Polymarket 15m Crypto Up/Down — Late-Window Near-Certainty Harvest (Buy the Favored Side When Realized-Vol Says a Sign Flip Is Nearly Impossible in the Remaining Time, Hold to Resolution)

A LONG-ONLY (structurally — prediction markets cannot be shorted) strategy on the recurring POLYMARKET binary series BTC-updown-15m. The tradeable unit is each freshly-minted 15-minute up/down market (resolves 1 if the reference price at close > the reference price at open, else 0). The mechanism is a near-arbitrage grounded in REALIZED VOLATILITY, not a fitted price pattern: late in a given market's life, if the underlying has already moved decisively in one direction, the probability that the remaining minutes reverse the SIGN of the move is small and directly computable from the coin's short-horizon realized volatility. When the market-implied probability of the already-favored outcome is materially BELOW that vol-implied fair probability, we buy the favored token and hold to resolution. Signal source is BINANCE spot 1-minute bars of the same coin (BTCUSDT.BINANCE_SPOT), used to (a) track the running move vs. the market's open reference and (b) estimate the per-minute realized-move standard deviation. This is NOT a directional bet on crypto and NOT an OHLCV momentum/mean-reversion strategy — it is a mispricing-capture on a binary whose fair value we can compute better than the market prices it, using data we already collect. It deliberately fills the single emptiest venue in the corpus (POLYMARKET = 0.0% of 3043 experiments) with a mechanism that has never been tried, and it clears the ~1% round-trip spread with a wide margin by design.

Direct implementation of the hypothesis's realized-vol near-arbitrage. Signal source is BTCUSDT.BINANCE_SPOT 1-minute bars exactly as specified: they supply both the running move versus the market's open reference (p_0 = spot close at the window-open minute, resolved once per market and cached) and the per-minute realized-move standard deviation (rolling 60-bar std of log returns, maintained with O(1) running sums). The decision window is parsed from each market's own slug (BTC-UPDOWN-15M-<open_epoch>-UP), so the 15-minute cadence is derived, not hardcoded. calculate of fair value is Phi(z) on the residual normal move; market_signal returns the continuous fair-minus-market edge on every decidable bar and the entry decision thresholds it. Offline validation on the 477 resolved UP markets in the catalog (2026-08-12 to 2026-08-19) at these defaults produced 109 entries, a 93.6% win rate, average entry price 0.80 and average fair 0.954 (well calibrated), for a mean per-trade return of about +19.8% of premium before the ~1% spread charge. Two deliberate design choices beyond the hypothesis text, both risk controls rather than added edge: (1) a market-agreement gate (token must already trade at or above 0.50), because the only way the model can be badly wrong is a window where Binance spot and Polymarket's own oracle disagree about the SIGN of the move, and requiring the market to agree on direction removes exactly that failure mode while leaving the underconfidence we are harvesting intact; (2) entry_cutoff_frac is pinned to 1.0 because these markets are listed ~24h before their window, so the base template's life-fraction cutoff would reject every correctly-timed entry - the minutes-left window is the real, stricter gate. One structural limitation worth stating plainly: the expansion layer resolves ONE outcome series per spec, so this strategy can only buy the UP token. Down-favored windows (the mirror half of 'buy the favored side') would need a second strategy with outcome: 'Down'; Polymarket cannot be shorted, so they are skipped here rather than faked. Performance: ~480 markets are listed at once and each emits a bar per minute for a full day before its window opens, so on_bar pre-filters a market's bars until its decision window opens (one dict lookup plus two integer comparisons), dropping ~98% of bars before any work and delegating everything that might matter untouched to the base dispatcher.