IPO Market Intelligence
Four frontier models predict IPO outcomes — then get scored against reality
Four leading AIs predict IPO outcomes, then get scored against what actually happened. 800+ IPOs backtested, running live on a public dashboard anyone can check.
800+
IPOs in research universe
4
LLM arena models
19
Technical indicators computed daily
$100/mo
Arena cost cap
See the real product
The problem
IPO research is fragmented across SEC filings, market data vendors, news, and social sentiment — and there's no honest, systematic way to know whether AI model predictions about deal outcomes are actually any good.
What we built
Built a multi-source ingestion pipeline: SEC filings, full price history (D-30 to D+252 per deal), GDELT news sentiment, Reddit archives, and survivorship tracking via SEC Form 25 delistings.
Designed the LLM arena: each of four frontier models scores every IPO with structured output — direction, day-1 and day-30 percentage predictions, confidence, reasoning, and risks — precomputed in scheduled jobs so the public site never touches a model key.
Wrote an event-driven backtest engine treating 800+ IPOs as discrete events with entry/exit rules, realistic slippage and IPO-spread friction, benchmarked against SPY/QQQ with Sharpe, CAGR, and drawdown.
Hardened the public surface: read-only API behind Cloudflare with rate limiting and Turnstile, admin routes IP-allowlisted and key-gated.
Architecture
Outcomes
- ▸Live and public — a working product anyone can open, not a private repo claim
- ▸Model predictions scored against actual market outcomes: a real accountability loop most 'AI finance' projects never build
- ▸Backtests use realistic friction (baseline + IPO spread), avoiding the inflated returns of naive simulations
- ▸Built to answer real questions: how should the SpaceX, OpenAI, and Anthropic IPOs be played when they arrive?
Stack