noahroboros

Autonomous trading strategy optimization using AI autoresearch loops. Inspired by karpathy/autoresearch and Nunchi's 103-experiment run.

An AI agent modifies a trading strategy, backtests it against historical data, scores the result, keeps improvements, reverts regressions, and repeats — autonomously, indefinitely.

Current Best Strategy

RSI(32) momentum with 3-layer exit system — discovered after 88 automated experiments.

Architecture

Entry:  RSI(32) > 50 → Long    |    RSI(32) < 50 → Short

Exit Layer 1:  4% profit target          (locks in gains)
Exit Layer 2:  RSI reversal at 77/23     (catches momentum shifts)
Exit Layer 3:  RSI extreme at 85/15      (safety net)

Backtest Results

BTC / ETH / SOL, 1-hour candles, Sep 2025 — Mar 2026 (6 months)

Metric	Value
Composite Score	2.569
Sharpe Ratio	2.569
Total Return	6.37%
Max Drawdown	1.74%
Win Rate	83.1% (64/77 trades)
Profit Factor	7.3x
Annual Turnover	8.6x

Capital: $100,000 · Position size: 8% per trade · Fees: 5 bps · Slippage: 1 bps

Optimization Journey

Starting from a Nunchi-style 6-signal voting system (score 1.851), the autoresearch agent discovered that simplification wins — removing signals one by one until only RSI(32) remained, then layering precise exit mechanisms on top:

 1.851  Nunchi hybrid baseline (6 signals, 6 toggleable mechanisms)
   ↓    Remove momentum, EMA, MACD, BB — pure RSI outperforms
 2.316  RSI(32) with 80/20 exits
   ↓    Add 4% profit target
 2.418  RSI(32) + profit target
   ↓    Widen RSI exits to 85/15 (PT handles most exits now)
 2.483  RSI(32) + PT + wider exits
   ↓    Add RSI reversal exit at 77/23
 2.569  Current best ✓

88 experiments tested, 85 discarded. Full log in experiments/results.tsv.

Quick Start

# 1. Install Rust
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# 2. Clone and build
git clone https://github.com/upupnoah/noahroboros.git && cd noahroboros
cp .env.example .env
cargo build --release

# 3. Download historical data
cargo run --release -- download

# 4. Run backtest
cargo run --release -- backtest -d data/1h

# 5. Start autoresearch (requires Cursor CLI)
./run.sh -n 100

Downloading Historical Data

Data is fetched from Binance's public klines API (no API key needed for spot). Files are saved to data/{interval}/{SYMBOL}.csv.

# Default: 9 months of 1h BTC/ETH/SOL data
cargo run --release -- download

# Specific symbol and interval
cargo run --release -- download --symbols ETHUSDT --interval 1m

# Custom date range
cargo run --release -- download --symbols ETHUSDT --interval 5m \
  --start 2025-09-01 --end 2026-03-01

# Multiple symbols
cargo run --release -- download --symbols BTCUSDT,ETHUSDT,SOLUSDT --interval 15m

Supported Intervals

Interval	Flag	6-month data size (per asset)	Download time
1 second	--interval 1s	~15M candles / 720 MB	~90 min
1 minute	--interval 1m	~260K candles / 13 MB	~2 min
5 minutes	--interval 5m	~52K candles / 3 MB	~20 sec
15 minutes	--interval 15m	~17K candles / 936 KB	~8 sec
1 hour	--interval 1h	~4.3K candles / 232 KB	~3 sec
4 hours	--interval 4h	~1.1K candles / 56 KB	~2 sec
1 day	--interval 1d	~180 candles / 10 KB	~1 sec

Download Options

Flag	Description	Default
--symbols, -s	Comma-separated trading pairs	BTCUSDT,ETHUSDT,SOLUSDT
--interval, -i	Candle interval	1h
--months, -m	Months of history (from now)	9
--start	Start date (YYYY-MM-DD)	(auto from --months)
--end	End date (YYYY-MM-DD, exclusive)	now
--output, -o	Output base directory	data

Running Backtests

# Backtest on 1-hour data (default)
cargo run --release -- backtest -d data/1h

# Backtest on higher-frequency data
cargo run --release -- backtest -d data/1m

Example output:

Backtesting 13035 candles, capital=$100000, position=8%, fee=5bps, slip=1bps
---
score:              2.569
sharpe:             2.569
total_return_pct:   6.369
max_drawdown_pct:   1.735
num_trades:         77
win_rate_pct:       83.117
profit_factor:      7.295
annual_turnover:    8.6
---

When multiple symbol CSVs exist in the data directory, the engine splits capital equally and runs the strategy independently per asset, then aggregates results.

Autoresearch

The core loop: AI modifies src/strategy/baseline.rs → builds → backtests → keeps improvements / reverts regressions → repeats.

# Run N experiments autonomously
./run.sh -n 100

# Interactive mode
./run.sh

# Cloud: push to Cursor Cloud Agent
./run.sh --cloud

See AGENTS.md for the full agent instructions.

Scoring

Composite score (Nunchi-style, higher = better):

score = sharpe × √(trade_count_factor) − drawdown_penalty − turnover_penalty

Where:

• trade_count_factor = min(num_trades / 50, 1.0) — penalizes fewer than 50 trades
• drawdown_penalty = max(max_dd% − 15, 0) × 0.05 — free below 15% DD
• turnover_penalty = max(annual_turnover − 500, 0) × 0.001 — free below 500x

Hard cutoffs: score = −999 if trades < 10, max DD > 50%, or equity drops below 50%.

Project Structure

AGENTS.md                AI agent instructions (= program.md)
.env.example             Configuration template
run.sh                   Automation script
src/
  main.rs                CLI entry point
  config.rs              Config loader (.env)
  strategy/
    mod.rs               Strategy trait (Signal: Long/Short/Flat/Hold)
    baseline.rs          ** AI modifies this file **
  backtest/mod.rs        Backtest engine (per-bar equity, fees, slippage)
  scoring/mod.rs         Composite scoring (Sharpe, DD, turnover)
  market/
    mod.rs               CSV data loader
    download.rs          Binance klines downloader
  trading/mod.rs         Exchange traits (Binance, Lighter.xyz)
data/
  1h/                    1-hour candles (BTC, ETH, SOL)
experiments/
  results.tsv            Experiment log (88 experiments)

Configuration

Copy .env.example to .env to customize:

cp .env.example .env

Key settings: INITIAL_CAPITAL, POSITION_SIZE_FRAC, FEE_BPS, DOWNLOAD_SYMBOLS, DOWNLOAD_INTERVAL, scoring parameters, exchange API keys. See .env.example for the full list.

License

MIT