A Polymarket question like “Will BTC be up this hour?” is a binary option on spot. Your edge there rarely comes from being right about the close — it comes from being early. The prediction-market book is thinner and slower than a major perp venue, so when real money hits BTC on Binance or Bybit, the spot tape moves first and the YES book follows. That gap is the trade. This playbook reads it with features that are live today, at the cadence they are actually served.
Be honest about cadence first
QUANT_API serves its live features at roughly a one-minute cadence today: each signal publishes about one value per minute. That is the right tool for the up/down binaries that resolve over fifteen minutes, an hour, four hours, or a day — horizons where a mispricing lives for minutes, not milliseconds. The five-minute binary is a different animal: it rewards second-scale reaction, and second-by-second live streaming is on our roadmap, not shipped. So treat m5 as something you research point-in-time today, and the 15m/1h/4h/daily binaries as something you can poll and act on now.
The features that lead the book
Pull these in one request, all addressed as category.signal@window:transform:
flow.ofi@5m:zscore— order-flow imbalance: which side is paying to cross the spread. The earliest read on directional pressure.liquidations.liq_imbalance@15m:zscore— long-vs-short forced selling. A deep flush is non-informational selling that tends to snap back.funding.funding_avg@1h— cross-venue funding: is the crowd still leaning the wrong way, leaving fuel for a squeeze?polymarket.pm_microprice@5mandpolymarket.pm_spread— where the YES book actually sits, and whether it is liquid enough to act on without paying the edge back in slippage.
The play is a comparison, not a prediction. When the spot tape has clearly turned — positive OFI reinforced by a one-sided liquidation print — but pm_microprice has not repriced and pm_spread says there is depth to cross, the binary is mispriced relative to the underlying. Poll once a minute; when spot leads and the book lags, you press the side spot is pointing at.
Why leak-free is non-negotiable here
Binary options resolve on a known outcome, which makes them dangerously easy to backtest dishonestly: it is trivial to let post-resolution information, or a revised liquidation print from five minutes later, leak into the row that “decided.” Every value QUANT_API returns is computed strictly from data with a timestamp at or before your as_of — live or historical — and the same resolver answers both GET /v1/features/live and POST /v1/features/historical. So the feature row you backtested on is the row you would have seen live. There is no separate research path that quietly sees more than production.
Research it before you size it
Take a list of past binaries. For each, build one decision timestamp and POST those timestamps to /v1/features/historical with the feature set above. You get back a point-in-time matrix — every cell resolved as of that row’s instant, leak-free by construction. Feed it to /v1/backtest and read the hit-rate against the 50/50 prior. Crucially, read span_days too: a rule that looks strong over a short window is a curve fit, not an edge.
We are not going to quote you a win rate. We do not invent performance numbers — the entire point is that your own matrix tells you the truth. Leak-free data does not promise a profitable strategy; it promises that your backtest and your live results measure the same thing, so when they diverge you know the cause was your model or the regime, not a timestamp that lied.
Where to start
Wire the five features above, poll the 1h binary once a minute, and watch the lag between a spot lead and the YES book. When you are ready to go faster, the same addresses already resolve at sub-minute windows in the historical endpoint for research — and lower-latency live streaming is coming. Browse the full set in the catalog and the request shapes in the docs.