elective-stereophonic
elective-stereophonic
Helix - Dapps | Test - C&P Repos  
Please login or register.

Login with username, password and session length
Advanced search  

News:

Latest Stable Nxt Client: Nxt 1.12.1 Upgrade before block 2870000 is mandatory!

Pages: 1 [2] 3  All

Author Topic: Helix - Dapps | Test - C&P Repos  (Read 10945 times)

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #20 on: November 18, 2014, 11:05:10 pm »

user could form a very powerfull financial grid (supercomputer), connected via blockchain to compute massiv data and store the best
AI back into the blockchain, envolving step by step via GA into an optimal computation grid.

You could reward computational nodes based on their submissions.

Last night I had a vision from BCNext...  This is what he told me...

"2Kool... tell them to build a trading bot controlled by a neural net on the blockchain... use computational nodes to train the neural net... proportionately distribute the trading bot's revenues to the computational nodes based on their submissions.
BUILD IT and THEY WILL COME!"



not bad 2Kool4Skewl, good attitude  ;D

yes, why not some visions or just an unsual approach. the streets also only cooking with water,
there is no special magic there (except much money, ok) but crypto has huge resources 
(human and tech) available to compete. there is no need to be shy.

just to mention:
neural nets are not good in predicting but they are very good (with long proven results)
in patter recognition. what boinc[1] is for folding will be FinHive for finance.

here is an very,very audacious example setup.
assuming some helix clients are running (no full nodes necessary, just a small exec).

1. 50% are running CNNs[2] ( deep learning, overfitting reduced due to RNN )
   used for pattern recognition. CNNs having the best MNIST[3] error rate so far (0.3%).
2. 25% are running BAMs/Hopefield[4] (bidirectional associative memory)
   used to store and recall RNN detected pattern and/or correlations.
3. 25% are running SOMs[5] (self-organizing map)
   used to run on complete uncorrelated RNN output just to ensure nothing is missed
4. and now a nice layer on top: each of those net's are dynamically bounded to an modified RNN[6] as ONE layer,
   forming a metalayer, enabling result based reorganisation and cascading for the whole layer set.
   the signaling (data about weights, connections and main topology) is stored and transferred via
   nxt blockchain, which is slower but provides simple and reliable distributed/simultaneous access
   from all sublayers. so, instead adding aditional layers into the RNN net it adds whole NNs, feeding
   them recursiv.

then, let the beast run on some terrabyte finance data, i am sitting here on more than several TB and
accumulating more every day, not only quotes and TA but also fundamental, economic and some other timeseries.
store the results e.g latest after 500.000 epochs or error rate reaches the threshold into the blockchain.
if meaningfull patterns or correlations are found use them and go hunting at the street. if not, adjust/reform the input params,
reorg the layers and run again.

as far as i know, nobody created that level of nested and selforganized net layers and if so not public.
so why not try? crypto have all resources to do so. nothing to loose here but if lucky much to win.

the main task is not to create such a nested layer topology but to prepare meaningfull input data.
it's simple, garbage in garbage out but i am sure, if this would be built, it will detect hidden pattern.
i would even take a serious bet on it.

and here, a little more crazy idea for you:
instead using different net's, all nodes are running a RNN forming a population where every RNN is viewed as
an individual. then, the RNNs are doing cross over mutation, driven by GA[7]. instead e.g. swapping byte-positions
(chromosoms) by individuals within a population, they are changing weights and full topology, or e.g. complete
layers on each RNN.
mutating CNN pairs are defined and selected by fitness ratio. since this is a evolutional approach the
outcome is somehow undefined. i guess you would like it, surprises are guaranteed. just let them compete
and fight for a rare digital resource which is needed to fuel them and extract the best RNNs after a while.
well, unless they don't have killed each other to fast.

finaly: the whole AI stuff in FinHive is just a research section. FinHive will make it's profit from
selling traditional but very, very intense and complete financial analytic, so FinHive is very grounded here.
however, i really would like to see this crazy stuff, just because i am curious for the possible outcome.


[1] http://en.wikipedia.org/wiki/Berkeley_Open_Infrastructure_for_Network_Computing
[2] http://en.wikipedia.org/wiki/Convolutional_neural_network
[3] http://yann.lecun.com/exdb/mnist/
[4] http://en.wikipedia.org/wiki/Bidirectional_associative_memory
[5] http://en.wikipedia.org/wiki/Self-organizing_map
[6] http://en.wikipedia.org/wiki/Recurrent_neural_network
[7] http://en.wikipedia.org/wiki/Genetic_algorithm

« Last Edit: November 18, 2014, 11:07:14 pm by nexern »
Logged

jl777

  • Hero Member
  • *****
  • Karma: +718/-123
  • Offline Offline
  • Posts: 6170
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #21 on: November 18, 2014, 11:20:20 pm »

user could form a very powerfull financial grid (supercomputer), connected via blockchain to compute massiv data and store the best
AI back into the blockchain, envolving step by step via GA into an optimal computation grid.

You could reward computational nodes based on their submissions.

Last night I had a vision from BCNext...  This is what he told me...

"2Kool... tell them to build a trading bot controlled by a neural net on the blockchain... use computational nodes to train the neural net... proportionately distribute the trading bot's revenues to the computational nodes based on their submissions.
BUILD IT and THEY WILL COME!"



not bad 2Kool4Skewl, good attitude  ;D

yes, why not some visions or just an unsual approach. the streets also only cooking with water,
there is no special magic there (except much money, ok) but crypto has huge resources 
(human and tech) available to compete. there is no need to be shy.

just to mention:
neural nets are not good in predicting but they are very good (with long proven results)
in patter recognition. what boinc[1] is for folding will be FinHive for finance.

here is an very,very audacious example setup.
assuming some helix clients are running (no full nodes necessary, just a small exec).

1. 50% are running CNNs[2] ( deep learning, overfitting reduced due to RNN )
   used for pattern recognition. CNNs having the best MNIST[3] error rate so far (0.3%).
2. 25% are running BAMs/Hopefield[4] (bidirectional associative memory)
   used to store and recall RNN detected pattern and/or correlations.
3. 25% are running SOMs[5] (self-organizing map)
   used to run on complete uncorrelated RNN output just to ensure nothing is missed
4. and now a nice layer on top: each of those net's are dynamically bounded to an modified RNN[6] as ONE layer,
   forming a metalayer, enabling result based reorganisation and cascading for the whole layer set.
   the signaling (data about weights, connections and main topology) is stored and transferred via
   nxt blockchain, which is slower but provides simple and reliable distributed/simultaneous access
   from all sublayers. so, instead adding aditional layers into the RNN net it adds whole NNs, feeding
   them recursiv.

then, let the beast run on some terrabyte finance data, i am sitting here on more than several TB and
accumulating more every day, not only quotes and TA but also fundamental, economic and some other timeseries.
store the results e.g latest after 500.000 epochs or error rate reaches the threshold into the blockchain.
if meaningfull patterns or correlations are found use them and go hunting at the street. if not, adjust/reform the input params,
reorg the layers and run again.

as far as i know, nobody created that level of nested and selforganized net layers and if so not public.
so why not try? crypto have all resources to do so. nothing to loose here but if lucky much to win.

the main task is not to create such a nested layer topology but to prepare meaningfull input data.
it's simple, garbage in garbage out but i am sure, if this would be built, it will detect hidden pattern.
i would even take a serious bet on it.

and here, a little more crazy idea for you:
instead using different net's, all nodes are running a RNN forming a population where every RNN is viewed as
an individual. then, the RNNs are doing cross over mutation, driven by GA[7]. instead e.g. swapping byte-positions
(chromosoms) by individuals within a population, they are changing weights and full topology, or e.g. complete
layers on each RNN.
mutating CNN pairs are defined and selected by fitness ratio. since this is a evolutional approach the
outcome is somehow undefined. i guess you would like it, surprises are guaranteed. just let them compete
and fight for a rare digital resource which is needed to fuel them and extract the best RNNs after a while.
well, unless they don't have killed each other to fast.

finaly: the whole AI stuff in FinHive is just a research section. FinHive will make it's profit from
selling traditional but very, very intense and complete financial analytic, so FinHive is very grounded here.
however, i really would like to see this crazy stuff, just because i am curious for the possible outcome.


[1] http://en.wikipedia.org/wiki/Berkeley_Open_Infrastructure_for_Network_Computing
[2] http://en.wikipedia.org/wiki/Convolutional_neural_network
[3] http://yann.lecun.com/exdb/mnist/
[4] http://en.wikipedia.org/wiki/Bidirectional_associative_memory
[5] http://en.wikipedia.org/wiki/Self-organizing_map
[6] http://en.wikipedia.org/wiki/Recurrent_neural_network
[7] http://en.wikipedia.org/wiki/Genetic_algorithm
Why not add SVM's to the mix?
I find they are better at NN for training with large number of features
Logged
There are over 1000 people in SuperNET slack! http://slackinvite.supernet.org/ automatically sends you an invite

I am just a simple C programmer

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #22 on: November 18, 2014, 11:30:53 pm »

jepp, you are right. SVMs generating good results in MNIST also and would be an additional option.
there are tons of different tops possible. the best would be to test a selection and define a possibe mix right after.
e.g. just a simple overfitted MLP could also be used instead a hopefield. the combinations are endless.
Logged

jl777

  • Hero Member
  • *****
  • Karma: +718/-123
  • Offline Offline
  • Posts: 6170
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #23 on: November 18, 2014, 11:40:39 pm »

jepp, you are right. SVMs generating good results in MNIST also and would be an additional option.
there are tons of different tops possible. the best would be to test a selection and define a possibe mix right after.
e.g. just a simple overfitted MLP could also be used instead a hopefield. the combinations are endless.
SVM using NN outputs as features
or NN using SVM outputs as its inputs
yes, all combinations can be tested and evolved using genetic algos, I like pareto pruning

nothing can stop SKYNET!!
Logged
There are over 1000 people in SuperNET slack! http://slackinvite.supernet.org/ automatically sends you an invite

I am just a simple C programmer

Tosch110

  • Hero Member
  • *****
  • Karma: +211/-18
  • Offline Offline
  • Posts: 2365
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #24 on: November 18, 2014, 11:54:29 pm »

Just discovered this thread and this is what perfectly describes my current impression:



Lol +1 ;D

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #25 on: November 18, 2014, 11:56:56 pm »

jepp, you are right. SVMs generating good results in MNIST also and would be an additional option.
there are tons of different tops possible. the best would be to test a selection and define a possibe mix right after.
e.g. just a simple overfitted MLP could also be used instead a hopefield. the combinations are endless.
SVM using NN outputs as features
or NN using SVM outputs as its inputs
yes, all combinations can be tested and evolved using genetic algos, I like pareto pruning

nothing can stop SKYNET!!

optimizing NN via GA is great if you are using mainly unstructured input.
the input data i am feeding the NNs later is very clean, defined and uncorrelated,
which should reduce unwanted output noise.

however, if we like to see skynet we need GA or better ES for sure ;-)
Logged

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #26 on: November 19, 2014, 12:04:33 am »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?

edit:
and one more, because iirc you have also wrote some indicators and TA related stuff.
i was looking for a good TA lib a couple of months ago and all i have found was TAlib
which is, beside written in C++ and i would prefer ansi C, very outdated (7-8 years?)
and more important not as complete as i need.

therefore i started to write my own lib. my indicator list is complete now (+150) and only
candlestick and formations are needed. i guess you know how demanding this work is
so do you have any C based resources worth to look at, except TAlib?

not asking for code, will do it by myself, just for resources in case i missed something.
it is strange, that there is such a shortage on good and fresh TA related code out there.
« Last Edit: November 19, 2014, 12:57:39 am by nexern »
Logged

jl777

  • Hero Member
  • *****
  • Karma: +718/-123
  • Offline Offline
  • Posts: 6170
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #27 on: November 19, 2014, 12:31:38 am »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?
I've used FANN, but dont like the lack of convergence, so I switched to SVM
super optimized the fastest SVM lib out there
Logged
There are over 1000 people in SuperNET slack! http://slackinvite.supernet.org/ automatically sends you an invite

I am just a simple C programmer

gs02xzz

  • Hero Member
  • *****
  • Karma: +56/-12
  • Offline Offline
  • Posts: 1101
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #28 on: November 19, 2014, 12:43:03 am »

That is cool.
(What's it for? Skynet?)
its not for Skynet
it IS Skynet!

Eh, I am starting to wonder if this is BCNext's real project.

Logged
Nxt Mission is to commercialize the crypto technology and build new commerce and society.

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #29 on: November 19, 2014, 12:47:35 am »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?
I've used FANN, but dont like the lack of convergence, so I switched to SVM
super optimized the fastest SVM lib out there

cool, homebrew is always the best. i switched also from FANN to Torch7, very nice framework.
just the tensor handling is worth using it.

i have edited my above post adding a TA question. could you take a look?
Logged

abuelau

  • Sr. Member
  • ****
  • Karma: +74/-1
  • Offline Offline
  • Posts: 461
    • View Profile
    • mynxt.info
Re: Helix - Dapps | Test - C&P Repos
« Reply #30 on: November 19, 2014, 12:50:58 am »

this is really cool! The possibilities....

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #31 on: November 19, 2014, 12:55:49 am »

this is really cool! The possibilities....

hopefully you mean the dapp examples and not the AI stuff, this part is just at a - talked into a plastic bag stage - and needs much more research. ;D

ps: will add one last dapp example tomorrow, to test various code fraqment recombinations from different locations, with a simple selfmodifying and
then a persistant (freeze the app) write back to storage by a followed reload from another location. similar than teleporting a VM to another host but
just for apps.
« Last Edit: November 19, 2014, 01:05:06 am by nexern »
Logged

jl777

  • Hero Member
  • *****
  • Karma: +718/-123
  • Offline Offline
  • Posts: 6170
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #32 on: November 19, 2014, 02:55:32 am »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?

edit:
and one more, because iirc you have also wrote some indicators and TA related stuff.
i was looking for a good TA lib a couple of months ago and all i have found was TAlib
which is, beside written in C++ and i would prefer ansi C, very outdated (7-8 years?)
and more important not as complete as i need.

therefore i started to write my own lib. my indicator list is complete now (+150) and only
candlestick and formations are needed. i guess you know how demanding this work is
so do you have any C based resources worth to look at, except TAlib?

not asking for code, will do it by myself, just for resources in case i missed something.
it is strange, that there is such a shortage on good and fresh TA related code out there.
I spent 6 years coding all sorts of filters, signals, predictors, etc. probably half a million lines of code all told. all in various states. never found any TA lib that I could directly use

Next year, I will have time to port my source base to crypto (it is currently all forex oriented). Hopefully we can work together. just need the raw data and I can make 90% accurate predictors for 2hours ahead, maybe 85% for crypto, not sure.

James
Logged
There are over 1000 people in SuperNET slack! http://slackinvite.supernet.org/ automatically sends you an invite

I am just a simple C programmer

jl777

  • Hero Member
  • *****
  • Karma: +718/-123
  • Offline Offline
  • Posts: 6170
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #33 on: November 19, 2014, 02:57:30 am »

this is really cool! The possibilities....

hopefully you mean the dapp examples and not the AI stuff, this part is just at a - talked into a plastic bag stage - and needs much more research. ;D

ps: will add one last dapp example tomorrow, to test various code fraqment recombinations from different locations, with a simple selfmodifying and
then a persistant (freeze the app) write back to storage by a followed reload from another location. similar than teleporting a VM to another host but
just for apps.
GAUL is a decent genetic algo framework. Quite simple to use

James
Logged
There are over 1000 people in SuperNET slack! http://slackinvite.supernet.org/ automatically sends you an invite

I am just a simple C programmer

abuelau

  • Sr. Member
  • ****
  • Karma: +74/-1
  • Offline Offline
  • Posts: 461
    • View Profile
    • mynxt.info
Re: Helix - Dapps | Test - C&P Repos
« Reply #34 on: November 19, 2014, 10:03:52 am »

this is really cool! The possibilities....

hopefully you mean the dapp examples and not the AI stuff, this part is just at a - talked into a plastic bag stage - and needs much more research. ;D

ps: will add one last dapp example tomorrow, to test various code fraqment recombinations from different locations, with a simple selfmodifying and
then a persistant (freeze the app) write back to storage by a followed reload from another location. similar than teleporting a VM to another host but
just for apps.

I meant the idea of storing code in AM...

Evan

  • Sr. Member
  • ****
  • Karma: +18/-2
  • Offline Offline
  • Posts: 333
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #35 on: November 19, 2014, 01:36:58 pm »

its not for Skynet
it IS Skynet!
lol   the legend about skynet is real  :o
Logged
8897015223734827770
NXT-BYRU-8NKZ-PEYJ-9E6YQ

himmelsfaller2

  • Jr. Member
  • **
  • Karma: +4/-0
  • Offline Offline
  • Posts: 64
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #36 on: November 19, 2014, 03:21:24 pm »

Logged

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #37 on: November 19, 2014, 09:52:23 pm »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?

edit:
and one more, because iirc you have also wrote some indicators and TA related stuff.
i was looking for a good TA lib a couple of months ago and all i have found was TAlib
which is, beside written in C++ and i would prefer ansi C, very outdated (7-8 years?)
and more important not as complete as i need.

therefore i started to write my own lib. my indicator list is complete now (+150) and only
candlestick and formations are needed. i guess you know how demanding this work is
so do you have any C based resources worth to look at, except TAlib?

not asking for code, will do it by myself, just for resources in case i missed something.
it is strange, that there is such a shortage on good and fresh TA related code out there.
I spent 6 years coding all sorts of filters, signals, predictors, etc. probably half a million lines of code all told. all in various states. never found any TA lib that I could directly use

Next year, I will have time to port my source base to crypto (it is currently all forex oriented). Hopefully we can work together. just need the raw data and I can make 90% accurate predictors for 2hours ahead, maybe 85% for crypto, not sure.

James

impressive, even lower percentages combined with a solid money management should be sufficient to
monetarize this heavily. would be more than interesting to see this running on other instruments
like stocks or commodities. with those numbers, the only thing you need is a 'direct/real' market access,
at least for currencies, otherwise i guess as soon a so called market maker is in-between, the
slippage/inhouse matching algos are triggered to slow things down. e.g scalping would be very
difficult without direct access.

yes, a cooperation would be good. raw or precalculated data is no problem. i can provide much from
EOD (many years) to intraday 1min (weeks to months) for ~100.000 instruments plus crypto. crypto
is framed at 5min. atm and the current fetchers needs some adjustment due to restricted serverload
on some exchanges but i think this can be lowered to 1min too. crypto needs also additional quality
treatment (e.g detecting spikes), making them more reliable. i am on this soon due to FinHive.

moreover, as soon the TA lib is completed, you can use it to hunt for signal confirmations and other
filters. used on the above mentioned data, integrated into your codebase could be a mighty tool.
Logged

jl777

  • Hero Member
  • *****
  • Karma: +718/-123
  • Offline Offline
  • Posts: 6170
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #38 on: November 19, 2014, 10:02:13 pm »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?

edit:
and one more, because iirc you have also wrote some indicators and TA related stuff.
i was looking for a good TA lib a couple of months ago and all i have found was TAlib
which is, beside written in C++ and i would prefer ansi C, very outdated (7-8 years?)
and more important not as complete as i need.

therefore i started to write my own lib. my indicator list is complete now (+150) and only
candlestick and formations are needed. i guess you know how demanding this work is
so do you have any C based resources worth to look at, except TAlib?

not asking for code, will do it by myself, just for resources in case i missed something.
it is strange, that there is such a shortage on good and fresh TA related code out there.
I spent 6 years coding all sorts of filters, signals, predictors, etc. probably half a million lines of code all told. all in various states. never found any TA lib that I could directly use

Next year, I will have time to port my source base to crypto (it is currently all forex oriented). Hopefully we can work together. just need the raw data and I can make 90% accurate predictors for 2hours ahead, maybe 85% for crypto, not sure.

James

impressive, even lower percentages combined with a solid money management should be sufficient to
monetarize this heavily. would be more than interesting to see this running on other instruments
like stocks or commodities. with those numbers, the only thing you need is a 'direct/real' market access,
at least for currencies, otherwise i guess as soon a so called market maker is in-between, the
slippage/inhouse matching algos are triggered to slow things down. e.g scalping would be very
difficult without direct access.

yes, a cooperation would be good. raw or precalculated data is no problem. i can provide much from
EOD (many years) to intraday 1min (weeks to months) for ~100.000 instruments plus crypto. crypto
is framed at 5min. atm and the current fetchers needs some adjustment due to restricted serverload
on some exchanges but i think this can be lowered to 1min too. crypto needs also additional quality
treatment (e.g detecting spikes), making them more reliable. i am on this soon due to FinHive.

moreover, as soon the TA lib is completed, you can use it to hunt for signal confirmations and other
filters. used on the above mentioned data, integrated into your codebase could be a mighty tool.
I have found that 1 min data is best.
I got a lot of additional accuracy due to forex major 8 currencies having a large amount of constraints.
Things really got into the 90% territory when I made some primitive TA based features. So each minute I recalculated a bruteforce set of TA primitives, in addition to all the other features. I think those would work across all markets.

just from home computer being almost 1 second away from markets, got a fully automated trading system to breakeven, considering the slippage, fees, etc. and not being very optimized, it was actually quite a feat. The remaining problem was some noise that caused spurious trades and I was just about to fix that when ....

I discovered crypto!

James
Logged
There are over 1000 people in SuperNET slack! http://slackinvite.supernet.org/ automatically sends you an invite

I am just a simple C programmer

nexern

  • Sr. Member
  • ****
  • Karma: +83/-11
  • Offline Offline
  • Posts: 496
    • View Profile
Re: Helix - Dapps | Test - C&P Repos
« Reply #39 on: November 19, 2014, 11:08:05 pm »

btw, jl777, do you writing your NNs from scratch or do you use libs like FANN?

edit:
and one more, because iirc you have also wrote some indicators and TA related stuff.
i was looking for a good TA lib a couple of months ago and all i have found was TAlib
which is, beside written in C++ and i would prefer ansi C, very outdated (7-8 years?)
and more important not as complete as i need.

therefore i started to write my own lib. my indicator list is complete now (+150) and only
candlestick and formations are needed. i guess you know how demanding this work is
so do you have any C based resources worth to look at, except TAlib?

not asking for code, will do it by myself, just for resources in case i missed something.
it is strange, that there is such a shortage on good and fresh TA related code out there.
I spent 6 years coding all sorts of filters, signals, predictors, etc. probably half a million lines of code all told. all in various states. never found any TA lib that I could directly use

Next year, I will have time to port my source base to crypto (it is currently all forex oriented). Hopefully we can work together. just need the raw data and I can make 90% accurate predictors for 2hours ahead, maybe 85% for crypto, not sure.

James

impressive, even lower percentages combined with a solid money management should be sufficient to
monetarize this heavily. would be more than interesting to see this running on other instruments
like stocks or commodities. with those numbers, the only thing you need is a 'direct/real' market access,
at least for currencies, otherwise i guess as soon a so called market maker is in-between, the
slippage/inhouse matching algos are triggered to slow things down. e.g scalping would be very
difficult without direct access.

yes, a cooperation would be good. raw or precalculated data is no problem. i can provide much from
EOD (many years) to intraday 1min (weeks to months) for ~100.000 instruments plus crypto. crypto
is framed at 5min. atm and the current fetchers needs some adjustment due to restricted serverload
on some exchanges but i think this can be lowered to 1min too. crypto needs also additional quality
treatment (e.g detecting spikes), making them more reliable. i am on this soon due to FinHive.

moreover, as soon the TA lib is completed, you can use it to hunt for signal confirmations and other
filters. used on the above mentioned data, integrated into your codebase could be a mighty tool.
I have found that 1 min data is best.
I got a lot of additional accuracy due to forex major 8 currencies having a large amount of constraints.
Things really got into the 90% territory when I made some primitive TA based features. So each minute I recalculated a bruteforce set of TA primitives, in addition to all the other features. I think those would work across all markets.

just from home computer being almost 1 second away from markets, got a fully automated trading system to breakeven, considering the slippage, fees, etc. and not being very optimized, it was actually quite a feat. The remaining problem was some noise that caused spurious trades and I was just about to fix that when ....

I discovered crypto!

James

->PM
Logged
Pages: 1 [2] 3  All
 

elective-stereophonic
elective-stereophonic
assembly
assembly