b15bc71fb6
The previous adaptive LOSS controller (rl_reward_clamp_controller) drove
LOSS toward 1.0 floor via observed L/W trade ratio, killing Q's loss
aversion and producing wr=0.27 trend-follower. The dd049d9a4 baseline
achieved wr=0.567 with LOSS=3.0 STATIC, but disabling adaptation
caused NaN at step 4 (other components depend on LOSS-driven signals).
Solution: build a SPECIALIZED ISV controller that drives LOSS toward
maintaining target wr, decoupled from observed L/W trade ratio.
New: rl_loss_aversion_controller.cu
- Single-thread Schulman bounded-step controller
- Input: wr_ema (slot 590, EMA of trade outcomes on done events)
- Target: wr_target (slot 591, bootstrap 0.55 — surfer)
- Output: LOSS clamp (slot 453)
- ±10% adjustment per fire, asymmetric dead-zone (loss aversion bias)
- Bounds [1.5, 5.0], sparse-aware (skips if wr_ema = 0)
Modified rl_reward_clamp_controller.cu:
- Added wr_ema update on done events (single source of truth alongside
win/loss magnitude EMAs)
- LOSS slot 453 ownership transferred to rl_loss_aversion_controller
Local smoke (1000 steps, b=128, seed=16962):
- wr_ema trajectory: 0.32 → 0.42 → 0.53 → 0.49 (climbing toward 0.55)
- LOSS auto-adjusts: 4.71 → 4.38 → 4.29 → 3.52 (tightening as wr rises)
- Action entropy stable 2.0-2.2 (no collapse)
- Reward clamp confirmed active (r_min saturates at -LOSS)
- No NaN, completed_clean: true
Per pearl_loss_clamp_controls_entropy_stability: LOSS≥3.0 correlates
with high wr + stable entropy. Specialized controller maintains this
structural property regardless of observed trade ratio noise.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Foxhunt
Production HFT trading system in Rust.
Architecture
The workspace contains 32 crates organized as follows:
Core Libraries (16)
| Crate | Purpose |
|---|---|
trading_engine |
Order processing, FIX 4.4, IB TWS, SIMD, RDTSC timing |
risk |
VaR, Kelly, circuit breakers, kill switches, compliance |
risk-data |
Risk data types and shared structures |
trading-data |
Trading data types |
ml |
DQN Rainbow, PPO, TFT, Mamba2, ensemble inference |
ml-data |
ML data types and feature definitions |
data |
Market data ingestion and storage |
backtesting |
Replay engine, strategy tester |
adaptive-strategy |
Ensemble execution, microstructure analysis |
common |
Shared types, resilience, error handling |
storage |
S3 and local model storage |
model_loader |
Model serialization and loading |
market-data |
Market data feed handlers |
database |
PostgreSQL access layer (SQLx) |
config |
Configuration management |
tli |
CLI commands and tooling |
Services (8)
| Service | Purpose |
|---|---|
backtesting_service |
gRPC backtesting service |
broker_gateway_service |
FIX routing, broker connectivity |
trading_service |
Core trading operations |
ml_training_service |
Model training orchestration |
data_acquisition_service |
Market data acquisition |
trading_agent_service |
Autonomous trading agents |
api_gateway |
gRPC API gateway with auth |
web-gateway |
Axum REST + WebSocket gateway |
Frontend
web-dashboard/ -- React 19 + TypeScript + Vite + TradingView charts.
Building
# Check compilation (no PostgreSQL required)
SQLX_OFFLINE=true cargo check --workspace
# Run tests for a specific crate
SQLX_OFFLINE=true cargo test -p <crate> --lib
# Clippy
SQLX_OFFLINE=true cargo clippy --workspace
ML Models
Four production model architectures on Candle v0.9.1 with CUDA:
- DQN Rainbow -- Deep Q-Network with prioritized replay, dueling heads, noisy nets
- PPO -- Proximal Policy Optimization with GAE and LSTM policies
- TFT -- Temporal Fusion Transformer for multi-horizon forecasting
- Mamba2 -- State space model for sequence prediction
Each model has a standalone trainer and a UnifiedTrainable adapter for the hyperopt pipeline.
Infrastructure
- Git: Gitea at
git.fxhnt.ai(Tailscale-only), Scaleway DEV1-S - Observability: OpenTelemetry OTLP (env
OTEL_EXPORTER_OTLP_ENDPOINT) - Database: PostgreSQL with SQLx offline mode for CI
License
Proprietary. All rights reserved.