Pages:
Author

Topic: [ANN] cointelligence.net - Create / test strategies for free online bot (Read 2804 times)

newbie
Activity: 21
Merit: 0
I just deployed a major update.... Please let me know if you run into any issues Wink

Update 5
  • Fixed rounding error that caused market API requests to fail
  • We are now combining and minifying the JavaScript files.
  • We are now combining and minifying the CSS files
  • We are now using a better modular implementation pattern that will allow us to soon run our code both in the browser and the server.
newbie
Activity: 21
Merit: 0
Thanks for reply. Hm.. i cant find AROON/MFI strategy on this page http://www.cointelligence.net/#/strategies. Is it possible to view it logic?

Hi, I put it on the strategies list now Smiley Maybe you can check it out and tune it a bit Wink
full member
Activity: 224
Merit: 100
Thanks for reply. Hm.. i cant find AROON/MFI strategy on this page http://www.cointelligence.net/#/strategies. Is it possible to view it logic?
newbie
Activity: 21
Merit: 0
2csupnig what time interval do you mean saying "aroon strategy does not work very well on short periods"?

Cause there is example, where aroon works well on 1h interval - https://cryptotrader.org/backtests/LucuYKZW59fvjZqpC

I was talking about a 1-5 minute interval. But I guess that I have to partly take that back, because yesterday a 5 minute interval on the ARRON/MFI combination was working very well for me.
full member
Activity: 224
Merit: 100
2csupnig what time interval do you mean saying "aroon strategy does not work very well on short periods"?

Cause there is example, where aroon works well on 1h interval - https://cryptotrader.org/backtests/LucuYKZW59fvjZqpC
newbie
Activity: 21
Merit: 0
Does anyone know a proper way to package, combine and minify javascript with node.js so that it can easily be done during a deployment process? I would love to implement this, so that you don't have to clear the caches/hard refresh on every update.

I would be very thankful for suggestions Wink
newbie
Activity: 21
Merit: 0
Update 4
  • Fixed the volume fetchend from btce. Up to now we always had the volume of the last 24 hours. Now we have the correct volume of the period.
  • We are now using a proper online javascript editor (http://ace.c9.io/) for creating/editing strategies. It should make things easier now.
  • Implemented MFI (it is now available in the MathUtils)
  • Added a test strategy for AROON and MFI (it is not tuned yet... maybe you want to help?)
member
Activity: 73
Merit: 10
Wow, yes that was a fantastic explanation. Exactly what I was looking for - I can take a look at what the aroon outputs are (now that I know how the method works) and try to better range the entry and exit conditions. Thanks for your reply - I'll see what I can do with this info and post back if I have any success.
newbie
Activity: 21
Merit: 0
So I started toying with strategies again and while I wasn't able to actually turn a profit using the preset Aroon strategy, I did see some positive results when I applied it to intervals of a few days.
It  seemed to do a very good job of detecting entry and exit points (in backtesting); I think that the only reason it wasn't profitable was because of the binary trade execution (i.e. sell 100% once signal is determined and vice versa). I'd like to use the code as a starting point and try to add a basic algorithm to make incremental buys and sells rather than all or nothing. I'm pretty sure that by just adding a set of a few simplistic rules you could turn the algo profitable.

The one thing that I'm a little confused on is the MathUtil.Aroon method. First, in the code I see three parameters: high, low, 22 - what is the 22 for? Also I'm not sure exactly how the overarching code uses the call to this method to make trades. From the code, it seems like you're just getting the aroon.up and aroon.down, finding the difference and then calling a buy or sell method depending on whether the result is positive or negative.

I have to imagine that's pretty close to what's happening, but I know I'm missing something. For instance, if that was the case then a trade would be initiated at every interval except for the case where aroon.up = aroon.down, which I would think would be fairly uncommon. But the strategy is obviously not making trades at every interval so I know that there has to be more to it than that. Maybe you could help point me in the right direction? Thanks!

Hi Smiley

You're correct, that the aroon strategy does not work very well on short periods. I implemented it as a first draft to test my aroon code in the MathUtil class. I think if you combine it with a volume based oscilator like the MFI (that I will hopefully implement tomorrow - depending on how my time schedule and the work on some other features work out), it could work very well!

To answer your questions:
The parameters for AROON are - high (an array of the high-elements of the last candles), low (an array of the low-elements of the last candles, 22 is the period (on how many candles the AROON should look at).
You are absolutely correct, that in this implementation, the aroon strategy would be making a trade at every interval it is even calling the buy/sell methods on the portfolio, but the portfolio will not execute the buy/sell order if you don't have the needed funds (e.g. if you already bought LTC and don't have any USD left). So it will always execute the orders on the first change in trends. If you want to read up more on this technical indicator, you can do that here: http://stockcharts.com/school/doku.php?id=chart_school:technical_indicators:aroon.

I hope that I could give you the right pointers, so that you can continue on your investigation Wink
member
Activity: 73
Merit: 10
So I started toying with strategies again and while I wasn't able to actually turn a profit using the preset Aroon strategy, I did see some positive results when I applied it to intervals of a few days.
It  seemed to do a very good job of detecting entry and exit points (in backtesting); I think that the only reason it wasn't profitable was because of the binary trade execution (i.e. sell 100% once signal is determined and vice versa). I'd like to use the code as a starting point and try to add a basic algorithm to make incremental buys and sells rather than all or nothing. I'm pretty sure that by just adding a set of a few simplistic rules you could turn the algo profitable.

The one thing that I'm a little confused on is the MathUtil.Aroon method. First, in the code I see three parameters: high, low, 22 - what is the 22 for? Also I'm not sure exactly how the overarching code uses the call to this method to make trades. From the code, it seems like you're just getting the aroon.up and aroon.down, finding the difference and then calling a buy or sell method depending on whether the result is positive or negative.

I have to imagine that's pretty close to what's happening, but I know I'm missing something. For instance, if that was the case then a trade would be initiated at every interval except for the case where aroon.up = aroon.down, which I would think would be fairly uncommon. But the strategy is obviously not making trades at every interval so I know that there has to be more to it than that. Maybe you could help point me in the right direction? Thanks!
newbie
Activity: 21
Merit: 0
We had some problems with the btc-e api today. This resulted in updates not being properly recorded and pushed to the browsers.

The requests seem to get through now and I hope the issues are resolved. I apologize for any inconvenience Wink

newbie
Activity: 21
Merit: 0
Update 3
  • Changed the default interval to 5 minutes
  • Changed the order of the default strategies, so that the EMA Rebalance Portfolio (works best for me) is on top
  • Implemented AROON and RSI (they are now available in the MathUtils)
  • Implemented an ichimoku class in the custom combined strategy
  • Fixed some chart bugs in Firefox
newbie
Activity: 21
Merit: 0
Hi,

im pretty new in the whole strategy business, however i allready traded a little bit on btc-e.com.
does your side make automatical trades via btc-e.com or how does it excatly works, because i dont get it lol
sorry for beeing a noob on this

Hi there,

yes, if you provide it with your API secret and your API key from BTC-e, it will be able to perform automatic trades for you. Just be sure to select and test a proper strategy before! Use the back test and the live test to do that.

The API secret will never be sent to the server (so your account is save), it will just be used to perform client side signing of the requests, that will be sent to BTC-e.

If you're into coding, you can even create your own strategies in JavaScript.

If you have any further questions, don't hesitate to ask!
newbie
Activity: 21
Merit: 0
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 Smiley

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
newbie
Activity: 1
Merit: 0
Hi,

im pretty new in the whole strategy business, however i allready traded a little bit on btc-e.com.
does your side make automatical trades via btc-e.com or how does it excatly works, because i dont get it lol
sorry for beeing a noob on this
newbie
Activity: 10
Merit: 0
I would be curious to know how your work on a neural network pans out. Keep us posted if you work on it past theory!
member
Activity: 73
Merit: 10
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 Smiley

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.
newbie
Activity: 21
Merit: 0
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 Smiley
member
Activity: 73
Merit: 10
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!!!
newbie
Activity: 21
Merit: 0
What does the ANN stand for? Artificial Neural Network?

That is just the type of post => Announcement.

Up to now I have not coupled any neural network to a strategy. Would be interesting though!
Pages:
Jump to: