metadata
language: en
tags:
- trading
- candlestick
- transformer
- btc
- cryptocurrency
license: mit
Candle Transformer 46M
Custom transformer LLM for BTC/USDT trading using candlestick data.
Model Details
- Parameters: 46,355,576
- Architecture: 12 layers, 8 heads, 512d, 2048ff
- Training: 1-year BTC data (1d + 4h + 1h = 11,315 candles)
- Tokenizer: Hybrid (price bins + pattern tokens + indicator tokens)
Usage
from inference import CandlePredictor
predictor = CandlePredictor("Ashu9675/candle-transformer-46m/best_model.pt")
pred = predictor.predict(candles, timeframe="1h")
# Returns: signal (BUY/SELL/HOLD), confidence, probabilities, next_candle
Results
- Overall accuracy: 46.7% (backtest, 30 windows)
- Signal: BUY 67%, SELL 31%, HOLD 1%
- Inference: ~4.5s on CPU
See benchmark_report.md for full details.