AutoResearch Strategic Stock Orchestration

AutoResearch Strategic Stock Orchestration is a self-improving multi-agent trading system inspired by Andrej Karpathy's nanochat methodology. Three trading strategies (momentum, mean-reversion, buy-and-hold) compete for capital from a meta-allocator. An overnight AutoResearch loop sends the source code to the Claude API and asks for one targeted improvement per iteration as a unified diff. The loop applies the diff, runs a backtest on held-out validation data, and keeps the change only if validation Sharpe improves by at least 0.02. Otherwise it rolls back with git reset. Every experiment is logged to JSON; every accepted change becomes a git commit with a structured message. Across 105 unattended iterations on Lightning AI, the agent produced five accepted improvements that compounded validation Sharpe from a 1.0911 equal-weight baseline to 3.6260. The agent rediscovered classical quantitative techniques from first principles by reading source code alone: inverse-volatility weighting, momentum-of-strategies allocation, EWMA-based scoring, regime-conditional weights, and a shorter momentum lookback. Held-out final-test Sharpe came in at 6.5319, which we discuss honestly as a combination of generalization, sample-window luck, and adaptive-allocator effects. The full audit trail (every proposed diff, kept or rejected) is committed to git and visualized in a Vercel-deployed Next.js dashboard. The system demonstrates a methodology rather than claiming alpha: small codebase (1,145 lines total), measurable metric (validation Sharpe), and self-improvement driven by readable git diffs that any reviewer can inspect.