heranliu / AlphaResearchOS

AlphaResearchOS 
 Agentic Quant Research 
 Turn market data into research hypotheses, predictive models, and auditable strategies. 
 English  |  简体中文 
 Quick start · Benchmark · User guide 
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**
## Features
| 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.
## Quick start
Python 3.11–3.13 · macOS / Linux · Windows through WSL2
 [code] 
Open http://127.0.0.1:8765, then:
1. **Inspect and import your data.** Click **Import CSV / 导入 CSV**, review coverage and validation results, then confirm the import.
2. **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.
3. **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:
 [code] 
Use 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 →
## Benchmark
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