Autonomous Multi-Vertical Trading on Kalshi & Robinhood

24/7 Autonomous Trading Agent Across Sports, Weather, Crypto & Macro Markets

App: Axiom Trading Agent | Industry: Quantitative Trading | Client: Independent quantitative retail trader (event-contract markets) | Company size: Solo operator, ~$8.4K live bankroll | Duration: Multi-month iterative R&D, daily production iteration

Summary

An autonomous multi-vertical trading agent that finds, validates, and executes positive-EV trades on Kalshi prediction markets and Robinhood crypto end-to-end, with no human in the loop.

The Challenge

Prediction markets like Kalshi expose hundreds of new contracts daily across wildly different domains (a Yankees game, NYC's high temp, BTC's 15-minute close, the next CPI print). No single human can monitor every market, model fair value, size positions correctly, and execute fast enough to capture edge before it decays.

Our Approach

Outcomes & ROI

Reduced operator intervention from 'restart backend whenever it crashes' to effectively zero. Hard 10%-of-bankroll exposure cap enforced at the API layer prevents any single bug from blowing up the account.

Technologies Used

GPT-4 Swarm, LightGBM, LSTM, Transformers, Platt/Isotonic Calibration, Kalshi REST + WebSocket, Robinhood Crypto API, AccuWeather API, Outlier.bet + Sportradar feeds

Key Takeaways

  1. Stacking specialized per-vertical models with an LLM debate layer outperforms either approach alone for noisy, multi-domain prediction problems
  2. Calibration drift detection (ECE-based throttling) is the unsung hero of unattended trading — it saves capital faster than any single trade decision
  3. Self-healing infrastructure isn't optional for 24/7 autonomous systems; a watchdog process is the cheapest insurance you can buy

Frequently Asked Questions

How does the agent decide which trades to take?

Each cycle pulls every open contract, prices it with a domain-specific model, runs a 6-persona LLM debate, applies a final-veto LLM, then passes the survivor through a Kelly-sized risk gate. Typically 1,000+ signals collapse to 0–3 actual orders.

What happens if a model starts drifting?

Per-segment calibration tracking automatically reduces position sizes (or halts a vertical entirely) when expected calibration error exceeds thresholds. The system self-throttles before bad predictions become losses.

Is it fully autonomous?

Yes — it runs end-to-end with no human in the loop. The watchdog supervisor respawns the backend within 30 seconds of any crash, and a 10% bankroll exposure cap is enforced at the API layer as a final safety net.

Related Services

More Case Studies