Skip to content

finance-trading.md

Latest commit

 

History

History
46 lines (41 loc) · 10.4 KB

finance-trading.md

File metadata and controls

46 lines (41 loc) · 10.4 KB
Raw

Finance / Trading

Use this category for examples where autoresearch is applied to markets, investing, trading agents, or financial decision systems with measurable performance feedback.

Submission format

- [Name](URL) - Industry: one-sentence description of the autoresearch use case.

Entries

  • atlas-gic - Trading: applies Karpathy-style autoresearch to a swarm of market agents, rewriting the worst-performing prompts and keeping changes only when rolling Sharpe improves.
  • autoresearch-trading - Options trading: applies an autoresearch-style keep/revert loop to SPY strategy parameters, logging each experiment against backtest metrics.
  • autoresearch-trading - Trading research: combines Karpathy-style autoresearch with classical optimization so the agent iterates on strategy structure while an optimizer tunes parameters and walk-forward validation decides what survives.
  • binance-trading-karpathy-autoresearch-results - Crypto trading research: publishes results from 4 years of Binance trading data optimized through Karpathy's autoresearch method, with an iterative strategy proposal → backtest → keep/discard loop across multiple assets.
  • BTCautoresearch - Bitcoin forecasting: uses Karpathy-style autoresearch to mutate a single formula file, score walk-forward out-of-sample RMSE, and keep only forecasting rules that beat the baseline power law.
  • autoresearch-skfolio - Portfolio optimization: edits a single portfolio-research script, runs fixed out-of-sample validation across multiple datasets and reversed-return variants, and keeps only Deflated Sharpe Ratio gains.
  • AutoHypothesis - Portfolio research: runs a Karpathy-style autoresearch loop on agent.py, iterating on DEV data and keeping only stock-selection hypotheses that clear one-shot holdback and walk-forward gates before final holdout evaluation.
  • autoresearch-glm - Credit scoring: adapts autoresearch to Taiwan credit-default prediction by editing feature-policy code and keeping only validation AUC gains in a fixed logistic-GLM benchmark.
  • Feature Research at Machine Speed - Credit underwriting: adapts Karpathy's autoresearch to compute_features.py, scoring each iteration by Information Value and ending the round after 25 runs once no new strong features emerged, yielding 58 strong features and a +0.023 AUC lift at a matched 20-feature budget.
  • autoresearch-markets - Prediction-market trading research: adapts Karpathy's single-file keep/revert loop to Kalshi data, editing train.py and optimizing val_logloss on held-out resolved markets.
  • Simmer Autoresearch - Prediction-market trading: lets agents mutate skill configs, measure P&L or edge over live trading cycles or historical replays, and auto-commit only the configurations that improve results.
  • Autonomous Trading Strategy Research - Crypto trading research: adapts Karpathy's single-file autoresearch loop to Hyperliquid perpetual futures, backtesting each strategy.py change on fixed historical data and keeping only score improvements across 103 autonomous experiments.
  • PolyEdge AutoResearch - Prediction-market arbitrage: applies a Karpathy-style keep/discard loop to Polymarket Up/Down paper trading, mutating one strategy parameter at a time and scoring each multi-window run on P&L, fill rate, and trading frequency.
  • AutoResearch — Autonomous DEX Strategy Discovery - DEX trading research: applies Karpathy-style autoresearch to Base DEX strategies, backtesting one mutation at a time against real Uniswap V3 and Aerodrome data and lifting composite score from 0.421 to 8.176 over 230+ experiments.
  • Winning the Paradigm Prediction Market Challenge with Claude Code - Prediction-market market making: uses parallel Claude Code agents as an autoresearch swarm to generate 1,039 strategy variants, run 2,000+ evaluations, and optimize mean edge to a first-place finish in Paradigm's challenge.
  • Autoresearch Trading Strategy Optimizer - Crypto trading research: applies Karpathy's autoresearch to one editable strategy.py, hill-climbing on deterministic historical backtests and keeping only commits that improve final_portfolio_value / max_drawdown.
  • Investing Autoresearch - Trading strategy research: uses an autonomous Claude loop to rewrite strategy.py, backtest on held-out market data, and keep only strategies that improve out-of-sample Sharpe under walk-forward, slippage, and fee validation.
  • EMA Crossover Autoresearch - Equity trading research: adapts Karpathy's three-file autoresearch loop to an SBIN EMA strategy, mutating only strategy.py, backtesting a fixed 10-year Indian equities dataset, and keeping only changes that improve a composite return, Sharpe, and drawdown score.
  • NSE AutoResearch - Equity trading research: adapts Karpathy's single-file keep/revert loop to Indian stock backtesting by mutating only strategy.py and keeping only lower composite-score strategies over a fixed 10-year NSE harness.
  • delu-agent - Live crypto trading: runs a self-improving treasury agent on Base with 5 parallel autoresearch loops that evolve the scoring model through 9,000+ backtested experiments 24/7 and execute trades via Bankr without human intervention.
  • investment-autoresearch - Trading strategy research: a Claude Code plugin that runs parallel agents testing single-hypothesis strategy variants in isolated git worktrees, scoring each against baseline buy-and-hold, and resetting from verified insights when incremental search plateaus.
  • autoresearch-crypto - Crypto trading research: applies Karpathy's single-file, backtest-driven keep/revert loop to crypto strategy development with fixed historical data and metric-based acceptance gates.
  • feat(autoresearch): autonomous experiment engine in Juspay Neurolink - Financial infrastructure research: merges an 8,700-line autonomous experiment engine with Karpathy-style sub-agents and git-backed keep/revert decisions into the Juspay Neurolink product codebase.
  • ml-vs-leadlag-jp-stock - Japanese equity trading: applies Karpathy's autoresearch framework to Japanese stock prediction, comparing ML-driven Sharpe maximization against a classical lead-lag PCA strategy on reproducible benchmarks with published write-up.
  • autoresearch-backtesting - Equity trading: uses an autonomous LLM agent to iteratively rewrite its PyTorch network for SPY directional prediction, backtesting each mutation and keeping only Sharpe improvements in a train → backtest → keep/discard loop.
  • IPL Odds Multi-Agent Autoresearch - Sports betting research: deploys 7 Claude Code subagents to forecast IPL match outcomes, grades each forecast by Brier score, and autonomously mutates reasoning rules across a full season without human intervention.
  • Clio - Prediction-market forecasting: implements autoresearch as a multi-agent, Pareto-scored, time-evolving backtest harness for prediction markets with adversarial validation and bankroll-as-ground-truth.
  • numerai-mcp-autoresearch - Quantitative finance: applies an agentic autoresearch harness to the Numerai classic tournament with XGBoost, Bayesian optimization, time-series cross-validation, and a custom MCP server for autonomous weekly retraining and submission.
  • Credit-autoresearch - Credit risk modeling: pairs an OpenAI coding agent that iteratively generates raw-credit-variable transformations with a deterministic Statsmodels GLM pipeline that evaluates each candidate by validation Gini, applying keep/discard to feature engineering.
  • trading-autoresearch - Intraday trading research: applies Karpathy-style autoresearch to a PatchTST transformer and Kelly-sized policy, iterating overnight and gating changes by Sharpe lower-CI and drawdown floor with detailed per-iteration reports.
  • quanti-autoresearch - Economic research: applies Karpathy-style autoresearch to econometric specification search for GDP per capita growth prediction, iterating on feature selection, estimators, and functional forms against held-out RMSE with structured 4-phase exploration.
  • AutoResearch-based-Trading-Strategy-Generation-and-Testing - Crypto trading research: applies Karpathy-style autoresearch to crypto futures strategy discovery with honest simulation guarantees, an LLM agent writing strategies from a knowledge base of quant techniques, and autonomous keep/discard decisions.
  • Auto-Quant - Crypto trading research: applies Karpathy's autoresearch pattern to FreqTrade strategies on a 5-pair crypto universe, running backtest-driven keep/discard loops that evolved multi-strategy lineups across five versioned runs.
  • noahroboros - Crypto trading research: applies Karpathy-style autoresearch in Rust to BTC/ETH/SOL strategy optimization over 88 backtest experiments, discovering a simplified RSI(32) momentum strategy with a 2.569 composite score.