Machine Learning Pipeline

Explore the architecture of our predictive models. We employ ensemble methods combining deep sequence learners with highly tuned gradient boosting.

Causal GRU & Transformer

Our flagship neural network processes 120-step historical sequences of our 45-dimensional feature vectors. The 3-layer Gated Recurrent Unit architecture incorporates self-attention mechanisms to weigh the importance of different historical states without future-peeking.

Handles non-linear relationships in market microstructure.
Trained across multiple prediction horizons simultaneously (1h to 1w).
Live inference running in production with sub-second latency.

XGBoost Ensemble

Complementing the neural nets, our XGBoost layer handles structured tabular data exceptionally well. It provides robust baseline predictions and is used heavily in feature importance ranking, highlighting which signals actually drive price momentum.

Continuous Retraining

Crypto markets suffer from extreme concept drift. Our pipeline supports continuous live-retraining configurations (via Celery workers), ensuring our models adapt to new volatility regimes and market conditions dynamically.

View Live Model Performance