English ｜ 中文
From idea to audited backtest. One command.
Quick Start · Features · Web Workbench · CLI Usage · Architecture · Roadmap
QuantBench is a local-first AI research workbench for quantitative researchers. Describe a strategy idea in natural language — QuantBench turns it into a reproducible, auditable research run: data sourcing, factor code, backtest, quality checks, charts, and a research note, all saved to a local artifact directory.
It is not an auto-trading system. It is not a chatbot. QuantBench produces research artifacts — the kind you can review, reproduce, and hand to a colleague.
| Tool | Version | Install |
|---|---|---|
| Python | 3.11+ | python.org |
| uv | latest | curl -LsSf https://astral.sh/uv/install.sh | sh |
| Node.js | 22+ | nodejs.org |
| LLM API key | — | DeepSeek-compatible key configured for LiteLLM |
# 1. Install dependencies
uv sync
# 2. Seed example research sessions (free, no API key needed)
uv run python -m quantbench examples seed
# 3. Launch the workbench
uv run python -m quantbench serveOpen the printed URL. Four pre-generated research sessions are ready to explore — complete with metrics, charts, Reviewer reports, and research notes.
First launch auto-detects
uv/node/npm, runsnpm installif needed (~1 min), and prints actionable hints if anything is missing.
Describe what you want to research in natural language. The Coordinator Agent interprets your intent, pulls data, writes factor code, runs the backtest, triggers quality checks, and produces a full research artifact — no boilerplate.
uv run python -m quantbench "在标普500成分股里测试20日动量因子的截面表现，2022-01-01到2024-12-31"Every run is automatically audited. The Reviewer flags lookahead bias, overfitting, cost sensitivity, survivorship bias, capacity constraints, and 20+ other statistical and structural issues. Results carry a verdict: STRONG, PROMISING, WEAK, or REJECT.
Each run saves everything to runs/<run_id>/:
config.yaml # Full experiment configuration
signal.py # The exact factor/signal code that ran
backtest_result.json # Metrics, bootstrap CIs, execution assumptions
review_report.json # Reviewer findings and verdict
equity_curve.png # Interactive equity curve chart
drawdown.png # Drawdown visualization
research_note.md # Auto-generated research summary
manifest.json # Provenance: data hashes, LLM usage, skill injections
Browse, filter, and compare past runs by verdict, asset class, factor family, or Sharpe. Save validated factors to a reusable library. Fork any run to iterate with different parameters while preserving lineage.
- Single-asset and cross-sectional factor backtests
open_t+1default execution (realistic, not optimistic)- Liquidity-aware spread/participation-cap costs, borrow costs, funding costs
- Long/short contribution decomposition, beta/size/sector neutralization
- Decile portfolios, turnover tracking, parameter perturbation, regime decomposition
- US equities: S&P 500 current snapshot and point-in-time constituents
- Crypto: Top-N USDT perpetual contracts via CCXT/Binance
- Explicit survivorship/snapshot bias labeling via trust tiers
Extend QuantBench with MCP servers and workflow skills — paste a mcpServers JSON config, toggle servers on/off, all hot-reloaded without restart. Fully aligned with Claude Code's claude mcp experience.
A React + Vite local workbench for browsing runs, inspecting artifacts, and launching new research.
More screenshots
Research Artifacts — every run produces reproducible code, charts, and a research note:
Customize Panel — manage workflow skills and MCP connectors from the UI:
Experiment Library — filter and compare past research runs:
Zero-dependency hand-drawn SVG charts with hover tooltips: equity curve, drawdown, turnover, decile returns, cost sensitivity, parameter perturbation, regime decomposition, symbol concentration, and returns correlation matrix.
Run a research experiment:
# Cross-sectional factor backtest on S&P 500
uv run python -m quantbench "在标普500成分股里测试20日动量因子的截面表现，2022-01-01到2024-12-31，等权十分位多空组合"
# Crypto perpetual cross-sectional research
uv run python -m quantbench "构建 top 30 USDT 永续合约的截面 universe，测试20日动量因子，2023-01-01到2024-12-31"Experiment Library:
uv run python -m quantbench library list --verdict PROMISING,STRONG --sort sharpe
uv run python -m quantbench compare run_A run_BFactor Library:
uv run python -m quantbench factor save run_A --name momentum_20d
uv run python -m quantbench factor list --family momentum --min-verdict PROMISING
uv run python -m quantbench factor use momentum_20d --param lookback=60 --on "在AAPL上测试，2020-2024"Workflow Skills:
uv run python -m quantbench skill list
uv run python -m quantbench --skill reviewer-weak-triage "我上一个因子被打成 WEAK，帮我看看下一步"MCP Servers:
uv run python -m quantbench mcp add-json filesystem '{"command":"npx","args":["-y","@modelcontextprotocol/server-filesystem","/data"]}'
uv run python -m quantbench mcp list
uv run python -m quantbench mcp enable fetchquantbench/
agent/ Coordinator Agent, LLM wrappers, prompts
api/ FastAPI backend, run state management, artifact serving
artifact/ Per-run artifact storage
data/ Data providers, universe builders, cache, DuckDB warehouse
engine/ Single-asset & cross-sectional backtest engines, metrics
factors/ Factor library entries, parameter extraction, local JSON store
library/ Experiment library index, filtering, comparison, lineage, fork
review/ Reviewer engine, verdict logic, structured reports
skilldocs/ Workflow skill document parsing, matching, prompt injection
skills/ Code execution, plotting, reporting, data quality skills
skills_docs/ Workflow skill Markdown documents
web/ React + Vite local workbench
tests/ CLI, API, data layer, backtest engine tests
Runtime state lives in ~/.quantbench/ by default (override with QUANTBENCH_HOME):
~/.quantbench/
data_cache/ Downloaded market data
runs/ Research run artifacts
factors/ Saved factor library
literature/ Imported papers
api_token Local API authentication
QuantBench is a single-user local research tool. The API binds to 127.0.0.1 with a local token (QUANTBENCH_API_TOKEN or ~/.quantbench/api_token). Do not expose the port to a network. Cross-origin access is restricted to configured localhost origins.
Research quality — expand point-in-time universe coverage, Reviewer stress tests, and statistical guardrail calibration.
Data & execution — more asset classes (futures, macro, alternative data), dataset versioning, optional sandboxed code execution, exportable experiment bundles.
Product — tag governance, per-symbol Factor IC Heatmap, multi-factor risk attribution, multi-session research workflows.
See CHANGELOG.md for recent releases.
QuantBench produces research artifacts, not investment advice. All backtests are subject to survivorship bias, look-ahead bias, data quality limitations, transaction cost assumptions, overfitting, and regime change. Results must be reviewed, reproduced, and stress-tested before informing real decisions.
Copyright © 2026 QuantBench contributors.
This project is licensed under the GNU Affero General Public License v3.0 — you may use, modify, and distribute this software, but any modified version you distribute or deploy as a service must also be released under AGPL-3.0 with full source code.
See LICENSE for the full text.