Agentic Quant Research
Turn market data into research hypotheses, predictive models, and auditable strategies.
English | 简体中文
AlphaResearchOS brings the quantitative research loop into one local workbench. Import a standard OHLCV CSV, connect Codex or a compatible model API, and describe the strategy you want to investigate.
Hypothesis → Features + model → Development evaluation → Independent review → Iteration → Holdout
| Research | Review | Analyze |
|---|---|---|
| Generate hypotheses and 1–6 feature expressions | Review each valid candidate in a fresh model session | Compare strategies with equal-weight and momentum baselines |
| Compare rank ensembles, Ridge, and histogram gradient boosting | Inspect parent trajectories, rationale, and follow-up experiments | Explore equity, drawdown, turnover, and cost sensitivity |
| Guide search with development memory and UCB parent selection | Set request, candidate, time, and token-admission budgets | Manage a factor library and export complete research reports |
Choose a dark or light theme from the top bar. The workbench follows your system on first use and remembers your choice.
Dark and light theme preview
Interface preview with synthetic demonstration data.
Python 3.11–3.13 · macOS / Linux · Windows through WSL2
uv sync --locked
uv run alphaos serveOpen http://127.0.0.1:8765, then:
- Inspect and import your data. Click Import CSV / 导入 CSV, review coverage and validation results, then confirm the import.
- Connect a model. Choose your authenticated Codex CLI, a DeepSeek/Gemini/Ollama API preset, or a custom compatible endpoint. Enter a model ID, save, and test its structured response.
- Start research. Select the imported dataset, enter a direction and budget, and follow candidates as they are proposed, trained, reviewed, and evaluated.
Your CSV uses these columns:
date,symbol,open,high,low,close,volumeUse UTF-8, 3–100 assets, aligned daily observations, and consistently adjusted prices. Research with the default settings needs at least 468 common trading days; the importer accepts datasets from 400 days. Each date–asset pair must be unique. See the data guide.
New research defaults to 6 candidates, 12 model requests, and 900 seconds. The workbench keeps the features, fitted-model settings, review decisions, and development results together. Model connections and controls →
On nine US sector ETFs, AlphaResearchOS's selected two-feature Ridge strategy achieved 117.53% total return, 26.22% CAGR, and a 1.5608 Sharpe ratio over 2023-05-11 to 2026-09-17, after 10 bps one-way transaction costs.
The dataset spans 2010-01-04 to 2026-09-17. Model training and candidate selection use history through 2023-05-08; the table and charts below cover the subsequent 841-session holdout. Strategies rebalance every five sessions. AlphaResearchOS and the momentum baselines hold the top three assets; equal weight holds all nine.
| Strategy | Total return | CAGR | Sharpe | Max drawdown | Annual volatility |
|---|---|---|---|---|---|
| AlphaResearchOS · Ridge | 117.53% | 26.22% | 1.5608 | -16.21% | 15.71% |
| Equal weight | 63.22% | 15.81% | 1.2662 | -15.61% | 12.18% |
| 20-day momentum | 58.22% | 14.74% | 1.0615 | -21.21% | 13.86% |
| 60-day momentum | 65.63% | 16.32% | 1.1823 | -18.14% | 13.57% |
The return lead was +54.32 pp over equal weight, +59.31 pp over 20-day momentum, and +51.90 pp over 60-day momentum. Sharpe was higher than all three baselines, alongside higher volatility; maximum drawdown was 0.60 pp deeper than equal weight.
Across 39 complete months, the strategy beat equal weight in 28 months (71.8%), 20-day momentum in 22 (56.4%), and 60-day momentum in 23 (59.0%). It led all three in 2024, 2025 and 2026 through September 17, but trailed them in the partial 2023 window. The chart preserves every year and the full drawdown path.
Implementation: costs and turnover
At 40 bps one-way costs, the strategy returned 92.15%, compared with 61.15% for cost-matched equal weight. At the 10 bps benchmark setting, average daily turnover was 4.92%: 62.50% lower than 20-day momentum and 37.39% lower than 60-day momentum, but above equal weight's 0.51%.
Methods, development results, and chart reproduction · Immutable benchmark data · Derived diagnostics
Features run through a bounded OHLCV expression language. Predictive training uses matured labels, with model fitting fixed before each development fold and the final holdout. A separate model session reviews development evidence and suggests the next experiment. Approved candidates compete by development score before the final strategy is evaluated.
The workbench supports executable window, field, and turnover constraints, plus pause/resume and a searchable factor library. Research methods →
uv sync --locked --extra dev
uv run pytest -q
uv run ruff check src/alpharesearchos scripts
node --check src/alpharesearchos/static/app.js
uv buildContributing · Roadmap · Changelog
MIT. Third-party attribution is maintained in NOTICE, with numerical component details in UPSTREAM.md.