gavindiaz/NexQuant
1
0
Fork 0
mirror of https://github.com/NicolasBohn/NexQuant.git synced 2026-07-27 15:37:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

docs: update README — Kronos-small, test depth, daemon setup, project structure

Browse Source 
- Kronos: auto-integrated into fin_quant, 3 horizons, model size table
- Tests: 1125+ collected (was 134), property-based + fuzzing
- CLI: updated commands, removed outdated Kronos commands
- llama-server: reduced to 18 GPU layers for Kronos co-existence
- Daemon setup: auto-restart commands for fin_quant, autopilot, live trader
- Project structure: scripts/, test/ details, current architecture
This commit is contained in:
TPTBusiness
2026-05-09 09:39:06 +02:00
parent ba7458fba5
commit 85b56b8179
1 changed files with 82 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+82 -46
View File
@@ -77,14 +77,15 @@ rdagent predix
**Predix** is an autonomous AI agent for quantitative trading strategies in the EUR/USD forex market. Built on a multi-agent framework, Predix automates the full research and development cycle:
- 📊 **Data Analysis** – Automatically analyzes market patterns and microstructure
- 💡 **Strategy Discovery** – Proposes novel trading factors and signals
- 🧠 **Model Evolution** – Iteratively improves predictive models
- 📈 **Backtesting** – Validates strategies on historical 1-minute data
- 📊 **Factor Generation** — LLM proposes novel alpha factors; Kronos foundation model generates OHLCV-based predictions
- 💡 **Strategy Discovery** — Autopilot generates + backtests trading strategies 24/7
- 🧠 **Model Evolution** — CoSTEER iteratively improves predictive models through code evolution
- 📈 **Backtesting** — Unified engine with 10 runtime invariants on 1-min EUR/USD data (2020–2026)
- 🔄 **Auto-Restart** — All services run as daemons with automatic crash recovery
Predix is optimized for **1-minute EUR/USD FX data** (2020–2026) and uses Qlib as the underlying backtesting engine.
Predix is optimized for **1-minute EUR/USD FX data** (2020–2026) and supports both local LLMs (llama.cpp) and cloud backends (OpenRouter).
> **Backtest Verification**: Every backtest result is automatically verified at runtime against mathematical invariants (MaxDD ∈ [-1,0], WinRate ∈ [0,1], Sharpe finite, sign consistency, etc.). 479 unit tests + 10 ground-truth validation tests ensure ~99% metric correctness. See [Backtest Integrity](#backtest-integrity).
> **Backtest Verification**: Every backtest result is automatically verified at runtime against mathematical invariants (MaxDD ∈ [-1,0], WinRate ∈ [0,1], Sharpe finite, sign consistency, etc.). 1125 collected tests with deep property-based, fuzzing, and hypothesis tests ensure metric correctness. See [Backtest Integrity](#backtest-integrity).
## Acknowledgments
@@ -222,10 +223,10 @@ QLIB_DATA_DIR=~/.qlib/qlib_data/eurusd_1min_data
```bash
~/llama.cpp/build/bin/llama-server \
--model ~/models/qwen3.6/Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf \
--n-gpu-layers 24 \
--n-gpu-layers 18 \
--no-mmap \
--port 8081 \
--ctx-size 240000 \
--ctx-size 260000 \
--parallel 2 \
--batch-size 512 --ubatch-size 512 \
--host 0.0.0.0 \
@@ -234,10 +235,10 @@ QLIB_DATA_DIR=~/.qlib/qlib_data/eurusd_1min_data
```
> **Important flags:**
> - `--ctx-size 240000 --parallel 2` — allocates **2 slots × 120,000 tokens each**. `fin_quant` prompts can reach 80k+ tokens with full factor history; a smaller slot causes silent overflow and empty responses.
> - `--ctx-size 260000 --parallel 2` — allocates **2 slots × 130,000 tokens each**.
> - `--reasoning off` — **critical**: completely disables Qwen3 chain-of-thought. `--reasoning-budget 0` is not sufficient and produces empty JSON responses.
> - `--n-gpu-layers 24` — 4 fewer than maximum on RTX 5060 Ti (16 GB), freeing ~500 MB VRAM for the larger KV cache.
> - `-ctk q4_0 -ctv q4_0` — quantises the KV cache to 4-bit, reducing VRAM from ~5 GB to ~1.3 GB at 240k context.
> - `--n-gpu-layers 18` — reduced from max (33) to free ~7 GB VRAM for Kronos-small GPU inference alongside llama-server.
> - `-ctk q4_0 -ctv q4_0` — quantises the KV cache to 4-bit, reducing VRAM usage.
### Data Configuration
@@ -334,13 +335,19 @@ rdagent server_ui --port 19899 --log-dir git_ignore_folder/RD-Agent_workspace/
python predix.py best
```
### 3. Run Continuously
### 3. Run Continuously (Auto-Restart)
```bash
while true; do
rdagent fin_quant
sleep 5
done
# Start all services with auto-restart daemons:
# fin_quant — factor R&D loop
nohup bash -c 'while true; do rdagent fin_quant --loop-n 10 --model local >> /tmp/fin_quant_daemon.log 2>&1; sleep 10; done' &
# Autopilot — 24/7 strategy generator (Kronos factors auto-selected)
nohup python scripts/predix_autopilot.py >> /tmp/autopilot_daemon.log 2>&1 &
# Live Trader — FTMO FIX API (requires credentials)
nohup python git_ignore_folder/live_trading/ftmo_live_trader.py >> ftmo_live_trader.log 2>&1 &
```
---
@@ -365,18 +372,19 @@ done
| `python predix.py best` | Show top strategies by composite score |
| `python predix.py best -n 20 -m sharpe` | Top 20 by Sharpe ratio |
| `python predix.py best --show NAME` | Full metadata for one strategy |
| `python predix_gen_strategies_real_bt.py` | Generate 10 strategies with LLM + real backtest |
| `python predix_gen_strategies_real_bt.py 20` | Generate 20 strategies |
| `python predix_gen_strategies_real_bt.py 10` | Generate 10 strategies with LLM + real OHLCV backtest |
| `python predix_gen_strategies_real_bt.py 20` | Generate 20 strategies (parallel workers) |
| `python scripts/predix_autopilot.py` | 24/7 Auto-Pilot: endless strategy generation |
| `python predix_continuous_strategies.py` | Continuous generation with ML training
### Kronos Foundation Model
| Command | Description |
|---------|-------------|
| `python predix.py kronos-factor` | Generate Kronos predicted-return factor (daily stride, ~15 min GPU) |
| `python predix.py kronos-factor --pred 30` | 30-bar prediction horizon |
| `python predix.py kronos-factor --device cpu` | CPU inference (slower) |
| `python predix.py kronos-eval` | Evaluate Kronos IC / hit rate vs LightGBM baseline |
| `python predix.py kronos-eval --pred 96` | Daily horizon evaluation |
| `rdagent fin_quant` | Kronos factors auto-generated on startup (3 horizons) |
| Model size: `KRONOS_MODEL_SIZE=small\|mini\|base` | Configurable via env (default: small) |
Kronos runs automatically — no separate command needed. Factors are regenerated if missing from `results/factors/`.
### Factor Evaluation
@@ -450,30 +458,38 @@ Real-time dashboard for monitoring:
### 🤖 Kronos Foundation Model Integration
Predix integrates [Kronos-mini](https://github.com/shiyu-coder/Kronos) — a 4.1M parameter OHLCV foundation model pretrained on 12+ billion K-lines from 45 global exchanges (AAAI 2026, MIT):
Predix integrates Kronos — an OHLCV foundation model from the NeoQuasar team (AAAI 2026, **MIT License**) — for alpha factor generation:
- **Option A — Alpha Factor**: Rolling daily inference generates a `KronosPredReturn` factor. Every 96 bars (one trading day), Kronos predicts the next day's return from the previous 512 bars of EUR/USD OHLCV data. The factor is forward-filled to 1-min frequency and plugs directly into Predix's factor evaluation pipeline.
| Model | Params | p24 IC | Best For |
|-------|--------|--------|----------|
| **Kronos-small** (default) | 25M | \|IC\| ≈ 0.09 | 1-min EUR/USD |
| Kronos-mini | 4.1M | \|IC\| ≈ 0.07 | Low-resource |
| Kronos-base | 102M | \|IC\| ≈ 0.002 | Daily/weekly data only |
- **Option B — Model Evaluation**: Kronos runs alongside LightGBM as a standalone predictor. IC (Information Coefficient), IC IR, and directional hit rate are computed over the full dataset for direct comparison with LightGBM-generated models.
Kronos generates 3 prediction-horizon factors automatically on `fin_quant` startup:
- `KronosPredReturn_p24` — 24-minute horizon
- `KronosPredReturn_p48` — 48-minute horizon
- `KronosPredReturn_p96` — 96-minute horizon (best performer)
The model runs on GPU (CUDA) alongside the llama-server, using CPU as fallback.
Factors are persisted in `results/factors/` for use by the strategy orchestrator.
```bash
# One-time setup
git clone https://github.com/shiyu-coder/Kronos ~/Kronos
# Kronos runs automatically with fin_quant (no separate command needed)
rdagent fin_quant --loop-n 10 --model local
# Generate factor (Option A) — saves to results/factors/
python predix.py kronos-factor
# Evaluate as model (Option B) — prints IC vs LightGBM reference
python predix.py kronos-eval
# Model size is auto-detected and configurable via env
# Set KRONOS_MODEL_SIZE=base to use the 102M-param model
```
### 🔒 Security & Quality
Automated quality assurance:
- **134+ Tests** — all features tested automatically on every commit
- **1,125+ collected tests** — deep property-based, fuzzing, and hypothesis tests on every commit
- **Bandit Security Scanner** — pre-commit security checks
- **Weekly Dependency Audit** — automated vulnerability scan via GitHub Actions
- **Closed-source detection** — CI verifies no local/ files are accidentally committed
---
@@ -483,22 +499,42 @@ Automated quality assurance:
predix/
├── rdagent/ # Core agent framework
│ ├── app/ # CLI and scenario apps
│ │ └── qlib_rd_loop/ # Quant R&D loop (factor + model generation)
│ ├── components/ # Reusable agent components
│ │ ├── backtesting/ # Backtest engine & protections
│ │ │ ├── backtest_engine.py
│ │ │ ├── vbt_backtest.py # Unified backtest engine
│ │ │ ├── vbt_backtest.py # Unified backtest engine (1-min bars)
│ │ │ ├── verify.py # Runtime backtest invariant checker
│ │ │ ├── results_db.py
│ │ │ └── protections/ # Trading protection system
│ │ └── coder/ # Factor & model coding (CoSTEER + Optuna)
│ │ │ └── protections/ # Trading protection system
│ │ ├── coder/ # Factor & model coding
│ │ │ ├── CoSTEER/ # LLM-based code evolution engine
│ │ │ ├── factor_coder/ # Factor-specific coders
│ │ │ ├── model_coder/ # Model-specific coders
│ │ │ └── kronos_adapter.py # Kronos foundation model adapter
│ │ └── workflow/ # R&D loop workflow
│ ├── core/ # Core abstractions
│ ├── scenarios/ # Domain-specific scenarios
│ ├── oai/ # LLM backend (LiteLLM, streaming, retry)
│ ├── log/ # Logging infrastructure
│ ├── scenarios/ # Domain-specific scenarios (qlib, kaggle, rl)
│ └── utils/ # Utilities
├── test/ # Test suite (134 tests)
│ └── backtesting/ # Backtest unit tests
├── web/ # Web UI frontend
├── scripts/ # Daily operation scripts
│ ├── predix_autopilot.py # 24/7 auto strategy generator
│ ├── predix_gen_strategies_real_bt.py # Parallel strategy generation
│ ├── predix_parallel.py # Multi-instance parallel R&D
│ ├── predix_continuous_strategies.py # Continuous strategy generation
│ ├── predix_fast_rebacktest.py # Fast strategy re-evaluation
│ └── predix_rebacktest_parent.py # Parallel rebacktest orchestrator
├── test/ # Test suite (1,125+ collected)
│ ├── backtesting/ # Backtest engine deep tests
│ ├── qlib/ # Quant loop, factor, model tests
│ ├── oai/ # LLM backend tests
│ ├── log/ # Logger tests
│ ├── local/ # Closed-source tests (autopilot, ML, strategies)
│ └── integration/ # End-to-end pipeline tests
├── data_config.yaml # Walk-forward split configuration
├── pyproject.toml # Project metadata
└── requirements.txt # Dependencies
├── requirements.txt # Dependencies
└── AGENTS.md # Agent configuration & workflow guide
```
---
@@ -584,11 +620,11 @@ The verifier runs in **<1ms** and catches corrupted/missing/flipped metrics befo
### Test suite (CI + pre-commit)
```bash
pytest test/qlib/ -q # 479 tests, 0 failures
pytest test/backtesting/ -q # backtest engine tests
pytest test/ -q # 1,125+ collected, property-based + fuzzing
pytest test/backtesting/ -q # backtest engine deep tests
```
**Coverage**: IC linear invariance, forward-return alignment, cross-implementation validation, ground-truth hand-computed scenarios, look-ahead bias detection, edge cases (all-NaN, constant, zero-variance), Monte Carlo p-value, walk-forward rolling, buy-and-hold equality.
**Coverage**: IC linear invariance, forward-return alignment, cross-implementation validation, ground-truth hand-computed scenarios, look-ahead bias detection, edge cases (all-NaN, constant, zero-variance, 1-bar, empty series), Monte Carlo p-value, walk-forward rolling, buy-and-hold equality, property-based testing (hypothesis: cost monotonicity, signal inversion, max-DD invariants), fuzzing (1,000 random backtest results), autopilot failure recovery, threshold rescaling, API key distribution, ML model acceptance criteria.
---