The Hackworth Series is AlphaNet's flagship quantitative engine. It is not a single model but an ensemble orchestration layer that combines up to 20 distinct alpha signals — each produced by its own specialized deep learning model — into one unified, dynamically hedged portfolio. Below is an overview of the architecture, data pipeline, search methodology, and trading mechanics.
Model-Agnostic Architecture
The engine runs a heterogeneous ensemble of deep learning model families, dynamically weighted by a model router that allocates compute based on prevailing market entropy:
Transformers handle long-range regime context. LSTM/GRU variants model sequential momentum and trend persistence. CNN extractors identify local microstructure patterns — support/resistance clusters, order book imbalance artifacts. The router shifts weight in real-time as market conditions change.
Data Pipeline
Multi-modal ingestion processing ~50,000 data points per minute per pair:
Price & Volume (L1/L2 OHLCV)
Order Book (bid/ask depth, imbalance)
Funding & Open Interest (perps)
Cross-Market Correlation (10+ pairs)
On-Chain Flows (exchange in/out)
Volatility Surface (realized / implied)
Features are normalized through adaptive z-scoring with rolling windows that auto-adjust to regime volatility.
Smart Parameter Search
Rather than brute-force enumeration, Hackworth uses an intelligent genetic algorithm with early stopping to navigate a high-dimensional configuration space. The search covers:
Model Architectures:
Transformer, LSTM, GRU, CNN,
CNN-Transformer hybrid, Autoencoder,
Attention-only, Gated residual variants
Prediction Horizons:
1H, 2H, 4H, 8H, 12H, 1D, 3D, 7D
Feature & Dataset Combinations:
256 subsets across price, volume,
order book, funding, on-chain,
cross-market, volatility surface
Signal Mechanisms:
Threshold crossing, z-score mean reversion,
momentum breakout, volatility expansion,
pattern completion, flow divergence
Trading Layer Config:
Position sizing (8 risk models),
Entry/exit logic (64 threshold combos),
Ensemble weighting (16 schemes),
Drawdown circuit breaker levels
The smart search prunes 99.7% of candidates before full backtesting, surfacing high-probability configurations without exhaustive computation.
Dynamic Trading Rules
A meta-learning controller adjusts trading rules in real-time through four integrated mechanisms:
- Regime Detection: Proprietary unsupervised learning model classifies market regime across multiple dimensions: trend strength, entropy, volatility, macro correlation, and volume profile. This multi-factor classification enables finer-grained adaptation than single-signal approaches
- Alpha Selector: Re-weights the 20 constituent alphas based on regime — trend alphas ramp in directional markets, mean-reversion alphas dominate in ranges
- Kelly Sizing: Fractional Kelly criterion with drawdown circuit breakers; max single-trade risk capped at 2%
- Correlation Hedge: When inter-alpha correlation exceeds 0.4, exposure is reduced automatically to prevent cluster risk
Multi-Alpha Composition: 20 Distinct Strategies
Each "alpha" is a complete, independently trained deep learning strategy with its own pipeline, architecture, and signal generator. They are trained on different horizons, features, and objectives to ensure low cross-correlation (target <0.15 pairwise):
Alpha 01-05: Momentum / Trend Following (LSTM/GRU)
Alpha 06-10: Mean Reversion (CNN + Transformer)
Alpha 11-14: Volatility Expansion (autoencoder anomaly)
Alpha 15-17: Cross-Market Arbitrage (correlation break)
Alpha 18-20: Microstructure (order book, flow prediction)
The ensemble layer combines all 20 signals via a learned attention mechanism that adapts to which alpha types are performing best in the current regime. When one alpha class underperforms, others typically compensate, producing smoother equity curves than any single-strategy approach.
Online Training
Models update every 4 hours via continual learning with Elastic Weight Consolidation (EWC), preventing catastrophic forgetting while adapting to evolving microstructure. All updates are validated in a shadow sandbox before production promotion.
Risk Control Layer
A dedicated risk network monitors portfolio heat: gross exposure (max 130%), net exposure (variant-dependent 17-32%), daily drawdown circuit breakers (-5% reduce, -10% halt), and VaR-based dynamic leverage scaling.
