88 days, 17.84M snapshots in — which of these 4 backtests is actually worth running on prediction-market data?

[manja316]

The collector just crossed 88 continuous days of 15-minute Polymarket snapshots. Live DB this morning: 17.84M price snapshots across 21,402 markets, plus 1.78M order-book (top-of-book) snapshots — Mar 28 → Jun 24. Category mix: ~3k sports, ~2.6k crypto, ~1.4k politics, ~656 geopolitics, rest "other".

(Honest footnote so nobody gets surprised: the downloadable export lags the live DB slightly — the last published file is ~17.26M / 20,585 markets / 85 days. The numbers above are the live collector. The free GitHub sample + schema reflect the published export.)

Prediction markets are an unusually clean backtest surface: every contract has a known terminal value (0 or 1 at resolution), so you get ground truth that equities never give you. Four things I think are worth testing on this, and I want to know which you'd actually run:

1. Calibration — when the market says 70%, does it resolve YES ~70% of the time? (favorite-longshot bias lives here — earlier discussion When Polymarket says 70%, does it happen 70% of the time? Calibration on 19.3M snapshots #5 has a first cut on 19.3M snaps)
2. Mean reversion after sharp moves — a market craters 20% on a headline; how often does it bounce within N hours?
3. Cross-market arbitrage — correlated markets (same event framed two ways) that drift apart
4. Order-book microstructure — does spread/depth widen before a big move?

Questions for anyone who backtests for a living:

• Which of these is worth the time, and which is a known dead end?
• What's the one field you'd need that I probably don't have? (I have prices, top-of-book, volume, liquidity, timestamps, category, resolution flags.)
• Is 15-min granularity workable, or a hard limit for what you do?

Free sample + schema + loader: https://github.com/LuciferForge/polymarket-historical-data