Bitcoin Candle Predictor

Simple neural network that tries to predict if the next BTC hourly candle will be green or red. Pulls data from Binance and trains on historical patterns.

What it does

Looks at the last 10 hourly candles and uses 3 features per candle:

• Return (open to close movement)
• High/Low range relative to open
• Volume normalized against 20-period average

Gets around 60-85% accuracy which is honestly pretty decent for something this simple.

Setup

Install dependencies:

pip install torch pandas numpy scikit-learn requests matplotlib

Then run in order:

python download_data.py    # grabs historical data
python train_model.py       # trains the model
python use_model_live.py    # make predictions

The live prediction script lets you either predict once or keep it running to update every hour.

Model

Pretty straightforward neural net:

• Input: 30 features (10 candles × 3 features)
• Hidden layers: 128 → 64 → 32 neurons with dropout
• Output: sigmoid for binary classification

I tried adding more features like momentum and volatility but they actually made it worse. Also bigger networks just overfit.

Warning

Don't actually trade with this. It's just a fun project to mess around with ML and crypto data. If you do trade based on this and lose money that's on you.