Neural Alpha Engine
Transformer-based multi-asset trading system that processes alternative data feeds to generate alpha signals across equity and crypto markets.
System Overview
The Neural Alpha Engine (NAE-v3) is a production trading system that generates alpha signals across 15 equity and 20 crypto assets using a modified Temporal Fusion Transformer architecture. This is the third major iteration of the system — v1 was a simple LSTM-based predictor, v2 introduced multi-asset attention, and v3 added alternative data ingestion and a hierarchical signal aggregation layer.
The system operates on 1-hour and 4-hour timeframes, generating directional predictions with associated confidence scores. Signals above a calibrated confidence threshold are forwarded to a separate execution engine that handles order routing, position sizing, and risk management.
Data Pipeline
NAE-v3 consumes four categories of data, each processed through dedicated ingestion pipelines:
Market Data — Real-time OHLCV data from multiple exchanges (Binance, Coinbase, Interactive Brokers) normalized into a unified schema. Cross-exchange price discrepancies are flagged and arbitrage-filtered to prevent training on stale quotes.
Alternative Data — Aggregated social sentiment scores from Twitter/X and Reddit (processed via a fine-tuned FinBERT model), GitHub commit activity for crypto protocols, and on-chain metrics including active addresses, exchange inflows/outflows, and whale transaction alerts.
Macro Features — VIX, DXY, 10Y Treasury yields, and Fed Funds futures implied probabilities. These are sampled daily and broadcast to all intraday prediction windows as static covariates.
Orderbook Microstructure — Level 2 orderbook snapshots sampled every 10 seconds, compressed into features: bid-ask spread, depth imbalance at 5 levels, and trade flow imbalance using a modified VPIN calculation.
All raw data flows through Apache Kafka into a feature store backed by Redis and PostgreSQL. Features are computed with strict point-in-time correctness — every feature value is timestamped and immutable once written. This eliminates the lookahead bias that plagues most ML trading systems.
Model Architecture
The core model is a Temporal Fusion Transformer with several modifications:
- Variable Selection Networks — Learned gates that automatically weight the importance of each input feature. This is critical because feature relevance shifts across market regimes. During high-volatility periods, orderbook features dominate; in trending markets, momentum indicators carry more weight.
- Hierarchical Attention — Two levels of attention: intra-asset (temporal patterns within a single instrument) and inter-asset (cross-asset dependencies like BTC’s influence on altcoins).
- Quantile Regression Heads — Instead of point predictions, the model outputs the 10th, 50th, and 90th percentile of expected returns. This provides a natural confidence interval that the execution layer uses for position sizing.
The model is retrained weekly on a rolling 18-month window using purged walk-forward validation. Training runs on a 4x A100 GPU cluster and takes approximately 3 hours per full retrain cycle.
Production Deployment
The live system runs on a Kubernetes cluster with the following components:
- Data Ingestion Service — Python workers consuming from exchange WebSocket feeds and writing to Kafka
- Feature Engine — Rust-based feature computation service that processes raw data into model-ready tensors, optimized for low-latency throughput
- Inference Server — FastAPI service wrapping the PyTorch model with TorchScript optimization, serving predictions at sub-10ms latency
- Execution Engine — Handles order routing, partial fill management, and position tracking across exchanges
- Monitoring Stack — Grafana dashboards tracking model prediction accuracy, signal decay rates, portfolio P&L, and system health metrics
All components are containerized with Docker and orchestrated with Kubernetes. Deployments use blue-green switching to ensure zero-downtime model updates. A canary deployment process runs new model versions on paper trading for 48 hours before promoting to live capital.
Performance and Risk Management
NAE-v3 has been live since Q1 2025 with the following verified metrics: Sharpe ratio of 2.1, maximum drawdown of 8.3%, and a 62% win rate on closed positions. The system enforces hard risk limits including a 2% per-position stop loss, a 5% daily portfolio drawdown halt, and automatic deleveraging when rolling 30-day Sharpe drops below 0.5. These risk controls operate independently of the model and cannot be overridden programmatically.