Finetuning and traning

[traderpedroso]

Thank you for making this fascinating project available in a way that is applicable to LNN in the stock market. Is there any chance you could share the training code for testing with different assets?

[lorenzmeis]

please find here a notebook I used to train another ticker (SPY). Tested OK with app.py.
It is using the same model config as the AAPL and TSLA models. Based on it you should be able to train new ticker until Hussein releases his code.
I don`t fully understand the reasoning of the regularizer settings of the first Dense layer and how they were calculated.
Hoping that @HusseinJammal can give us some enlightenment here ;-)

model.add(Dense(
units=64,
activation='linear',
kernel_initializer=GlorotUniform(),
bias_initializer=Zeros(),
kernel_regularizer=L1L2(l1=9.999999747378752e-06, l2=9.999999747378752e-05),
bias_regularizer=L1L2(l2=9.999999747378752e-06),
activity_regularizer=L1L2(l1=9.999999747378752e-05),
name='dense'
))

[traderpedroso]

please find here a notebook I used to train another ticker (SPY). Tested OK with app.py. It is using the same model config as the AAPL and TSLA models. Based on it you should be able to train new ticker until Hussein releases his code. I don`t fully understand the reasoning of the regularizer settings of the first Dense layer and how they were calculated. Hoping that @HusseinJammal can give us some enlightenment here ;-)

model.add(Dense( units=64, activation='linear', kernel_initializer=GlorotUniform(), bias_initializer=Zeros(), kernel_regularizer=L1L2(l1=9.999999747378752e-06, l2=9.999999747378752e-05), bias_regularizer=L1L2(l2=9.999999747378752e-06), activity_regularizer=L1L2(l1=9.999999747378752e-05), name='dense' ))

Thanks, buddy! I'm going to start implementing this right now. The cross-validation logs are giving me a good idea. I'll train it and integrate it into backtesting to see how it performs in a more realistic environment. I'll let you know if there are any updates and share the code.

[HusseinJammal]

Hello guys! @lorenzmeis @traderpedroso
First of all, thank you for your feedback and thanks @lorenzmeis for your contribution as well. I really appreciate it!
Regarding the training code, I'm sorry that I couldn't provide for two reasons:

1. The institution I work with had some regulations regarding sharing the code and I didn't have permission to do so.
2. The project is going under some advancements in the model (as a second phase)
   I hope you guys understand this case.

If you need more help or have any concerns, please do not hesitate to contact me.
I'd be happy to assist!

[traderpedroso]

Thank you for the feedback, and no problem! I've made some progress on the code thanks to @lorenzmeis and of course, you @HusseinJammal. I'd like to share my progress in improvements. I did the following:

• Removed all technical indicators and simply used Fibonacci as the base.
• Instead of using it normally, I separated it into three parts: 0% (current price), -100%, +100%. This created a form of price range or channel.
• I mentioned Fibonacci because I used the same calculations, but for different price areas with these three points.
• I created what would be classes, not prices. These classes represent support and resistance and served as a basis for stop and take profit.
• The results are astonishing. I had something similar with Transformer Quantiles, but the liquid NN proved much more efficient, especially for the short term.
• Once I have a production-ready solution, I'll share the code.