InvestBot

A local, fake-money trading bake-off. Several approaches each manage their own virtual $100 over the same ~100-name market, so you get an apples-to-apples leaderboard of what actually works — before risking a real dollar on Robinhood.

Paper money only. bot/broker.py: RobinhoodBroker is an intentional stub; nothing here places a real order. Going live is gated — see Going live.

The competitors

Every competitor starts from $100 and trades the same ~100-name market; the dashboard ranks them with full click-through provenance, and the equity chart overlays the S&P 500 as a benchmark.

Competitor	Type	How it works	Cost
momentum_breakout	rule strategy	Buys a 20-day-high breakout confirmed by above-average volume; exits on a close below the 20-day SMA. A few big winners, many small losers.	free
mean_reversion	rule strategy	Buys oversold (RSI(14) < 30), sells overbought (RSI > 70). Trades less, wins more often, smaller edge.	free
blended_momo_rsi	rule strategy	Momentum breakout, but skips entries already overbought (RSI ≥ 70) and exits on a trend break or RSI > 75 — avoids chasing extended moves.	free
Research analyst	deep research	Runs Anthropic's Claude for Financial Services equity-research methodology each tick (screen → sector → comps → catalysts → thesis → portfolio) → target weights with a per-name rationale; grades its prior tick against the S&P 500 and adjusts.	Claude Code plan
llm_voters	LLM swarm	150 cheap LLMs (via OpenRouter), each a unique persona × risk × horizon × quirk profile voting on its own random ~20-name slice — an independent-voter election whose slices keep votes from herding.	~$0.20/run
mirofish_real	social swarm	Persona agents with memory that interact over rounds (a social simulation — the opposite of the independent vote); the book follows their rank-weighted consensus.	OpenRouter (more)
congress_mirror	politician mirror	Ranks members of Congress by the excess return of their disclosed trades (a free GitHub mirror of public STOCK Act filings), then buys what the top performers disclosed purchasing — on the disclosure date, which by law lags their actual trade by up to ~45 days.	free
S&P 500	benchmark	SPY bought all-in on day one and held — the market baseline. Drawn on the chart but never traded by the engine.	free
You	real account	The user's real Robinhood portfolio, rebased to the shared origin so it's comparable. Deposits/withdrawals are stripped via a time-weighted return, so transfers in/out aren't read as P&L. Performance only — a non-clickable line publishing the normalized curve + return/max-DD + a trade count, never the holdings or trades. Real $ stays in a gitignored file.	free

Run it

This project is agent-driven — a Claude Code agent fetches market data (the robinhood-trading MCP), runs the analyst + swarm, and advances the books. The easy button:

say “run the agents” — the run-agents skill: refresh data → analyst report → swarm → advance books → publish web state.

Manual pieces (stdlib-only Python, except httpx for the swarm):

python3 run.py                    # backtest only — prints the leaderboard
python3 tick.py                   # advance the rule strategies one session
python3 run_agents.py             # run the swarm + rebalance the AI agents' books (needs OPENROUTER_API_KEY)
python3 tools/build_dashboard.py  # publish web/public/state.json + history.json (no API calls)

The dashboard

The web/ app is the dashboard — a Next.js app (deployable to Vercel) that fetches the bake-off state (web/public/state.json), live-polls real prices (/api/quotes, Finnhub), pulls headlines (/api/news), and lets you click any ticker for its price chart with each method's buy/sell markers. The equity curves overlay the S&P 500 as a dashed benchmark; the decision trail is colour-coded by method with per-method filtering. Orders placed outside market hours queue and fill at the next open — those resting orders show in the decision trail and each competitor's popup until they fill. Each competitor has a holdings table, and a Stock pool section lists the universe with full company names. See web/README.md. (build_dashboard.py also publishes web/public/history.json for those charts and web/public/news.json, the daily headline cache.)

Standings, curves, and the decision trail are the live forward books (every competitor from $100, same method). Dollar/share figures are shown scaled to a $10,000 notional for readability — the real books are $100; scaling is applied once in build_dashboard.py (per-share prices and % stay real). A Dec–Jun walk-forward backtest is kept as per-strategy reference, not the board.

Risk controls (the “not gambling” part)

• Hard stop per position (STOP_LOSS_PCT, default 15%)
• Max open positions (MAX_POSITIONS, default 5) + a per-name cap for the agents (AGENT_MAX_WEIGHT)
• Equity circuit breaker (CIRCUIT_BREAKER_EQUITY, default $60) halts new buys
• Simulated slippage on every fill so paper results aren't flattering

Going live (later, gated)

Only after a competitor clears a graduation bar (survived a drawdown, enough decisions, tolerable max DD) and the Agentic cash account (••••) is funded do we wire RobinhoodBroker. That account is cash, no options, so this harness is equities/ETFs only. The account number is read from AGENTIC_ACCOUNT in a gitignored .env — never stored in source.

Configuration & secrets

• Knobs live in bot/config.py (universe, risk limits, signal params).
• Secrets go in a gitignored .env: OPENROUTER_API_KEY (swarm) and AGENTIC_ACCOUNT (go-live). The web app reads FINNHUB_API_KEY from web/.env.local (local) or a Vercel env var (deployed).

More

Architecture, conventions, and the agent workflow are in CLAUDE.md.