Inverse-Collateral Forced-Selling Absorption on COIN-M: buy a Binance COIN-M alt inverse perp at the close of a <=-2-sigma UTC day ONLY when its COIN-M premium sank >=1 sigma below the same coin's USD-M premium that day (the selling came from coin-margined longs), fixed 3-day clock (19 pre-2022 USD_PERP alts, 1-DAY, long-only event panel, 10% per leg, max 6 concurrent, flat ~97% per name)

Original Research Lead hypothesis (not seeded). Mechanism: a long position in a coin-margined (inverse) perpetual is collateralised in the coin itself, so when the coin falls the collateral and the position lose value together. Coin-margined longs therefore get liquidated sooner and more violently than USDT-margined longs for the same price move, and their forced selling hits the inverse contract specifically. When that happens, the COIN-M perp trades at an unusual discount to its index COMPARED WITH the USD-M perp of the same coin (whose longs have stable USDT collateral). That relative discount is a direct, contract-specific marker of forced (non-informational) selling — the conditioning variable the analyst asked for in the OKX stress-day post-mortem (2026-10-18: 'positioning/liquidation data that marks forced selling directly, rather than return z-scores'). Forced selling overshoots and is absorbed over the next days, so we provide liquidity: buy the inverse perp at the shock day's close and exit on a fixed 3-day clock. Difference from prior art: the pending 'Inverse-Margin Deleveraging Flush on COIN-M' conditions on COIN-M open interest (which has no stored history here); the promoted Binance-spot deleveraging panel conditions on USD-M OI; the abandoned 'Perp-Led Selloff Absorption' used a single perp-vs-index premium on USD-M 4H bars with a 12h clock. This hypothesis conditions on the CROSS-MARGIN-TYPE premium spread (COIN-M premium minus USD-M premium, both from Binance premiumIndexKlines, 1m history from late 2020) on daily bars and trades the low-fee COIN-M inverse perp (~0.06% RT taker). Fills: market order at the signal bar's close (24/7 venue: equivalent to the next bar's open — platform substitution for next-bar-open, stated here). Pre-registered event count (measured on the stored data before writing, per L237/L238; thresholds were NOT tuned on it): with daily log-return z <= -2 (vs trailing 60-day std, lagged) and spread z <= -1 (vs trailing 90-day mean/std, lagged), the 19-name universe plus ETH produced 177 qualifying events on 80 distinct UTC dates (2021: 23 events, 2022: 24, 2023: 35, 2024: 46, 2025: 29, 2026: 20; mean raw 3-day forward log return +1.56%, median +1.88%, mean 1-day +0.47%); ETH (11 events, mean -0.75%) is excluded from the traded universe as a major, leaving 166 events on at most 80 dates for the 19 names. Year means were negative in 2021-2022 and positive 2023-2026, so the effect is not uniform — the evidence unit is ~80 event dates, not 177 legs. Buy-and-hold comparison (lesson 170/224): the panel is flat ~97% of name-days and holds at most 60% gross, so its drawdown should be a fraction of an equal-weight alt buy-and-hold (which drew down >80% in 2022); report both.

Implements the hypothesis' event rule, universe, sizing, concurrency and fixed clock, with the plan's exact parameter names and fixed values. DATA DEVIATION (stated, not silent): the platform delivers premiumIndexKlines per leg only for USD-M perps (premium_index_by_instrument returns nothing for BINANCE_CM), so the COIN-M premium cannot reach the strategy. The cross-margin-type spread (COIN-M premium minus USD-M premium) is therefore measured from the two perps' own daily bars: log(TP_cm/TP_um) equals premium_cm - premium_um plus the log USD/USDT index ratio, which is common to all coins and slow-moving (absorbed by the 90-day z-score; a USDT depeg day such as May 2022 can still push every coin's spread down at once). The daily typical price stands in for the day's average premium. Consequently min_premium_minutes (plan-fixed at 1000) is carried but inert: there is no per-day minute count to check. shock_z uses r_t / lagged 60-day std (no mean subtraction, 'vs trailing 60-day std'). Capital note: the 19 USD-M reference legs add a second (BINANCE) $100k venue account that is never traded; reported total return is measured on the $200k sum, so it is about half the COIN-M account's own return (Sharpe and trade-level returns are unaffected; drawdown % is correspondingly diluted). point_in_time_universe=True because EOSUSD_PERP (2025-05-21) and ATOMUSD_PERP (2026-03-18) delisted and AAVE/AVAX/NEAR/SOL COIN-M listed in 2021; entries are gated by is_listed. Alignment: a leg contributes only when its newest daily bar is stamped exactly at the primary decision close. Risk control route (b): fixed time exit plus capped sizing (10% per leg, <=60% gross at 1x); worst plausible single-trade loss ~5% of equity, well inside the 15% profile cap.