Apppublic

LBJLincoln26/political-llm-trading-floor

sourceHugging Faceupdated 1h agoView on Hugging Face

0likes

App README

Nomos42 Political LLM Trading Floor

10 AI agents trade sector ETFs on daily political signals — every decision is a real LLM API call.

What it does

Each agent receives:

—Daily political events (insider trades, Fed rules, executive orders) from 2026-03-12 to 2026-03-26
—30-day sector trend baselines (avg excessreturn, winrate per sector)
—Agent-specific persona and system prompt

Each agent must decide: long, short, or hold cash on each event's sector ETF. After all 14 days and ~1120 events, we rank agents by final bankroll.

Leakage-safe: agents never see excess_return, y, or outcome — only signal metadata. Resolution: pnl_pct = direction_sign × excess_return × 5.0 (leverage), capped at ±50%.

Agents (10 real LLMs)

Agent	Model	Provider	Persona
Qwen Quant 235B	Qwen 3 235B	Cerebras	Regulatory-delta quant
Qwen Arb 235B	Qwen 3 235B	Cerebras	Cross-sector arbitrage
Llama Contrarian	Llama 3.1 8B	Cerebras	Consensus-fade
Gemini Analytical	Gemini 3 Flash	Google	Fed/SEC stats-first
Gemini Tactical	Gemini 3 Flash	Google	Calendar/schedule
Mistral Large	Mistral Large	Mistral	Ensemble meta-allocator
Mistral Medium	Mistral Medium	Mistral	Portfolio diversification
Mistral Small	Mistral Small	Mistral	Cash when no conviction
Mistral Nemo	Mistral Nemo	Mistral	Exec-order momentum
Ministral 8B	Ministral 8B	Mistral	Game-theory sizing

API Endpoints

Endpoint	Method	Description
`/api/status`	GET	Experiment status, agent bankrolls, running state
`/api/run`	POST	Start/resume experiment (non-blocking)
`/api/stop`	POST	Graceful stop (finishes current day)
`/api/reset`	POST	Clear saved state, start fresh
`/api/mutate`	POST	Change agent params mid-experiment
`/api/logs`	GET	Per-agent decision log stream
`/api/day-decisions`	GET	Day-level allocation details
`/api/leaderboard`	GET	Current standings JSON

Architecture

—TradingAgents (arXiv 2412.20138): structured agent reasoning framework
—Prediction Arena (arXiv 2604.07355): 1-bet-per-agent competition validation
—DMAD (Diverse Multi-Agent Debate): structurally distinct reasoning via different system prompts

Data

Political events schema: {date, ticker, event_type, signal_strength, agency, title, signal_type, signal_sector, donor_info, macro} Top event types: insider_trade (961), fed_rule (149), polymarket (8), exec_order (2) Top sectors: private_prisons, healthcare, energy, finance