Hyperquant — Lead Quant Agent (OSS prompt framework)

An open-source prompt framework for an AI quant researcher. You describe a trading strategy idea in plain English; a Lead Quant Agent turns it into runnable strategy code, fetches market data, backtests it with realistic costs, stress-tests it against the ways quant research fools itself, and iterates with you — or tells you honestly that the idea doesn't hold up.

This repo is instructions, not application code — the set of Markdown files that shape how the agent reasons, uses its tools, and avoids the classic pitfalls. Fork it, read it, and make it your own.

---

How it runs

        you (plain-English idea, in chat)
                    │
                    ▼
        ┌──────────────────────────┐     tools:
        │     Lead Quant Agent      │──▶  • web search / read papers
        │     (AGENT.md + docs/)    │     • fetch market & public data
        │                           │     • read / write / run code
        └──────────────────────────┘     • run backtest
                    │
                    ▼
   spec → data → signal → risk → backtest → self-critique → verdict
                    │
                    ▼
        you approve / iterate  →  versioned re-test

AGENT.md is loaded as the agent's system prompt; the agent reads the docs/ files on demand as it reaches each step.

---

The files

File	Role
AGENT.md	The agent's constitution — identity, the build loop, the hard rules, the skeptic discipline. Always loaded; kept lean.
docs/workflow.md	The phase-by-phase playbook: frame → data → signal → risk → backtest → review.
docs/pitfalls.md	The research-integrity checklist — lookahead, overfitting, data snooping, regime sensitivity, cost realism, and more.
docs/validation.md	How to prove an edge is real: pre-registration, OOS/walk-forward, the universal subtests, robustness, overfit tripwires.

Read order for a human: this README → AGENT.md → docs/workflow.md, then the rest as needed.

---

Design philosophy

A disciplined research process, not a strategy generator. The agent runs the full pipeline — framing, data, signal, risk, backtest, review — as a sequence of structured phases, each with a clear output and a gate it must pass. Built in is a mandatory self-critique: the agent attacks its own result, hunting for the flaw, before any of it reaches you.

Research integrity is the product. The goal is not a pretty equity curve — it's the truth about whether a repeatable edge exists. The framework is opinionated about killing bad ideas fast, modeling costs honestly, guarding the out-of-sample holdout, and never moving the goalposts after seeing results. A clean "this doesn't work, here's why" is a successful run.

Lean by design. A small always-on system prompt points to deeper docs that the agent reads only when needed, with a bias toward one well-designed backtest over many hopeful ones.

Model-agnostic. Nothing here depends on a specific LLM. The instructions are principles, not one model's idioms, so the same framework can run on different models.

---

Fork & extend

This is your starter repo — edit it.

• Tune the rules. Sharpen docs/pitfalls.md and the tripwires in docs/validation.md to your markets and risk appetite.
• Add data. Out of the box the agent works from public market data — OHLCV and common indicators; nothing stops you wiring in a richer source (on-chain, derivatives) and teaching the agent to use it.
• Keep AGENT.md lean. It's loaded every turn — put depth in docs/, pointers in AGENT.md.