Training deep learning architectures against financial time-series to extract alpha signals, model regime shifts, and generate synthetic data for robust backtesting.
Long Short-Term Memory networks trained on OHLCV candlestick data with attention mechanisms. Captures temporal dependencies in price action for short-horizon forecasting. The model ingests 60-day lookback windows of normalised returns, volume profiles, and technical indicators, outputting a 5-day directional probability distribution.
Transformer-based encoder for classifying market regimes — trending, mean-reverting, high-volatility — using self-attention over multi-asset correlation matrices. Trained on rolling 252-day windows across 500+ equities, the model identifies structural breaks in market dynamics before traditional statistical tests.
Generative Adversarial Network producing realistic synthetic order book data for strategy backtesting. Preserves fat tails, volatility clustering, and microstructure patterns. The discriminator is trained to distinguish real L2 order book snapshots from generated ones, pushing the generator toward statistical fidelity.
PPO-based agent trained against simulated market environments with realistic slippage and transaction costs. Optimises Sharpe-adjusted returns over variable holding periods. The reward function penalises drawdowns asymmetrically, producing agents that are conservative in adverse regimes and aggressive in favourable ones.
End-to-end process from raw market data to deployed prediction models.
Raw tick data, OHLCV, order book snapshots via exchange APIs
Outlier removal, gap filling, normalisation, feature engineering
Walk-forward validation, hyperparameter optimisation, ensemble methods
Out-of-sample testing, Sharpe analysis, regime-conditional performance
ONNX export, WASM compilation, real-time inference endpoint
Core language
Model training
Data processing
Browser inference
Classical ML
Visualisation
Time-series store
Containerisation