autoresearch-trading

A framework for building LLM-driven autoresearch trading systems: an immutable walk-forward evaluator with anti-overfit gates, and four swappable extension points — Broker, DataProvider, StrategyBase, CostModel — so you can target your own market, broker, and signal.

It ships with a complete reference implementation: a long-only, delivery-only swing-trading system for Indian equities (NSE) executing through the Dhan HQ Trading API, using only free public data. That reference is the repo itself — strategy.py at the root, scripts/ for operations — built on top of the same autoresearch package you install.

⚠️ Disclaimer — use at your own risk

autoresearch-trading is research/engineering software provided "AS IS", with no warranty, under the MIT License. It is not financial, investment, or tax advice, and its output (backtests, signals, orders) is for educational purposes only. Trading involves substantial risk, including total loss of capital, and past/backtested performance does not predict future results. You use it entirely at your own risk and are solely responsible for your trades, losses, credentials, and legal/regulatory compliance. To the maximum extent permitted by law, the author and contributors accept no liability for any losses or damages arising from its use. Full terms: DISCLAIMER.md. Start in paper mode and read the code before risking real money.

---

The idea

Three layers, decoupled by small interfaces:

1. Research harness (autoresearch.research, autoresearch.backtest) — a walk-forward evaluator with anti-overfit gates (Bonferroni correction, random-walk Monte-Carlo null, sub-period stationarity, parsimony budget, cost-aware Sortino). This is the immutable part you don't edit.
2. Strategy (StrategyBase) — your signal, as a backtrader strategy that changes positions only via order_target_percent.
3. Execution + data (Broker, DataProvider, CostModel) — how orders reach a venue, how prices/universe are sourced, and what trading costs apply.

The autoresearch loop (scripts/loop.py) closes the circle: an LLM reads a journal of past hypotheses, proposes one edit to your strategy.py, runs the evaluator, and KEEP/REVERTs against the gates — so your editable strategy.py and journal.md are your project's files, while the immutable machinery is the installed library.

---

Install

Requires Python ≥ 3.11. With uv:

uv add autoresearch-trading   # or: pip install autoresearch-trading

The distribution is autoresearch-trading; the importable package is autoresearch (e.g. from autoresearch.interfaces import StrategyBase).

To work on the reference system, clone the repo and sync:

git clone <this-repo>
cd autoresearch-trading-india
uv sync --extra dev

---

Start your own project

Not sure where to begin after installing? Just run autoresearch (or autoresearch-trading — same command, matching the install name) with no arguments and it prints exactly what to do next. The fastest way in is to scaffold a project that runs immediately, then edit three files:

autoresearch-trading              # prints "what to do next" (or: autoresearch)
autoresearch init my-quant        # scaffold a project (alias: autoresearch-init)
cd my-quant
python run.py                     # backtests on synthetic data — no setup

You now own strategy.py, provider.py, config.py, and journal.md. Edit them in that order — your signal, your data, your knobs. The library stays installed and untouched. (autoresearch-init is one command in this package, not a separate install.)

Quickstart (no data, no credentials)

Or run the bundled example — a toy backtest on deterministic synthetic data that shows the extension points wired together:

uv run python examples/minimal/run.py

This uses SyntheticDataProvider(DataProvider) + MinimalMomentum(StrategyBase) through the backtest engine. See examples/minimal/.

You can also generate a synthetic prices + universe pair directly:

uv run autoresearch-data            # writes storage/prices.duckdb + universe.duckdb

---

Getting the data (real market data)

All data is free and self-fetched — there is no dataset to buy. The reference system scrapes free public archives into local DuckDB stores. The only credential is a free FRED API key for macro series.

Data	Source (free)	Store
Stock prices	NSE bhav archive (daily ZIPs)	storage/prices.duckdb
Universe	derived point-in-time from price history	storage/universe.duckdb
Macro	FRED (FRED_API_KEY) + yfinance indices (India VIX / Nifty)	storage/macro.duckdb
Fundamentals	yfinance + NSE XBRL filings	storage/fundamentals.duckdb
News	MoneyControl / Pulse RSS / NSE filings / RBI / SEBI	storage/news.duckdb

The data stores are git-ignored and never shipped in the wheel — you regenerate them locally. Two steps:

cp .env.example .env            # add FRED_API_KEY (optional but recommended)

# 1. Fast bootstrap: last ~30 days, enough to start
uv run python -m scripts.bootstrap_ingest

# 2. Optional deep history for backtests (~hours, resumable)
uv run python -m scripts.backfill_5y --all
uv run python -m scripts.backfill_universe

Bringing your own market? Ignore NSE ingestion entirely and implement DataProvider for your source, returning the documented table shape (below).

---

Run the reference system

The India reference lives at the repo root (strategy.py) and in scripts/. It defaults to paper mode (DHAN_MOCK=1), so no broker account is needed to try it.

# Walk-forward backtest of the current strategy.py (+ anti-overfit gates)
uv run autoresearch-eval research
# equivalently: uv run python -m autoresearch.research.prepare research

# One end-to-end paper run for today
uv run python -m scripts.run_live --date $(date +%Y-%m-%d)

# One autoresearch loop iteration (LLM proposes a strategy.py edit, gated)
uv run python -m scripts.loop --iterations 1

The autoresearch loop needs an LLM CLI, not a Python SDK. It drives the model by shelling out to a local command, so there's no LLM package to install. Install and sign in to either Claude Code (the claude CLI) or the OpenAI Codex CLI (the codex CLI) and put it on your PATH — the loop (and the optional LLM classifiers) use whichever is available. Everything else — backtests, the evaluator, paper trading — needs no LLM at all.

The operational commands (run_live, daily_update, premarket_scan, daily_report, loop) live in scripts/ and run from a clone with python -m scripts.<name>. Enabling real live trading requires explicit, deliberate steps (a halt-flag gate + a paper-validation window) — see the code and CLAUDE.md.

---

Build your own

Implement the interfaces in autoresearch.interfaces and point the evaluator at your strategy. The India/Dhan classes are the reference implementations.

from autoresearch.interfaces import Broker, DataProvider, StrategyBase, CostModel

• StrategyBase — subclass it; change positions only via order_target_percent. Put your strategy in a strategy.py at your project root (this is the file the autoresearch loop edits).
• Broker — place_order / get_positions / get_holdings / get_cash / get_fills / …. Reference: DhanBroker, DhanMock.
• DataProvider — read_prices + pit_universe. Reference: NSE bhav ingest.
• CostModel — commission + round_trip_cost. Reference: IndiaCostModel.

DataProvider table-schema contract

The library reads adjusted daily bars in this shape (DuckDB daily_bars in the reference impl):

daily_bars(ticker TEXT, dt DATE, open DOUBLE, high DOUBLE, low DOUBLE,
           close DOUBLE, volume DOUBLE)   -- split/bonus-adjusted close

A point-in-time universe must be derivable from bar history alone (never a "current membership" list — that injects survivorship bias). See examples/minimal/provider.py for a working implementation and autoresearch/interfaces/data_provider.py for the full contract.

---

The research + anti-overfit harness

autoresearch.research.prepare is the immutable evaluator. It walk-forwards over train/validation folds, holds out a sealed test window, and scores each variant against atomic gates — a variant that fails any gate is rejected, not partially accepted. This is the point of the framework: the loop will try to overfit, and the gates are what catch it. Settings that a downstream user changes (capital, currency, universe size, cadence, vol target, dates, storage paths) live in autoresearch.config.Config with the India values as defaults.

---

Repository layout

autoresearch/            # the installable library
  interfaces/            #   Broker, DataProvider, StrategyBase, CostModel (ABCs)
  config.py              #   typed Config (India defaults)
  research/prepare.py    #   immutable walk-forward evaluator + gates
  backtest/              #   engine, metrics, risk, costs
  brokers/ data/ llm/ storage/   # India/Dhan reference implementations
strategy.py              # reference strategy (loop-editable, at repo root)
journal.md               # reference autoresearch memory (at repo root)
scripts/                 # reference operational CLIs
examples/minimal/        # runnable synthetic-data example
tests/                   # test suite

License

MIT.