Research

Whale Signals

I checked whether big bets on Polymarket move the stock market. Mostly they don't.

Why it exists

People on Polymarket are betting real money, so their prices should know things before the news does. If stocks pick that up late, you can measure how late. I couldn't find anyone who had actually checked.

How it works

Signal: YES price moving >7 points inside a rolling 60-minute window
Sample: 604 resolved markets, Jan 2023 to Dec 2024, ≥$100k volume
Dedup: 3,609 raw signals down to 2,623 after collapsing repeats inside 6 hours
Model: Market-model event study, OLS over 60 trading days against SPY
Windows: CAR at 1, 3, 7 and 30 days. CAAR tested with a cross-sectional t-test
Data: Gamma + CLOB APIs, yfinance. No auth, so anyone can rerun it

What broke

The CLOB endpoint's interval=max silently returns empty histories for resolved markets, and that's documented nowhere. For two days I thought every market I wanted was just missing its data. I ended up paginating backward from each resolution date in 13-day windows with explicit startTs and endTs.

The data

Geopolitical is the only category that beats chance at seven days. Everything else sits on it or below it.

Both lines drift down, which is the opposite of what I predicted. Crypto keeps going.

16,731 observations. If signal size mattered this would be a wedge. It's a cloud.

Artefacts

Paper: ~4,600 words, June 2026
Pipeline: detect_signals, sector_map, backtest, stats, visualize

Links