Trading signal platform that fetches multi-market data, computes features, trains ML models, evaluates rule-based strategies, runs backtests with walk-forward analysis, applies risk controls, and delivers standardized trading signals via Redis Streams.
┌──────────────┐ ┌──────────────┐ ┌─────────────┐
│ FastAPI │ │ Celery Beat │ │ Consumer │
│ (API :8001) │ │ (scheduler) │ │ (external) │
└──────┬───────┘ └──────┬───────┘ └──────▲──────┘
│ │ │
│ ┌──────▼───────┐ Redis Streams
│ │ CPU Worker │───────────┘
│ │ (fetch/back- │
│ │ test/rules) │
│ └──────────────┘
│ ┌──────────────┐
│ │ GPU Worker │
│ │ (train/pred) │
│ └──────────────┘
│
┌────▼───────┐ ┌───────────┐
│TimescaleDB │ │ Redis │
│(PostgreSQL)│ │ (broker) │
└────────────┘ └───────────┘
src/poseidon/
├── api/ # FastAPI routers
├── backtest/ # Backtest runner, walk-forward, optimizer
├── core/ # Config, schemas, events
├── data/ # Fetchers (FinMind, yfinance, CCXT), feature engine
├── ml/ # Model registry, training, artifacts
├── models/ # SQLAlchemy database models
├── risk/ # Risk rules engine, virtual portfolio
├── signals/ # Signal generation & Redis Streams delivery
├── strategies/ # Rule DSL, model strategy, voting strategy, regime router
├── autoresearch/ # Autonomous parameter search runner, guard, report
└── workers/ # Celery task definitions
# 1. Configure environment
cp .env.example .env
# Edit .env: set POSEIDON_API_KEY and POSEIDON_FINMIND_TOKEN
# 2. Start all services
docker compose up -d
# API available at http://localhost:8001| Market | Symbols | Data Source | Interval |
|---|---|---|---|
| Taiwan Stocks | 22 (TSMC, MediaTek, Hon Hai...) | FinMind | 1d |
| Taiwan Futures | 4 (TX, MTX, TE, TF) | FinMind | 1d |
| US Stocks | 21 (AAPL, NVDA, TSLA...) | yfinance | 1d |
| Crypto Spot | 10 (BTC, ETH, SOL...) | CCXT | 1d, 1h |
Symbol watchlist defined in config/symbols.yaml. Can be modified via API.
All endpoints except /health require X-API-Key header.
| Method | Path | Description |
|---|---|---|
| GET | /health |
Health check (no auth) |
| POST | /data/fetch |
Trigger market data fetch |
| POST | /data/backfill |
Start historical backfill |
| GET | /data/backfill/status |
Check backfill progress |
| GET/POST | /sentiment |
Sentiment scores |
| GET/POST | /strategies |
Strategy CRUD |
| POST | /strategies/{id}/activate |
Activate strategy |
| GET/POST | /api/risk-rules |
Risk rule management |
| GET | /api/risk-rules/portfolio |
Virtual portfolio state |
| GET | /models |
List model versions |
| POST | /models/train |
Train new model |
| POST | /models/{id}/predict |
Run prediction |
| POST | /backtest/run |
Execute backtest |
| POST | /backtest/optimize |
Hyperparameter optimization |
| POST | /autoresearch/run |
Dispatch autoresearch (returns task_id) |
| GET | /autoresearch/status/{task_id} |
Poll autoresearch progress |
| POST | /autoresearch/stop/{task_id} |
Graceful stop |
| GET | /autoresearch/experiments |
Query experiment results |
| GET | /signals |
Signal history |
POSEIDON_API_KEY=<secret> # API authentication key
POSEIDON_FINMIND_TOKEN=<token> # FinMind API token (Taiwan market data)
# Set automatically by docker-compose:
# POSEIDON_DATABASE_URL=postgresql://poseidon:poseidon@postgres:5432/poseidon
# POSEIDON_REDIS_URL=redis://redis:6379/0| Service | Image | Purpose |
|---|---|---|
| api | Dockerfile | FastAPI server (port 8001) |
| beat | Dockerfile | Celery Beat scheduler |
| cpu-worker | Dockerfile | Data fetch, backtest, rules (4 concurrency) |
| gpu-worker | Dockerfile.gpu | Model training & prediction (GPU) |
| postgres | timescale/timescaledb:latest-pg16 | TimescaleDB database |
| redis | redis:7-alpine | Message broker & signal delivery |
# Install dependencies
uv sync --extra dev
# Run tests
pytest
# Run API locally
uvicorn poseidon.main:app --reload --port 8000
# Lint
ruff check src/ tests/uv sync --extra gpu # PyTorch + XGBoostv2.0 放棄 ML 方向預測，改用簡單信號投票 + 自動化迭代搜索。
6-signal 投票策略：Momentum×2, EMA crossover, RSI, MACD histogram, Bollinger squeeze。min_votes 多數決 + ATR trailing stop 出場。
自動化參數搜索框架：
- ParameterSearchPipeline — Optuna TPE + holdout split + WFE validation
- AutoResearchRunner — 跨市場批次搜索，per-market 失敗隔離
- StrategyMutator — 隨機 + Optuna 變異策略參數
- ExperimentTracker — 實驗結果持久化到 PostgreSQL
- Immutability Guard — contextvar 保護 FeatureEngine/BacktestRunner 不被 autoresearch 修改
市場狀態分類系統（gated — 不贏就關）：
- RegimeRouter — 包裝 VotingStrategy，根據 regime 動態調整 min_votes/position_pct
- Regime Labels — 百分位 realized_vol_20 分 3 類 (low/medium/high vol)
- RegimeSearchPipeline — Per-regime Optuna 搜索（只調 2 參數）
- Outperformance Gate — holdout 上嚴格比較，不贏就 bypass
| Decision | Rationale |
|---|---|
| Modular monolith | Single developer, operational simplicity |
| Features computed on-the-fly | No feature store needed; avoids EAV anti-pattern |
| Redis Streams for signals | Guaranteed delivery with consumer groups |
| Celery Beat (redbeat) | Persistent scheduling, survives restarts |
| Strategy DSL | JSON condition trees with boolean combinators |
| Walk-forward backtest | WFE >= 50% threshold for overfitting detection |
| TimescaleDB | Native time-series compression for OHLCV data |
| Voting > ML prediction | Simple rule voting + automated search beats complex ML (v2.0) |
| Outperformance gate | Regime routing must strictly beat static baseline or auto-disable |
| Dynamic feature specs | VotingStrategy tells FeatureEngine exactly what it needs |