Quote from: thaReal on January 19, 2014, 06:38:58 PM

Quote from: csupnig on January 19, 2014, 03:43:10 PM

Quote from: thaReal on January 18, 2014, 04:11:01 AM

Wow, I am so glad I found this - some amazing work csupnig

This is the first open source trading & strategy testing I've found that I believe could seriously be more useful than what's currently possible with MT4. I've only been playing with it for a few minutes, but I can tell you definitely have some impressive coding skills - for having just launched the site has a very clean, user friendly interface and seems like it has a ton of possibility for customization. I'll definitely be following this discussion and commenting further once I get some time in playing with the system.

As ironically was mentioned in the last post, attempting to utilize a neural network strategy would be interesting. I recently was involved in a competition where neural networks were (attempted) to be used in order to predict the price of options over a period of two hours using a data set with anonymous options and market inputs. As it turned out, at the end of the competition no one was able to actually develop a predictive model that provided any significant improvement vs. simple traditional models. In fact, while the goal of the competition was to utilize the market data, most of the models that performed well simply ignored the market data and used only price and time to train the model.

However, even though it was a similar exercise, just because a neural network model wasn't feasible in that situation doesn't mean it's not in this market and I think this is a great platform to try and test that. I look forward to working with you and your site - I'm excited to see what the possibilities are!!!

Thanks for your feedback and I'm happy that you found it useful! If you come across any questions regarding the functionality or need some assistance in getting your strategies to work... I'm here to help!

The other day I talked to a guy on the bitcoin meetup in vienna, and he was working on a neural network model for the bitcoin market. When I talked to him, he was at a very early stage of his project... I'm not sure if applying neural networks to market data would actually work, because it mostly does not follow a pattern at all. But I would love someone to prove me wrong on this

Thanks for the response, I'll send you a message once I've actually spent the time working with the code. With the Nerual Network model - I don't think that there's much significance in using the traditional approach of training the model with common inputs like EMA, Volume, Price, etc. (well to a certain extent, these things obviously play a role and need to be accounted for somewhere in the model). My thought, which I admit I haven't spent the time to experiment with yet, is that maybe its possible to get better results by training a model based on the parameters of the coin and the market(s) it's being traded in.

More specifically, I think that to a certain extent a lot of the coins being released now have similar features to the original coins (since most of them were basically just copies anyway with a few modified parameters, i.e. any coingen coin) so theoretically the market on a macro level should to some extent mirror the markets of those previous coins. If you assume this to be true you could try to develop a NN based off of these training parameters to get a general model and then add inputs like EMA, Price, & Volume to hopefully get a reasonably predictive time-series model.

My reasoning is that the fundamental problem with neural networks are that they're very dependent on the dataset used to train the model, so they end up being highly dependent on the time interval you use to train your model, which means that while you can sometimes get it to fit to a specific case, the model is going to fall apart when anything rapidly changes, and with this market that's always. By doing the way I'm proposing, (theoretically) you're generalized model is more robust as it's not time-dependent and then you would likely have to derive several specific-models based on time-dependent market data.

Whew, well that was a mouthful - I know it's a lot of theory right now, but hopefully in the next week or two I can take a rough pass at trying to build a model and see if there's any promise here. Before I get started with that though, I'd like to spend more time playing with some of the existing strategies you developed. Admittedly, they do not perform well at the moment, but I understand how much work must have went in to just getting them working (along with everything else) so hopefully I can work on tweaking some of them until they at least start to perform more consistently and then work from there.

I'm with you on the thought that a neural network model would be highly dependent on the time and at least some recurring pattern. Maybe it would be easier to get a model going that works good for a longer period.

You're right, that most of the strategies that I developed do not perform very good. I simply created them to test the possibilities and see if my implementation of the technical indicators works. I think they would perform much better, if someone could take the time to tune them for a specific market and a period e.g. btc-e ltc/usd with a period of 5m. In the next days I'm trying to implement more technical indicators, so that you guys have a lot to play with Wink

When you test out strategies, be sure to keep us posted on your findings Wink