But the application of interest is in predicting the future. My point is that it's so easy to curve-fit and so many people have tried that approach that I would be very surprised if it has any predictive power today. Now if the inputs are something not common (say something not found in NinjaTrader and would be a real PITA or practically impossible to code up in any popular charting software) and not based on a moving average of price, then yes, maybe metaheuristic optimization algorithms can help. But if common computing hardware can process 16 million steps per second, my bet is that the indicators are not much more than moving averages or some simple arithmetic derivative of price.
I agree. You have to clearly define what problem you are trying to solve and then see what flavor of Machine learning, if any, can help you. For me, I want to automate as much as I can the process of finding and testing strategies. Machine learning techniques can definitely help with this.