Algorithmic Trading AI with Python

Usage Guide Training and Testing

Install Dependencies

pip install -r requirements.txt

HuggingFace Hub Login

For the Gemma Embedding model, you need to login to HuggingFace Hub

hf auth login

Model trains on 80% of data, tests on 20% of data.

python download.py ## (OPTIONAL) Download New Data
python train.py    ## (OPTIONAL) Re-Train the model
python test.py     ## Test the model ( recommend modify parameters in the script )

💡 The model is already trained, so running train.py is optional. Find the trained model saved as gemma_transformer_classifier.pth

To Profit from the model, probabaly??!?!?

In order to use the model, copy test.py to profit.py and modiy it to use Yfinance websocket data stream. https://ranaroussi.github.io/yfinance/reference/yfinance.websocket.html When a price signal is generated, load news (see download.py) process the data the same way in test.py and get the model output.

⚠️ Be carful, this is not financial advice, and the model is not guaranteed to be profitable.

Tuning and Optimization

• ✅ Save and load model
• ✅ Calcluate the Cost
• ✅ Separate training and validation data
• ✅ More data again!
• ✅ SGD
• ✅ Activations
• ✅ More DATA!
  • ✅ More Data ( pull in from download.py)
  • ✅
• Capture median price % deltas and use that as market indicator
• Reinforcement learning
• Reduce dimentionality of embeddings
• Change input data add more Symbols
• Labels threshold changes ( reduce from 0.1% to 0.05% )
• Batch Size
  • Multi-stage Batch Training
• Dropout Rate
• Learning Rate
• More Input
  • Time of day (vectorize, day of week )
• Learning rate adjustments
• Sequence Data ( merge sequences together )
• Embedding Cacheing ( save embeddings hash )
• ~~Instead of Gemma UniformScalar + NLP Embedding~~

Target Market Pattern

• Volatile Market
• Lots of news events
• High volume

Libraries

• Yfinance (datasource)
• Pytorch Transformer
• Gemma embedding 300m GOOGLE
• Numpy
• matplotlib
• Pandas
• Scikit-learn

Target Symbol

• BTC-USD (Bitcoin to US Dollar)
• AAPL (Apple Inc.)
• TSLA (Tesla Inc.)
• GC=F (Gold Futures)

Inputs and outputs

Input data

• Historical price data
• News headlines sentiment will be derived
• Foreign policy events

Outputs

• Sell/Hold/Buy signals

Model

• Gemma Embedding (300m parameters)
• Pytorch Transformer
• Classification Layer ([3] output)

Training

• Supervised learning
• Reinforcement learning

Input Data Format

Sorted by date, descending.

Prices: list of prices
Headline: list of news headlines

Prices: 10.0,11.0,14.0,8.0,10.0,5.0
Headline: Sector Update: Financial Stocks Rise Tuesday Afternoon

Prices: 10.0,11.0,14.0,8.0,10.0,5.0
Headline: Sector Update: Financial Stocks Rise Tuesday Afternoon

Prices: 10.0,11.0,14.0,8.0,10.0,5.0
Headline: Sector Update: Financial Stocks Rise Tuesday Afternoon

Ouptput Data Format

Sell/Hold/Buy signals [0, 0, 0 ]

Sell = [1, 0, 0] Hold = [0, 1, 0] Buy = [0, 0, 1]

Limit Triggers Indications

• Average up and down cycles
• 1% limit triggers 100 gain 1 dollar 101
• time window is 5 minutes