Do Polymarket probabilities lead crypto prices? A free cross-signal sample (BTC/ETH/SOL + funding + OI + PM sentiment) to test it

[manja316]

There's a free companion sample to this dataset that almost nobody knows about, and it answers a question I keep seeing in quant threads: do Polymarket probabilities lead crypto spot prices, or just follow them?

It's a separate Hugging Face drop — crypto-prediction-market-signals (CC-BY-4.0, free, no download wall) — that syncs crypto microstructure with prediction-market sentiment at 15-min intervals. Six parquet tables:

Table	Rows	What
candles	8,800+	15-min OHLCV — BTC, ETH, SOL (Binance Futures)
funding_rates	460+	8-hourly funding + mark price
open_interest	190+	OI in contracts + USD
gold	60+	XAU/USD spot (macro regime)
pm_crypto_signals	6,300+	99 Polymarket crypto markets, YES probabilities
cross_metrics	60+	One joined snapshot: prices + funding + OI + PM sentiment + BTC/Gold + ETH/BTC ratios

Honest caveat: this sample is a static snapshot (last rebuilt 2026-04-23), not the auto-updating feed — it's a sandbox for prototyping the join, not for live trading. The continuously-updated price history (now 18.6M+ snapshots / 18,600+ markets, verified live at api.protodex.io/stats) is the main dataset this repo documents.

Reproduce the lead/lag test yourself (~20 lines, no download)

import pandas as pd

base = "https://huggingface.co/datasets/manja316/crypto-prediction-market-signals/resolve/main"
cm = pd.read_parquet(f"{base}/cross_metrics.parquet")

# Inspect first — column names depend on the build, don't guess them:
print(cm.columns.tolist())
print(cm.head())

# Pick the PM-sentiment column and a BTC price/return column from the print above,
# then test whether PM sentiment at t-1 explains BTC return at t:
cm = cm.sort_values("ts")                      # use the actual timestamp column name
btc_ret = cm["<btc_price_col>"].pct_change()   # fill in from columns above
pm_lag  = cm["<pm_sentiment_col>"].shift(1)

print("lagged corr (PM[t-1] -> BTC ret[t]):", pm_lag.corr(btc_ret))
print("contemporaneous corr:               ", cm["<pm_sentiment_col>"].corr(btc_ret))

If the lagged correlation is meaningfully above the contemporaneous one, sentiment is leading price — that's where the edge would live. With only 60 cross_metrics rows in this sample it's underpowered (treat any number as directional, not significant), but the pipeline is the point: the full price-history dataset has 18.6M snapshots to run it at scale.

Why this exists

I run a public, live-traded crash-recovery bot (polymarket-crash-bot) and needed prediction-market sentiment lined up against crypto microstructure in one frame. Couldn't find it anywhere, so I built and published the join. Free sample is above; the full continuously-archived price dataset is on Hugging Face and Gumroad.

Curious what cross-asset signals others have found — funding→price, OI divergence, BTC/Gold regime flips. Drop findings below.