AI talking | Page 4 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

which fibonn was vegas talking about? 1 reply

Changing Your Self Talk By Talking to Yourself 2 replies

My left brain starts talking to me 7 replies

talking to myself(smile) 8 replies

Attachments: AI talking
Exit Attachments

AI talking

  • #61
  • Jan 20, 2023 11:54pm Jan 20, 2023 11:54pm
I wonder. Can you ask chatGPT for signals too?
  • #62
  • Jan 21, 2023 3:19am Jan 21, 2023 3:19am
Quoting xhiro
I wonder. Can you ask chatGPT for signals too?
Possible.
If that`s what you want, but for long term, I don`t think this is a solution.
It depends what you want.

Have a good trade and
More succes!
It is not about what you think, but what you see and your actions after!
  • #63
  • Jan 21, 2023 12:23pm Jan 21, 2023 12:23pm
Quoting ryuryu
Hey joyny! I'm playing with openai now hard. Yesterday I was googling, then stackoverflowing, then checking the things and today I just ask AI to make a code for me. And it works good. At least much faster than first approch. But I have faced some issues. For example I was trying to implement ClickHouse database. And I asked it how to delete all data in the database > 30 min for example. And you know what? I tried 100 times and it always told me "DELETE FROM...". But it is not correct! Correct is ALTER...
Openai is not 100% reliable. I had several issues with it..

for example ai could not say correct that trading 0.01 lots on 100 USD account is same risk as 1 lot on 10000 usd account. Had to ask additional questions till finally admits that risk % the same.

But worst incident so far was when asked top gold owners central banks and it first said that China 68 mill tons and Russia 46 mill tons and then US with 8 mill... When confronted ai with info from google it admited error. But could not provide sources for wrong info... This is bad.. should be posible to allways get source of info...
  • #64
  • Jan 21, 2023 12:24pm Jan 21, 2023 12:24pm
Quoting xhiro
I wonder. Can you ask chatGPT for signals too?
Openai have data on 2021. And no updating with fresh up to date info for now.
  • #65
  • Jan 25, 2023 10:57am Jan 25, 2023 10:57am
Quoting joyny
{quote} Openai have data on 2021. And no updating with fresh up to date info for now.
seems rather useless except for learning general/approved information.
The benefits being good understanding and fast and highlighted points.

It doesn’t want to give the same shortcuts to making money, such as the big banks use everyday to a commoner/ordinary citizen.

Am i using the wrong version? I use chatGPT. Or did it`s code change?
You have an IQ lower than 90 if you interact with this user.
  • #66
  • Jan 26, 2023 2:24pm Jan 26, 2023 2:24pm
Quoting Ferna
{quote} seems rather useless except for learning general/approved information. The benefits being good understanding and fast and highlighted points. It doesn’t want to give the same shortcuts to making money, such as the big banks use everyday to a commoner/ordinary citizen. Am i using the wrong version? I use chatGPT. Or did it`s code change?
openAI have data only till 2021.

But.. it is possible to use their API to fine-tune with fresh learning data for AI. This requires coding skills.
  • #67
  • Jan 26, 2023 8:04pm Jan 26, 2023 8:04pm
Quoting joyny
{quote} openAI have data only till 2021. But.. it is possible to use their API to fine-tune with fresh learning data for AI. This requires coding skills.
Is there anyone with a webcrawler associated with deepweb/darknet? It feels like i know the answers already when i ask something. I had hoped for more.
You have an IQ lower than 90 if you interact with this user.
  • #68
  • Jan 29, 2023 11:43am Jan 29, 2023 11:43am
Tried to train (fine tune) my OpenAI instance to trade EURAUD.. Here steps what i did:

1. Developed MQL4 script to export training data in jsonl file. Prompts and correct answers

2. Created 2 files - one for training AI (90000 H1 candles data) and one for validation (10000 H1 candles data)

3. Installed python on my windows, created .py files according to API documentation and then uploaded my 2 files to openai and created fine tune request for model "curie"

4. Retrieved tuning result and it was failed with this error: "Fine-tune failed. Fine-tune can not exceed $15 during free trial"

Google to get estimate how much could cost fine tuning..

Quote
Assuming you use a data set of 200,000 examples and training iterations of 10, the cost to fine-tune GPT-3 would be about $200

in my case i had 100000 examples (training + validation)

so i supose it would cost for me 100 usd.

will try to make smaller files for training and then validation but... this wont be usable for trading then.. just to see how training works..and then how AI responses will work..

will move to most cheapest model "Ada".. maybe that will help too (but again.. maybe worse trading results then..)
  • #69
  • Jan 29, 2023 11:51am Jan 29, 2023 11:51am
Quoting joyny
{quote} Openai is not 100% reliable. I had several issues with it.. for example ai could not say correct that trading 0.01 lots on 100 USD account is same risk as 1 lot on 10000 usd account. Had to ask additional questions till finally admits that risk % the same. But worst incident so far was when asked top gold owners central banks and it first said that China 68 mill tons and Russia 46 mill tons and then US with 8 mill... When confronted ai with info from google it admited error. But could not provide sources for wrong info... This is bad.. should...
It is a cunning liar and every 'fact' that you get from it needs to be double-checked. It can sound very reasonable and be completely wrong.
  • #70
  • Jan 29, 2023 11:58am Jan 29, 2023 11:58am
Quoting clemmo17
{quote} It is a cunning liar and every 'fact' that you get from it needs to be double-checked. It can sound very reasonable and be completely wrong.
Yeah i definitively experienced it myself. For me it made up books that most proably doesn't exist.
It's hard to let it state facts, and if you ask about anything censored it crashes.
You have an IQ lower than 90 if you interact with this user.
  • #71
  • Jan 29, 2023 12:00pm Jan 29, 2023 12:00pm
Here how would look my prompt message:

M11 D3 H18 RSI=(38 39 59 50 46 31 35 36 45 33 44 49 54 50 52 54 52 48 50 47 45 47 50 50 55 45 36 28 32 25 28 31 21 31 31 36 39 32 41 49 53 53 53 52 54 54 53 47) MA=(1.54100 1.54169 1.54202 1.54214 1.54254 1.54329 1.54403 1.54462 1.54504 1.54516 1.54533 1.54530 1.54521 1.54518 1.54525 1.54511 1.54472 1.54400 1.54340 1.54301 1.54260 1.54229 1.54198 1.54171 1.54155 1.54184 1.54252 1.54345 1.54441 1.54568 1.54722 1.54859 1.54989 1.55109 1.55204 1.55275 1.55337 1.55383 1.55408 1.55412 1.55415 1.55476 1.55558 1.55627 1.55698 1.55745 1.55767 1.55772) MA2=(1.54606 1.54653 1.54694 1.54728 1.54767 1.54809 1.54846 1.54872 1.54894 1.54911 1.54925 1.54932 1.54938 1.54946 1.54951 1.54948 1.54943 1.54935 1.54926 1.54923 1.54921 1.54916 1.54911 1.54904 1.54893 1.54882 1.54878 1.54881 1.54887 1.54894 1.54902 1.54905 1.54901 1.54894 1.54882 1.54863 1.54843 1.54823 1.54804 1.54778 1.54745 1.54712 1.54680 1.54648 1.54616 1.54582 1.54548 1.54518) ATR=(432 407 321 294 289 262 257 277 271 266 230 230 223 227 237 270 296 299 325 311 331 357 373 428 429 416 389 360 358 375 374 347 330 271 253 222 218 230 204 199 230 278 300 315 333 358 392 395)

And I expect 3 possible answers: BUY, SELL, WAIT

I train AI to trade with TP 20 and SL 20 pips.

Anyway this is pattern trading.. AI just would try to learn from my training data patterns I cant imagine. still patterns. no guaranties in future. but can combine portfolios of such AI patterns..

100 usd for 1 training is too expensive. 10 usd would be ok.
  • #72
  • Edited 1:11pm Jan 29, 2023 12:04pm | Edited 1:11pm
Quoting joyny
will try to make smaller files for training and then validation but... this wont be usable for trading then.. just to see how training works..and then how AI responses will work.. will move to most cheapest model "Ada".. maybe that will help too (but again.. maybe worse trading results then..)
seems this helped..

training 9000 H1 data
validation 1000 H1 data

most cheapest model "ada".

now in training progress.. 25% so far.

cost of this fine tune: 6.13 usd

processed 100% but detected that my training data was wrong...

Edited:
now prepared correct version of training data.. fine tuning started.. will cost additional $7.36
  • #73
  • Jan 29, 2023 1:55pm Jan 29, 2023 1:55pm
all done. fine-tuning ended.. but.. results are weird..

when I send my special prompt.. AI did not reply with one of 3 answers expected (as in the training data provided) but returned this mess instead:

SELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSEWAWAITSEWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAWAWAITWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWA

Conclusions:

The cheapest model "ada" seems to be not usable for trading purposes.

The more expensive (and smarter) model "curier" cost 5x more and "davinci" cost 50x more.

In order to to use 100000 training data rows for 1 fine tuning will cost:
1) ada ~70 usd
2) curier ~350 usd
3) davinci ~ 3500 usd

This is too expensive for me.. it is just 1 fine tune.. but maybe 100x more fine tunes needed to get something meaningful from OpenAI

And here is more... anyway, this is a patterns trading approach. Patterns fluctuate. AI walks thru training data and learns... it is the same as running MT5 optimizations. There are the same overfit issues for AI as it is for MT5.

I won't use so expensive AI. but maybe banks etc will use it and benefit a lot. maybe.. but they are doing it already with other AIs..

Quote
Microsoft and NVIDIA have a long history of working together to support financial institutions by providing cloud, hardware, platforms, and software to support algorithmic trading. Microsoft Azure cloud, NVIDIA GPUs and NVIDIA AI provide scalable, accelerated resources as well as routines, and libraries for automating quantitative analysis...

I can't test this openai for trading. I am a regular person with small capital

Therefore I stick to my self-made baby-neuron AI with just 10 neurons.. and this is still enough to find patterns for my portfolios.
  • #74
  • Jan 29, 2023 2:24pm Jan 29, 2023 2:24pm
Quoting joyny
all done. fine-tuning ended.. but.. results are weird.. when I send my special prompt.. AI did not reply with one of 3 answers expected (as in the training data provided) but returned this mess instead: SELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSELLSEWAWAITSEWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAITWAWAWAWAITWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWAWA...
The only way for smaller players like us to succeed with such is to cooperate. The big banks got thousands of people on a pay-roll to do it.
You should hook-up with people with similiar trouble and ambitions.
You have an IQ lower than 90 if you interact with this user.
  • #75
  • Edited 8:39am Jan 30, 2023 7:43am | Edited 8:39am
Spent about 3 hours with OpenAI guidance and learned how to make simple AIs with python.

Attached Image (click to enlarge)
Click to Enlarge Name: ai_python_learning.png Size: 24 KB

Here is what I did:

1. Made another MQL4 script to export CSV training data for H1 history

2. With OpenAI help (explanations and generated python code) made simple AI

3. Prepared 8000 rows (~1 year) with H1 data for training and 2000 rows (~3 months) for validation

4. Trained my simple AI and saved the trained/validated model in a file

5. Made another python script to load the saved model file and make predictions - AI tells me what to do: BUY, SELL or do nothing (WAIT)

Simple python AI model forward-tested with the unknown for AI data 1000 H1 candles (~40 days) and here results for EURAUD:

Attached Image (click to enlarge)
Click to Enlarge Name: ai_model_performance.png Size: 1 KB

62% of H1 predictions were accurate - AI predicted correctly BUY, SELL or WAIT.

Was done 455 deals of which 57% were profitable. As there is TP 20 and SL 20 then total trading result for 40 days is profits ~1300 pips.

A python AI can be integrated with MQL5 expert advisor. And it is possible to develop AI re-learning cycles for example every month re-train. All that can be automated.

Will continue to develop and continue to train/forwardtest various data sets (to find out correlations for large data sets or how deep history to take for training etc) and implement additional stats because I need to know drawdown too.

Edited:

Implemented more stats for forward-testing:

Attached Image (click to enlarge)
Click to Enlarge Name: ai_model_performance2.png Size: 3 KB


When trading 100 usd account with 0.01 lots max DD would be ~33$ and profits in ~40 days would be ~240 $.
  • #76
  • Jan 30, 2023 8:37am Jan 30, 2023 8:37am
Jesus take the wheel!
Quoting joyny
forwardtest
There's one catch of many catches. You trained your model on a training set and tested it on a test set. The moment you start to forward test one bar into the future your model begins to diverge. If you do not continuously retrain your model on the most recent data your model will accumulate divergence and perform worse and worse.
There's one cool example I mentioned in passing in my older neural network strategy post. You create a test model by training it on a set of predictable data where you predict one column that does simple addition of values in specific prior cells. The model is basically supposed to calculate 2+2=4. When you get that to work and plot the graph of the predicted results into the future you do get 2+2=4 (with negligible error) on the first forward increment and it starts to diverge on every sequential value even though the data really just has simple addition in it. So there's always a drift. In fact, a small drift may mean your model is good because it is not overfitted to data.
  • #77
  • Jan 30, 2023 8:45am Jan 30, 2023 8:45am
Quoting costos
Jesus take the wheel! {quote} There's one catch of many catches. You trained your model on a training set and tested it on a test set. The moment you start to forward test one bar into the future your model begins to diverge. If you do not continuously retrain your model on the most recent data your model will accumulate divergence and perform worse and worse. There's one cool example I mentioned in passing in my older neural network strategy post. You create a test model by training it on a set of predictable data where you predict one column that...
Yes, the divergence of a model when using it on future data is a common issue in machine learning and is known as overfitting. The model is only able to accurately predict values based on the data it was trained on, so as new data becomes available, the model may start to diverge. It's important to continuously retrain the model on the most recent data to reduce the divergence and improve its accuracy. A small drift may indicate that the model is not overfitted to the data, but it still needs to be monitored and retrained as necessary to avoid significant divergence.
  • #78
  • Jan 30, 2023 8:49am Jan 30, 2023 8:49am
Quoting joyny
{quote} .
Do you also get the same model performance if you train it on exactly the same training and test datasets? If you do, it's not because you've found the most amazing model. It's because it's not randomly seeded each time you run it. Once you fix the randomness of weight seeding you will see that every time you get a different model from the same training and test datasets that performs differently each time.
  • #79
  • Jan 30, 2023 9:13am Jan 30, 2023 9:13am
Quoting costos
{quote} Do you also get the same model performance if you train it on exactly the same training and test datasets? If you do, it's not because you've found the most amazing model. It's because it's not randomly seeded each time you run it. Once you fix the randomness of weight seeding you will see that every time you get a different model from the same training and test datasets that performs differently each time.
thank you! you make good points

Here is what OpenAI answered to me after several discussion iterations where I confronted OpenAI with your posts:

Attached Image (click to enlarge)
Click to Enlarge Name: default_params.png Size: 13 KB

I'm using the Decision Trees algorithm for my AI model and right now, I'm using the default parameters. However, I can create loops and experiment with different values for the parameters to see how it affects the model's performance. One of the parameters I can adjust is called "random_state", which you mentioned. By changing this parameter and others, I can get a variety of different models to compare and choose the best candidates - because I believe a portfolio of n AI models running will be needed anyway because this is patterns trading and patterns performance fluctuates.
  • #80
  • Jan 30, 2023 10:31am Jan 30, 2023 10:31am
Anyone else got trouble logging in?
or have my curiousity banned me once again,,,
You have an IQ lower than 90 if you interact with this user.
0 traders viewing now