Hypotheses

Derived by the Research Lead from kapsystem, SuperTrend Bot v6 Delta Exchange: test directional persistence of DOGE versus SHIB rather than outright coin direction. Source: https://www.tradingview.com/script/plslWgCa-SuperTrend-Bot-v6-Delta-Exchange/ (local archived full script read). Trade DOGEUSD.HYPERLIQUID against SHIBUSDT.BYBIT on synchronized 4-HOUR bars. This is an operationalization, not the author's strategy: replace candle Supertrend with an explicitly close-only relative-price version, choose the pair and venues, and replace fixed contract quantity with capped account-relative sizing. Market orders execute at the signal bar close, the registry's equivalent of next-bar-open execution on continuous venues. No independent opposing positions on one instrument, queue advantage or exchange-hosted protection is required. Dollar balance reduces common market exposure but does not guarantee beta neutrality.

Derived by the Research Lead from the abstract of Nico Achtsis et al., A Monte Carlo method for optimal portfolio executions (https://arxiv.org/abs/1312.5919): operationalize the interaction of liquidity, execution pressure and price risk as an event-driven directional hypothesis. Only the locally archived abstract was read; the full paper and its execution algorithm were not read or reproduced. The abstract motivates liquidity-dependent execution costs, not the return prediction proposed here. Test whether two successive same-direction price advances, with the later advance generating substantially greater price change per normalized trading volume despite LOWER volume, identify retreating opposing liquidity and subsequent continuation. This is a bar-derived price-response proxy, not measured signed order flow, identified institutional execution or a causal estimate of impact. Trade four independent 4-HOUR perpetual streams on CRV, COMP, SAND and TRX. Exact local bar directories exist; continuous usable spans and event counts remain unmeasured. Use all available history, admitting each named instrument only after its historical listing and full warm-up; never substitute missing instruments or signals. Market orders execute at the completed signal bar close, the registered faithful equivalent of next-bar-open execution on a continuous venue. Platform-matched intrabar protective stops are supported. No option legs, inverse contracts, queue priority or independent opposing positions are required.

Derived by the Research Lead from Chang, Lizardi and Shah: retain the six-stock daily Hurst selector between SMA momentum and Bollinger reversion, but re-register a smaller fixed allocation. Use ADBE, AMGN, CSCO, INTC, MSFT and QCOM on USEQ, long/cash, 1-DAY split-adjusted bars, without dividend income. These are six explicitly selected case studies, not a historical Nasdaq constituent universe or a reproduction of the paper. Change risk_per_name to 0.0025 of equity; with the retained 8% stop this targets 3.125% of equity per name and 18.75% across six concurrent entries before rounding. These allocations describe entry sizing, not guaranteed losses or continuously maintained weights. Retain all signal definitions and the 15% drawdown limit. Daily-close decisions fill at the next session open; protective stops trigger intrabar in the platform, including gap-through fills at the open. No exchange-hosted offline protection is required. Freeze the revision before collecting new forward evidence; prior optimization, holdout exposures and lineage looks remain counted. Previously exposed periods are development diagnostics only, never a fresh exam. Engineering must restore the bound optimization artifact and repair the contradictory no-report gate before another eligibility decision. The supplied source paper's full text was not provided or independently reviewed here; this revision preserves the existing operationalization rather than asserting the author's exact method.

Derived by the Research Lead from Chung-Han Hsieh, Jow-Ran Chang and Hui Hsiang Chen, Compounding Effects in Leveraged ETFs: Beyond the Volatility Drag Paradigm: test whether trailing positive return persistence identifies better holding periods for daily-reset leveraged ETFs. This is my operationalization, not an executable strategy published by the authors. Use actual TQQQ and SQQQ daily prices and QQQ as signal-only; do not synthesize fund returns by multiplying QQQ returns. Baseline: estimate rolling AR(1) with an intercept on 90 daily QQQ return pairs; require coefficient above 0.05 and follow the sign of its 20-session cumulative return. Hold only the corresponding bull or inverse fund, otherwise cash. All close-based decisions execute at the next session open. Fund expenses, financing and tracking are embedded in observed prices; do not subtract an invented expense rate again. Holdings do not reset daily merely because the funds internally do. The source supplies no causal estimator, entry threshold, exit, or allocation: those choices and a 20-session maximum hold are mine. Use common observed history after both funds actually launched, with warmup; never backfill pre-inception synthetic prices. Report the return-persistence gate against the identical direction/holding/size rules with that gate removed, and against equal-risk buy-and-hold TQQQ. Compare drawdown and notional-weighted expectancy, not just unweighted trade means. No performance pre-study was run.

Original Research Lead hypothesis, prediction-market archetype 1 (crypto signal -> binary pricing). The tradeable unit is the recurring POLYMARKET series eth-updown-4h: a new binary market for every 4-hour window. It resolves UP=1 if ETH's reference price at the window close is above its price at the window open, otherwise 0. The catalog holds about 388 resolved ETH 4H markets (collection began 2026-08-13), each with about 1,670 one-minute probability bars over its listing life. No strategy in the corpus has traded any 4H series. The only Polymarket runs so far are BTC 15m/1h, which died on infrastructure before a backtest (data_unavailable, not evidence about the idea), and SOL 15m fair-value lag, now in backtest_review. This is the same pricing archetype, but on a different horizon, coin and series. The 4H horizon matters for the mechanism: a 15-minute binary is dominated by noise and the spread, while over 4 hours the underlying's partial move carries real information. The question is whether the thinner 4H books reprice to it promptly. MODEL: under a driftless random walk, the fair UP probability at time t inside a window is p_fair = Phi( ln(S_t / S_open) / (sigma_1m * sqrt(m_remaining)) ). S_t is the latest ETHUSDT.BINANCE 1-minute close, S_open is the ETHUSDT.BINANCE 1-minute close at the bar ending at the window's start epoch (the epoch in the market slug), sigma_1m is the standard deviation of 1-minute log returns over the trailing vol_lookback_min minutes, and m_remaining is the number of minutes to the window close. TRADE: buy the UP token when p_fair minus the token's latest 1-minute probability bar close exceeds edge_min, then hold to settlement. A 'Down is underpriced' signal is not traded: the template trades one outcome's series, and shorting is impossible. EXECUTION (registry check): a market buy on the Polymarket bar fills via the platform's fill model; holding to settlement pays the taker fee once and settlement is free. No stops, no next-open requirement, no session venue. All capabilities used are 'supported'. FEE ARITHMETIC: the crypto up/down taker fee is 0.07 x (1 - p) per dollar staked: 3.5% at p=0.50, 2.1% at p=0.70, 4.9% at p=0.30. Add about 0.5% one-way spread (the ~1% round-trip assumption, paid once because we hold to settlement). The expected gross gain per dollar staked is (p_fair - p) / p. A gap of 0.08 at p=0.50 is about +16% expected per dollar against about 4% cost. Allowing the model's error (Chainlink reference vs Binance perp basis, fat tails), I predict a realized average of about +3% per trade, still above cost. The entry gate requires (p_fair - p)/p - 0.07*(1-p) - 0.005 > 0 in addition to edge_min. SAMPLE (lessons L237/L238): events are the markets themselves, non-overlapping 4H windows, one entry per market at most. A 60-day lookback holds about 360 markets. At an expected 30-50% firing rate that is about 110-180 independent trades. CONTROL (lesson L217): the analyst should compare against buying UP at the same elapsed time in every market without the Binance signal. The incremental expectancy attributed to the external signal is the hypothesis, not the raw P&L. CAVEAT: about two months of history is short for nested walk-forward. If the optimizer cannot form windows, I ask that the result be judged on fixed baseline parameters as inconclusive-but-informative rather than retuned.

Original Research Lead hypothesis. It is a cross-venue test of a NEW conditioning variable for the one design family that keeps surviving here (flat-most-of-the-time down-shock absorption with a fixed clock, lessons L224/L230), answering the analyst's 2026-10-18 OKX post-mortem request for 'positioning/liquidation data that marks forced selling directly, rather than BTC return z-scores'. The promoted Binance-spot panel used the OPEN-INTEREST drop (leverage being closed). This hypothesis uses a different and complementary positioning feed: the Binance USD-M perp GLOBAL long/short ACCOUNT ratio (`long_short_ratio` in the Binance metrics mirror, 5-minute, since 2021-12), which counts ACCOUNTS, i.e. the retail crowd, not notional. A large down day on which the share of accounts that are long FALLS means the crowd capitulated: retail longs sold or were stopped out, price-insensitive selling that liquidity providers are paid to absorb (Grossman-Miller 1988; Nagel 2012). A down day on which the account ratio RISES means retail is buying the dip into informed or directional selling; the crowd still has to unwind, so it should not revert (the classic contrarian reading of retail positioning). The trade is placed on HYPERLIQUID perps (13.3% of the corpus vs the >=20% target; cross-venue 10% vs >=15%), with each coin's Binance USD-M perp carried as a SIGNAL-ONLY extra instrument that is never traded: its `global_account_ratio_by_instrument` / metrics feed supplies the account ratio. Hyperliquid 1-DAY candles in our catalog reach back to 2020, so the binding history is the Binance metrics feed (2021-12-02 onward, about 4.9 years). Universe (point in time, fixed list, chosen by rule not by result): the 27 HL perps that also have a Binance USD-M perp with metrics history and an HL 1-DAY history from before 2023-06: LINK, ADA, DOGE, XRP, LTC, DOT, ATOM, ETC, BCH, AVAX, UNI, AAVE, SOL, NEAR, APT, ARB, OP, INJ, SUI, FIL, TRX, BNB, LDO, CRV, RUNE, SNX, STX (names whose HL or metrics series do not exist on a date are simply ineligible on that date: `is_listed()` + data present, never a price-only fallback). Eligible at a bar only if: listed, >= vol_lookback+1 daily HL bars, and an account-ratio observation (> 0) at or before the current close and at or before the previous close, each within 3 hours of that close (point-in-time `supp_as_of`). Pre-registered event count (counted on HL 1-DAY closes and Binance metrics, 2021-12-02 to 2026-10-18, z from a 60-day trailing std of log returns excluding today): z <= -2.0 AND dlog(account ratio) <= -0.05 gives 405 name-events on 97 distinct UTC dates (about 20 dates/year). Per lesson L238 the honest evidence unit is the DATE: 97 dates. Neighbours: drop <= 0 -> 556 events / 121 dates; drop <= -0.08 -> 340 / 83; z <= -2.5 with drop <= -0.05 -> 256 / 59. The L237 floor of ~30 qualifying events is cleared at every neighbour. The comparator subset (z <= -2 with the ratio RISING) has 650 events on 198 dates. Incremental-expectancy pre-registration (lessons L214/L217): the comparator is the SAME down-shock rule on the SAME HL names when the account ratio ROSE. Prediction: capitulation events (ratio drop >= 5%) earn a mean 3-day after-cost return at least 0.4 percentage points higher than ratio-rising down-shocks. If they do not, the positioning condition adds nothing and the premise is falsified even if the panel is profitable. The developer must log, for EVERY down-shock with z <= z_threshold on an eligible name, the dlog ratio, the dlog OI, whether it was traded or skipped (and why: condition, concurrency cap), so the analyst can compute both subsets, and the overlap with the already-promoted OI-flush condition, from one run. Execution: decisions on the close of each HL 1-DAY bar (00:00 UTC). A market order at the signal bar's close is the faithful equivalent of a next-bar-open fill on this 24/7 venue (`next_bar_open_fill` is use_equivalent; substitution named here). No stops. Risk-control route (b): fixed time exit plus capped sizing: 10% of equity per leg, at most 6 concurrent legs (60% gross), leverage 1x, so no single position exceeds 100% of equity. Under the conservative risk profile: a 10% leg in an alt with ~100-130% annualized vol contributes ~10-13% annualized vol (cap 120%); the book stays far from liquidation at 0.6x gross; worst single-trade loss at 10% notional is bounded near 3-5% of equity even on a further -40% leg move (cap 15%). Intended realized vol ~15-25%/yr. Fee arithmetic: HL round trip ~0.09% taker plus ~3 days of hourly funding (after a crash funding is usually near zero or negative, i.e. longs are paid; budget 0.05%) plus modeled slippage: ~0.15-0.20% all-in. A <=-2-sigma day on these alts is typically -6% to -12%. The promoted absorption panels captured 0.9-1.9% per trade over 2-3 days; I expect a gross 3-day capture of about 0.9% on capitulation events, so ~0.7% net, more than 4x the round trip and above the 0.15% HL floor. Buy-and-hold comparison (lesson L224): an equal-weight buy-and-hold of these alts drew down roughly 80-90% in 2022. The panel is flat on roughly 85-90% of name-days with at most 60% gross, so I pre-register a 25% max drawdown (futures default 20% would be flattering for a long-only alt panel that buys crash days in clusters). What distinguishes this from abandoned neighbours: the HL capitulation-close next-day transfer (underpowered) conditioned on range/close-location only and held 1 day; the HL idiosyncratic down-shock rebound and HL volume-confirmed daily panel (promoted) condition on price/volume; the OKX stress-day/beta-overshoot runs (underpowered) condition on BTC z-scores. None conditions on retail account positioning, and none uses a second venue's positioning feed to time an HL trade.

Original Research Lead hypothesis. It is a fresh-venue test (lesson L228) of the forced-selling marker that the OKX stress-day post-mortem asked for and that was then promoted on Binance SPOT ('Deleveraging-Confirmed Spot Capitulation': 212 trades, avg trade 1.86%, Sharpe 0.50, max DD 20.5%). Here the trade is placed on HYPERLIQUID perps (13.3% of the corpus against a >=20% target, and the venue live trading is moving to), and the signal is read on a DIFFERENT venue: the one-day change in the coin's Binance USD-M perpetual open interest in CONTRACT (coin) units. Binance USD-M carries most of the leveraged alt open interest, so a large Binance OI drop on a large down day marks liquidations and margin-call selling across the market; Hyperliquid perps are arbitraged to the same price by cross-venue traders, so the price-insensitive selling and its rebound show up on HL. A down day with RISING OI is a different event (new shorts opening, i.e. directional selling) and is not traded. Why a cross-venue signal: Hyperliquid's own OI history in our store starts 2026-09, which is too short. The platform's per-leg OI loader (`open_interest_by_instrument`) only maps BINANCE / BINANCE_SPOT / BINANCE_CM legs. So every HL trading leg is paired with its Binance USD-M perp as a SIGNAL-ONLY extra instrument (never traded). Its daily bars are subscribed only so the OI series is attached under that leg's id. The strategy reads OI point in time with the OI observation at or before the decision bar's close (latest within 2 hours). If either OI reading is missing it skips the name; it never falls back to a price-only rule. Universe (fixed list, every coin has HL daily bars and Binance USD-M metrics OI from 2021-12): LINK, ADA, XRP, LTC, DOT, ATOM, ETC, BCH, TRX, XLM, FIL, AVAX, UNI, AAVE, SOL, NEAR, ALGO, SAND, DOGE. Each leg is gated by `is_listed()` and needs at least 61 daily bars. The effective window is 2021-12-03 to 2026-10 (OI starts 2021-12). Pre-registered event count (lessons L237/L238/L242; counted on HL 1-DAY closes and Binance metrics OI with a fixed rule, not a strategy backtest): z = today's log return / std of the prior 60 daily log returns. z <= -2 with an OI drop of at least 8% gives 252 name-events on 84 DISTINCT UTC dates (about 17 dates a year), by year 2022: 53/21, 2023: 76/22, 2024: 72/21, 2025: 27/11, 2026: 9/7 (2021: 15/2). The honest evidence unit is the 84 dates. Raw (unsized, pre-cost) 3-day forward HL close-to-close return: mean +2.45%, median +1.17%. Neighbours: z<=-2 with OI drop >=5% gives 359 events on 107 dates, mean +2.27%; z<=-2.5 with OI drop >=8% gives 157 events on 52 dates, mean +1.96%; z<=-1.5 with OI drop >=8% gives 401 events on 150 dates, mean +1.94%. The 2022 bear year was negative (-1.13% mean on 21 dates), so a sustained bear regime is the known weak spot. Incremental-expectancy pre-registration (lessons L214/L217): the comparator is the SAME z<=-2 down-shock with Binance OI RISING that day: 347 events, mean 3-day forward +0.26%, against +2.45% for flush events. Prediction: in the backtest, flush trades beat the rising-OI subset by at least 1.0 percentage point per event. The developer logs every z<=z_threshold down-shock (name, date, OI change, traded or not, 3-day forward return) so the analyst can split the two subsets from the same run. If the flush subset does not beat the rising-OI subset, the OI conditioning adds nothing and the premise is falsified even if the panel is profitable. Execution: decisions on the close of each 1-DAY HL bar (00:00 UTC). On this 24/7 venue a market order at the signal bar's close is the platform's faithful equivalent of a next-bar-open fill (`next_bar_open_fill`: use_equivalent). HL hourly funding paid or received while long is in PnL. Risk-control route (b): a fixed time exit plus capped sizing. Each leg is 10% of equity with at most 6 concurrent legs, so gross exposure is at most 60% and unlevered (leverage 1.0). No stop-loss. Risk profile (conservative): one 10% leg of a ~100%-annual-vol alt contributes about 10-13% annualized vol, far under the 120% per-position cap. The book is flat about 85% of days, so expected realized vol is about 20-30% annualized, well inside the 60% band. At 1x there is no liquidation risk. The worst plausible single trade (-40% on one 10% leg) is a 4% equity loss, under the 15% cap. Fee arithmetic: the HL round trip is ~0.09% taker plus modeled slippage and ~3 days of funding (typically 0.01-0.03% per day on longs, about 0.05-0.10% per trade). Total friction is about 0.2% against a raw measured mean of +2.45% and median of +1.17%. I expect a net capture of about 1.0-1.5% per trade, more than 5x friction and far above the 0.15% HL floor. Buy-and-hold comparison (lesson L224): an equal-weight buy-and-hold of these 19 alts drew down about 80-90% in 2022. This panel is mostly flat with at most 60% gross, so I pre-register a 25% maximum drawdown. Not a duplicate: the in-pipeline HL panels condition on price only (HyperliquidAltPanelCapitulationAbsorption4H, the HL capitulation-close transfer that ended underpowered on one crash cluster) or on a BTC residual (HL idiosyncratic down-shock, BTC-hedged). This is the first HL test whose trigger requires a direct forced-deleveraging marker, and the first that reads that marker on another venue.

Original Research Lead hypothesis. It is a fresh-venue, rule-defined-universe test of the one directional mechanism this factory keeps promoting (lesson L219): directional momentum traded only when a fast and a slow horizon agree, with an explicit flat state when they disagree. The promoted forms are single USD-M names (SOL 12H/3D avg trade 5.1%, ADA 4H/1D 1.17%, NEAR 4H/1D 1.36%, BNB 4H/1D 0.68%). The failures are BTC 4H/1D (no_edge), AVAX 4H/1D (premise_flawed) and the six-name OKX 4H/1D panel with a 3-ATR trail (no_edge: 0.19% average trade but 0.055% per traded dollar, because the fast 4H clock and the tight trail generated $70M of turnover). This design changes three things that the OKX failure points at. (1) It uses the SLOW clock of the best survivor: 12H bars with a 10-day fast horizon and a 60-day slow horizon (the slow sign equals 20 three-day bars, computed from 12H closes so no cross-timeframe alignment is needed). (2) There is no trailing stop. The exit is the signal (disagreement goes flat, agreement in the other direction reverses). The only stop is a wide fixed 4-ATR intrabar protective stop for risk control, so turnover comes from the signal, not stop churn. (3) It trades on Binance COIN-M inverse perps (~0.06% round trip, lowest-fee directional venue; COIN-M is 2.7% of the corpus against a 5% quota). UNIVERSE (point in time, chosen by rule, not by results): every Binance COIN-M <BASE>USD_PERP contract whose bars start on or before 2020-12-31: ADA, BCH, BNB, BTC, DOT, EGLD, EOS, ETC, ETH, FIL, LINK, LTC, TRX, XRP. EOS (last bar 2025-05) and EGLD (last bar 2025-11) were later delisted. They are kept, and the base template's listing calendar closes them at delisting (`point_in_time_universe`: true). Every leg is traded independently. There is no cross-sectional ranking, so the panel is 14 parallel single-name tests under one risk budget. PRE-REGISTRATION MEASUREMENT (disclosed, fixed rule, whole universe, nothing removed): sign(close/close[20]-1) == sign(close/close[120]-1) on 12H closes, position taken at the next bar, signal exit, no stop and no vol scaling, 2020-08 to 2026-10. Result: 4,161 trades pooled, average gross trade +0.50%, median trade -0.96% (a few large trend winners, as expected for trend following), PF 1.22. By name the average trade ranges from -0.37% (LINK, PF 0.87) and +0.01% (EOS) to +1.19% (ETH) and +1.16% (BNB). The weak names stay in. This measurement fixed the expectancy prediction; it did not select names or parameters. EXECUTION SUBSTITUTION: decisions are made at the 12H bar close. On this 24/7 venue a market order at the signal bar's close is the faithful equivalent of a next-bar-open fill. The protective stop is a platform-matched intrabar reduce-only stop, not an exchange-resting order. Inverse contracts are sized with contracts_for_notional() (USD-face contracts: 100 USD for BTC, 10 USD for the alts). Funding is paid or received through the engine's COIN-M funding ledger. Risk profile (conservative): each leg is sized to 4% annualized volatility contribution. That is ~0.04-0.07x equity per alt leg at 60-100% realized vol, giving a gross of about 0.6-1.0x when all 14 legs are active. With pairwise correlation near 0.6 and every leg aligned, the book runs at roughly 40-50% annualized vol, under the 60% ceiling. Leverage stays at or below 1.5x, far from liquidation. Risk control route (a): a fixed 4x ATR(14) stop on a ~0.05x-equity leg loses about 1% of equity at the stop, under the 2% cap.

Original Research Lead hypothesis. It is built on the decomposition of short-term reversal in Da, Liu & Schaumburg (2014, Management Science, 'A Closer Look at the Short-Term Return Reversal') and on Nagel (2012, RFS). Those papers find that reversal comes from the part of a loser's return that is NOT explained by news or fundamentals: price pressure from liquidity demand, which reverts. The firm-specific, information-driven part of a decline does not revert and tends to continue. In crypto alts the cleanest bar-level proxy for 'liquidity-driven' is CO-MOVEMENT. A coin that falls hard in the same week as the whole alt complex is mostly being sold as part of market-wide de-risking (deleveraging, fund outflows, basis-arb unwinds). A coin that falls hard while the complex is flat is mostly repricing its own news (hack, unlock, delisting rumour, team sale). This hypothesis buys only the first kind. Venue and universe. Execution is on OKX_SPOT (0.2% of the corpus, the least-used crypto venue with multi-year daily bars). Long-only and unlevered, so it belongs on spot. The universe is fixed in advance: the 12 OKX USDT spot pairs with daily bars from 2020-2021 in our catalog: LTC, ADA, LINK, XRP, DOGE, ATOM, ETC, BCH (from 2020-01), FIL, SOL, AVAX, DOT (from their 2020-21 OKX listings). A name becomes eligible only after it has 180 daily bars on OKX_SPOT (point in time via listed_since / universe_at). BTC and ETH are excluded from both the traded panel and the market factor. Survivorship caveat, stated up front: all 12 names are still listed today, and OKX keeps no delisted pairs, so coins that died (e.g. LUNA) cannot be in the panel. That residual bias favours a long-the-dip rule. The analyst should read 2022 as the stress check. PRE-STUDY (done before this proposal, fixed rules, OKX_SPOT 1-DAY closes 2020-01 to 2026-10, entry at the signal close, 0.20% round trip deducted, non-overlapping per name). Honesty note: I looked at about 60 cells across correlation-regime, per-name and two-arm variants before settling. The DSR lineage should count roughly 60 trials. Results: (a) Gated rule (coin z7 <= -1.5, market z7 <= -1.0, 7-day hold): 376 trades on 134 distinct entry dates, mean +2.49% net per trade, hit rate 59%, per-entry-date basket t = 2.35. Yearly mean per trade: 2020 +3.8, 2021 +7.6, 2022 -1.9, 2023 +0.3, 2024 +3.6, 2025 +4.4, 2026 +2.0. (b) Same trigger WITHOUT the market gate (the pre-registered comparator, per lesson L217): 414 trades on 182 dates, mean +1.70%, t = 1.26. The gate adds about +0.8% per trade and roughly doubles the t-statistic at this threshold. At a deeper -2.0-sigma coin threshold the gate adds nothing, because nearly every -2-sigma alt week is already a market week. So the claimed increment is specific to moderate drops. (c) The opposite cell (coin <= -1.5 sigma while the market is NOT down) had a negative forward 5-7 day return on OKX spot (-2.5% per trade, 60 dates), which is consistent with the continuation of idiosyncratic news. The short arm is NOT proposed: on the OKX swap panel the same cell was positive, so it does not replicate (lesson L230). (d) Portfolio simulation (10% per name, max 6 concurrent, deepest coin z first, 7-day clock, fees 0.1% per side): +118% over 6.8 years, daily Sharpe 0.63, max drawdown 23.9%, years +6/+37/-10/+2/+15/+30/+9. The ungated version gave Sharpe 0.47 and +76%. The predictions below are set below the pre-study numbers, because the pre-study is in-sample and the selection was made after looking. Execution capability notes: daily bars on a 24/7 venue. 'Next-bar-open' entries are written as a market order at the signal bar's close (the registry's use_equivalent for next_bar_open_fill: on a continuous market that is the next bar's open). No stop orders are used. Risk control is route (b): a fixed time exit plus capped unlevered sizing. Distinct from the capitulation-absorption family in the pipeline (1-day or 4-hour shocks, BTC down-day triggers, OI or volume confirmation). The event here is a 7-day cumulative drawdown. The conditioning variable is the coin's co-movement with the alt market over that week, which separates systematic from idiosyncratic drawdowns. Nothing in the corpus tests that split.

Revision of 'Cross-Venue Deleveraging-Confirmed Down-Shock Absorption on Bybit Perps' after the analyst's pre-optimization backtest review. THIS REVISION FOLLOWS A LOOK AT RESULTS: the iteration-1 development backtest (Dec 2021 to Aug 2025: 261 trades, avg -0.38%, PF 0.83) has been seen, so this lineage's significance hurdle rises and the DSR correction must count that look. What is unchanged (per the analyst: do not change instrument, venue or signal): universe, venues, signal definition and its parameters. Signal: on each UTC 1-DAY close, for each of the 16 Bybit linear perps (LINK, DOT, ADA, XRP, LTC, BCH, ETC, ATOM, AVAX, SOL, DOGE, FIL, UNI, AAVE, NEAR, TRX; USDT.BYBIT), compute z = today's log return / std of the prior vol_lookback=30 daily log returns; a shock is z <= -shock_z (2.0). The OI-flush condition reads the same coin's BINANCE USD-M perp open interest (open_interest_by_instrument via supp_as_of, point in time, staleness <= 6h): log(OI at this day's close / OI at the prior day's close) <= -oi_drop (0.05). require_oi_flush=true is the contract; require_oi_flush=false on identical code is the pre-registered comparator. Binance perps are signal-only legs, Bybit perps are the traded legs. Point-in-time universe: a name is eligible only when both its Bybit and Binance perps have been listed >= 90 days (universe_at / is_listed). WHAT CHANGED - the risk control only. The registered contract carried a 15% intrabar stop that the pre-study (+1.28% close-to-close, no stop) never measured; in the backtest the 42 trades stopped at -15% lost $56.9k against +$43.8k from all other trades, and trades reaching the clock averaged +2.47%. Post-crash days on mid-cap alts have long lower wicks; an intrabar stop placed 15% under the entry sells exactly the forced-liquidation wick the premise says to buy. The revised contract uses platform risk_control route (b): a signal-based time exit plus capped sizing, with no intrabar stop at all. A disaster check is evaluated ONLY on a daily close (not intrabar): if a position's close is >= 30% below its entry, exit at market at that close. With a 2-day hold this check can fire only on the first close after entry. Sizing: 10% of equity per position, max 6 concurrent positions (was 8), so gross exposure <= 60% of equity and no single position exceeds 10% of equity; leverage 1.0. When more than the free slots signal on one date, take the names with the most negative OI log change first (the mechanism), ties broken by the more negative z. Per-cascade loss budget: worst plausible case = 6 names x 10% x 30% disaster loss = 18% of equity in one cascade (a Luna/FTX-style day where every held name closes 30% under entry); a more typical bad cascade (6 names x 10% x -12% two-day loss) is ~7% of equity. Drawdown constraint set at 25% to cover the 2022-type tape honestly (see below). Execution: decisions on the UTC daily close; Bybit is 24/7, so a market order at the signal bar's close is the platform's faithful equivalent of a next-bar-open fill (next_bar_open_fill -> use_equivalent). No stop order is placed, so no intrabar-stop or exchange-resting capability is required. A diagnostic QA should confirm: the iteration-1 stop rate (42 of 261 = 16% of trades hitting -15% within two days) is high; confirm no exit is evaluated against the signal bar's own low. PRE-REGISTERED EXPECTATIONS (written before any run of this revision). (1) Full contract, no stop: the 42 formerly stopped trades now close on the clock; assuming they average about -10% to -12% at the clock (worse than the stop on some, better on most wick recoveries), expected per-trade net = (219 x ~2.3% - 42 x ~11%)/261 - 0.13% cost ~ +0.0% to +0.5% over the development period; I pre-register an expected full-history average net trade of +0.45% (well above the 0.15% Bybit floor but far below the pre-study's +1.28%, which I now regard as optimistic). (2) Incremental expectancy of the OI condition: pre-study claimed +0.69% gross per trade (1.28% vs 0.59%). Pre-register that require_oi_flush=false on identical code yields an average net trade at least 0.35% lower than require_oi_flush=true; if the difference is under ~0.2% the OI condition adds nothing and the analyst should treat the strategy as an unconditioned shock-absorption replication. (3) Clustered-date risk: the iteration-1 run showed 2022 at -2.84% per trade before costs (vs +0.23% in the pre-study). Without the stop I expect 2022 at about -1.0% per trade net (~80 trades x 10% sizing ~ -8% of equity for the year), and a 2022-type bear tape drawdown of 15-22% at these sizes; positive years expected 2023, 2024, 2025, with 2022 and possibly Dec-2021 negative. Distinct entry dates ~105 over ~4.8 years (~22/year); trades ~300 over full history, minus a few from the 6-slot cap.

Original Research Lead hypothesis. It tests forced-deleveraging capitulation on the venue where that forcing is structurally strongest: Binance COIN-M inverse perpetuals. A coin-margined long posts the same coin it is long as collateral. When price falls, the position and the collateral lose value together, so effective leverage rises convexly and liquidations cluster on COIN-M longs during down-shocks. COIN-M open interest is counted in USD-face contracts, so its daily change is a direct, price-orthogonal count of positions closed. A large contract-OI drop on a sharp down day therefore marks liquidation-driven selling that ran ahead of fundamentals. Liquidity providers who buy that flush should earn a short-horizon rebound (Nagel 2012 liquidity-provision returns, conditioned on a direct forced-selling marker, as the analyst asked in the OKX stress-day post-mortem of 2026-10-18). Relation to prior art. The promoted Binance SPOT deleveraging-confirmed capitulation panel (risk_review) conditioned spot alt down-shocks on the USD-M perp's OI drop. This is a separate venue and instrument-class test (lesson L228: a transfer is a fresh test, results reported separately). It trades the inverse contract itself, conditioned on that contract's own COIN-M OI from the BINANCE_CM Vision metrics feed (5-min, from 2021-07). CoinMAltPanelVolumeShockAbsorption1H (in pipeline) conditions on volume on 1H bars; this idea conditions on positioning on 1D bars. Pre-registered fixed-rule measurement (lessons L223/L237/L238), counted once by the Research Lead before writing this, with close-to-close 3-day forward log returns and no costs, for disclosure and the DSR ledger. Across 47 COIN-M USD_PERPs (BTC excluded), the rule z<=-2 AND OI change<=-8% fired 584 times on 109 DISTINCT dates from 2021-07 to 2026-10, mean +0.57% per event. The per-date mean is +0.34% (median +0.47%), and the per-date t is only 0.37 because events cluster: up to 36 names on one day, 29 dates with more than 6. The same z<=-2 down-shock WITHOUT the OI condition fired 1,957 times with mean -0.22%, so the incremental expectancy attributed to the OI-flush marker is about +0.8% per event (the lesson L217/L214 comparison). Year split of the conditioned events: 2021 +2.8% (66 events), 2022 -2.4% (236), 2023 -0.1% (112), 2024 +6.4% (53), 2025 +4.4% (84), 2026 +0.3% (33). The bear year 2022 (LUNA, FTX) is negative, so this is NOT a strong edge and the predictions below are modest. Effective evidence is roughly 109 entry dates, not 584 trades. Fee arithmetic. COIN-M taker is 0.05% per side, so 0.10% round trip, plus roughly 0.05% of funding paid by a long over 3 days (funding is often negative after a flush, which would be a credit). Gross capture is about 0.57% per event, so net is about 0.42% per trade: about 4x the round-trip cost. Execution. Decisions are made on the 1-DAY bar close (00:00 UTC). On this 24/7 venue a market order at the signal bar's close is the faithful equivalent of a next-bar-open fill (capability next_bar_open_fill = use_equivalent). Size inverse contracts in whole USD-face contracts with contracts_for_notional(). Universe, point in time: at each daily close the eligible set is every BINANCE_CM <BASE>USD_PERP in the instruments list that is listed at that bar (universe_at / is_listed, delisted names included until their delisting) and has at least 90 days of daily bars and of OI history. BTCUSD_PERP is excluded (the panel trades alts plus ETH). Risk profile (conservative). Route (b): a fixed time exit plus capped sizing. 8% of equity per leg, at most 6 concurrent legs (48% gross), 1x, no leverage. The intended strategy volatility is about 15-25% annualized, inside the 60% band. A single leg at 8% of a coin with about 100% annual vol contributes about 8% annualized vol, far under the 120% cap. The worst plausible single trade (-50% in 3 days) costs 4% of equity, under the 15% limit.

Derived by the Research Lead from Da, Gurun and Warachka (2014, RFS), 'Frog in the Pan: Continuous Information and Momentum' (https://academic.oup.com/rfs/article-abstract/27/7/2171/1578455; only the published abstract was consulted). The source finding: investors underreact to information that arrives in many small pieces (a 'continuous' path) more than to the same total news arriving in a few large jumps, so gradual winners keep drifting while jump-driven winners do not. The equity result does not establish a crypto edge. This is a fresh-universe, fresh-venue test (lessons L228 and L237): the mechanism was promoted at low confidence on six Hyperliquid perps (HlMatchedWinnerInformationPathSpread: 30 trades, average leg 2.09%, holdout Sharpe 1.17, but pooled OOS t 0.21 on only ~17 pairs). That run is underpowered, and its HL 4H history stops at about 2.3 years. This hypothesis tests the same signal on a DISJOINT 22-name Bybit universe with daily history from 2021, to find out whether the effect exists or the HL result was luck. None of the six HL names (BCH, LTC, ETC, XLM, DOT, ATOM) is in the panel. Universe (point in time): LINK, ADA, XRP, DOGE, TRX, FIL, UNI, AAVE, XTZ, ALGO, VET, THETA, COMP, NEAR, AVAX, SOL, AXS, SAND, CRV, ICP, HBAR and CHZ, all as <SYM>USDT.BYBIT. A name is eligible at a rebalance only when it has a complete 22-day window of daily closes before that rebalance (universe_at / is_listed; point_in_time_universe=true). Residual survivorship: Bybit serves no bars for closed contracts, so the panel is limited to names that are still listed, all listed by Dec 2021. Pre-registered sample size (lesson L238), counted from bar availability and the frozen signal rule, without looking at returns: of the 301 Monday rebalances from 2021-01-18 to 2026-10-19, 188 have a qualifying matched pair. That gives about 188 distinct entry dates and about 376 legs. Execution: decisions use the completed UTC daily bar stamped Monday 00:00 (Sunday's close). Orders are market orders at that bar's close, the supported equivalent of a next-bar-open fill on a 24/7 venue. Risk-control route (b): a fixed 14-day clock exit, 12% of equity per leg, at most 2 overlapping cohorts (48% gross), leverage 1. No intrabar stop is used: the source has none, and the pair is dollar-neutral. Fee arithmetic: one leg costs 0.11% round trip on Bybit taker fees, about 0.13% with modeled impact. The HL run averaged about 2% gross per leg. This test assumes a capture several times smaller, 0.6% per leg, which is still about 4.5x the cost. Funding is paid and received on both legs and largely nets out in a dollar-neutral pair. No preliminary performance test was run.

Original Research Lead hypothesis built on a published finding. Keyrock's study of more than 16,000 token unlocks (https://keyrock.com/from-locked-to-liquidity-what-16000-token-unlocks-teach-us/) reports three things. About 90% of unlocks put downward pressure on price. The decline usually begins about 30 days before the unlock date and intensifies into it. Team and investor allocations are the most negative (team about -25%), while ecosystem-development unlocks are mildly positive. The mechanism is a known, dated supply overhang. Recipients and their counterparties pre-hedge or sell OTC, perp shorts front-run the date, and buyers stand aside until the new float is absorbed. The calendar is public in advance, so the entry decision uses only information that existed at the bar. This is a new mechanism for the factory: the knowledge base contains no unlock or vesting hypothesis, and no strategy has used a supply-schedule feed. Operationalization (all free choices are tunable). Universe, point in time: BYBIT USDT linear perps that have been listed for at least 30 days on the entry date (universe_at()). Events: CLIFF unlocks (not continuous linear emissions) whose recipient category is team/core contributors or private investors, and whose size is >= unlock_min_pct_circ of circulating supply as known on the announcement date. Ecosystem, community and treasury unlocks are excluded because the source finds them neutral to positive. On the 1-DAY close of day T - entry_lead_days (T = unlock date, UTC), short the alt perp and buy BTCUSDT.BYBIT against it, sized as the alt's trailing 60-day beta to BTC times the alt notional. The hedge isolates the supply effect from market beta, following the analyst note on the OKX stress-day run that hedged alt books must separate beta from the effect. Cover both legs at the daily close of T + exit_lag_days. Execution: daily-close market orders. On this 24/7 venue a fill at the signal bar's close is the next bar's open, so the next_bar_open_fill capability is met by its use_equivalent substitution. An intrabar protective stop (platform-matched, reduce-only stop market) sits on the short alt leg at +stop_pct above entry. It triggers inside the bar in the simulated venue and does not require an exchange-resting order. Pre-registered event count (lesson L237/L223 discipline). This is an estimate from public schedules; the Data Scout's feed must confirm it before coding. Monthly team/investor cliffs on Bybit-listed perps include APT (from 2023-11, about 35 events), ARB (from 2024-03, about 30), OP (from 2023-05, about 38), SUI (from 2023-06, about 38), STRK (from 2024-04, about 30), plus quarterly or one-off cliffs on DYDX, IMX, TIA, SEI, WLD, PYTH, JTO, ZK, ENA, ZRO, W, ALT, PIXEL and others. That gives roughly 250-400 qualifying events from 2023-05 to 2026-10, spread over about 120+ distinct entry dates. The analyst should judge evidence by distinct entry dates, not legs. If the feed shows fewer than 60 qualifying events at the baseline threshold, decline the run rather than test. Fee arithmetic. Bybit round trip is about 0.11% per leg; with a beta-1 hedge, the cost is about 0.22% of alt notional per event. Expected gross residual (alt minus beta x BTC) over a T-7 to T+2 window: the source reports declines starting a month out and team unlocks at about -25% over the full window. I assume conservatively that only 1.5-2.5% of residual decline falls inside the final 9 days, and that the strategy captures about 1.2% net after costs per event, about 5x the cost. Per trade (alt and hedge legs counted separately), that is about 0.5% average. Risk (conservative profile). Each event's alt short is 8% of equity, with at most 6 concurrent events (48% gross short). The BTC hedge is capped so that total BTC notional is <= 1.0x equity. Route (a): an 18% stop on an 8% short loses 1.44% of equity, under 2%. A squeeze through the stop gaps, so assume about 3% worst case. Target annualized vol is about 15-20%, with gross leverage <= 1.5x, well inside 5x and the 8-daily-move liquidation buffer.

Original Research Lead hypothesis. A price rise can be paid for by new money (buyers opening longs, so open interest rises) or by shorts closing (forced or voluntary buy-to-cover, so open interest falls). A short-covering rally is buying that ends once the shorts are gone: no new holder stands behind the higher price, so the move should give back some of its gain. This is the up-move counterpart of the factory's down-shock flush-vs-build classifiers ('Who Sold?' spread, OkxOiFlushVsBuildDownShockSpread, Cross-Venue Deleveraging-Confirmed Down-Shock Absorption). Those all classify DOWN days; this trades only UP days, which is a different set of events. Rule: at each UTC daily close, for each of 20 Bybit USDT perps listed before 2022 (LINK ADA DOT ATOM LTC ETC FIL XRP AVAX UNI AAVE NEAR ALGO XLM BCH TRX DOGE SOL SNX CRV), compute the day's log return z = r / std(prior 30 daily returns) from the Bybit bars, and the same coin's Binance USD-M open-interest day change z_oi = dlog(OI) / std(prior 30 daily OI log changes). OI is in coin units, read point-in-time from open_interest_by_instrument for the signal-only BINANCE legs, using the last value at or before each day close. If z >= shock_z (2.0) and z_oi <= oi_z_max (-1.0), short the Bybit alt at 10% of equity and buy BTCUSDT.BYBIT for the same dollar notional. Close both legs on a fixed hold_days (3) clock. Bybit is a 24/7 venue, so the order is a market order at the signal bar's close. That is the platform's stated equivalent of a next-bar-open fill. PRE-REGISTERED MEASUREMENT (lessons L230 and L217; fixed rule, run by the Research Lead before writing this; Bybit closes, Binance OI, 2022-01-15 to 2026-10): 134 qualifying events on 62 distinct dates. BTC-hedged 3-day forward return of the alt: -1.11% (5-day -1.33%). By year: 2022 -1.99% (30 events), 2023 -4.00% (27), 2024 -0.22% (17), 2025 -1.20% (20), 2026 +1.16% (40). So the short earns in 4 of 5 years, and 2026 is the counter-year. Incremental value of the OI classifier over the same up-shock without it: up-shocks with OI roughly flat (|z_oi|<1) returned +0.25% hedged, and those with OI rising >=1 sigma returned +0.39%. So the OI fall is what separates the losing rallies, not the up-shock itself. On the Binance USD-M price series of a 30-name superset the same arm gave -1.36% over 197 events on 84 dates. The OI-rising 'fresh longs' arm was too thin on Bybit (+0.39% hedged against a 0.22% two-leg round trip), so it is deliberately NOT traded. Cost arithmetic: two legs x 0.11% Bybit round trip = 0.22% of the alt notional, against an expected gross capture of about 1.1%, so net ≈ 0.9% per event, about 4x the cost. The analyst should count the 62 distinct dates, not the 134 events (L238). Risk control route (a): each alt short carries an intrabar protective stop at stop_atr_mult (3.0) x ATR(14) above its fill (supported intrabar stop, platform-matched). With daily ATR around 5% that is about 15% x 10% notional = 1.5% of equity, so at most 2% of equity at the stop. When an alt short stops out, its BTC hedge is closed at the same time. At most 5 concurrent events (100% gross maximum: 50% short alts / 50% long BTC). Leverage is at most 1x gross, well inside the conservative profile. An alt whose Binance OI series is missing for the day is skipped, never traded on price alone.

Original Research Lead hypothesis built on the published 'high-volume return premium' (Gervais, Kaniel and Mingelgrin 2001, Journal of Finance 56(3): stocks whose trading volume over a day or week is unusually high relative to their own recent history earn higher returns over the following ~20 trading days, and unusually low-volume stocks earn lower returns; confirmed across 41 countries by Kaniel, Ozoguz and Starks 2012, JFE). The mechanism is the visibility/attention effect (Miller 1977, Merton 1987): a volume shock makes a stock visible to more investors, enlarging its investor base and demand, while a quiet stock drifts out of view. The effect is about VOLUME, not the price move, so the design neutralises the price move by excluding the formation-week return extremes. This is a mechanism the factory has never tested on equities: the only prior 'high-volume return premium' run was a single-asset ETH spot timing rule. It is also not one of the USEQ families that keep dying here (price reversal, OPEX/TOM calendars, cross-asset lead-lag, ETF pairs). UNIVERSE (point in time, survivorship-aware): at each rebalance, the S&P 500 members AS OF THAT DATE (FactoryStrategy.universe_at()/index_member() over sp500_membership.parquet, former members included where bars exist; candidates via `python -m src.data.listings sp500 --start 2006-01-01 --end <run end> --with-data`), that have at least 60 completed sessions of 1-DAY bars and a last close >= $5. SPY.USEQ is the primary bar clock (it sets the rebalance calendar) and is NEVER traded. Residual bias, stated honestly: Yahoo has no bars for some acquired/bankrupt former members, and those names are absent from the cross-section. Backtest window starts 2006-01-01 to keep the bar load tractable (~5,200 sessions x ~650 names). Execution: decisions are taken on the SPY session close and fill at the NEXT session's open on USEQ (next_open_fill, supported). No stops are needed for the risk route (see exit_logic); there are no intrabar or exchange-resting stop requirements. COST ARITHMETIC: USEQ round trip is ~0.02-0.05% (spread + impact, zero commission). GKM report a high-minus-low volume spread of roughly 0.5-1.0% over 20 days in their sample, smaller for large caps and probably smaller after 2006. I predict a 0.5% leg spread per 20-session period. That gives about +0.25% per trade averaged over long and short legs, versus at most 0.05% round-trip cost, so 5x the cost. Turnover is near 100% of each side per rebalance (volume shocks do not persist), so cost per period is about 2 x 0.04% = 0.08% of gross against an expected 0.25% book return per period. PRE-REGISTERED EVIDENCE COUNT (lesson L238: trades on the same date are not independent): about 260 distinct rebalance periods over 2006-2026 (about 210 in development, about 50 in the holdout), each with 30 legs. The unit of evidence is the period, not the ~7,800 leg trades. PRE-REGISTERED COMPARATOR (lesson L217): the incremental claim is that the volume rank adds expectancy over a random long/short pick from the same return-filtered PIT pool. That comparator's expected period return is ~0. The analyst should read the per-period book return and its HAC t-statistic, not the leg count. Separately reported: the long-leg and short-leg mean 20-session market-adjusted returns. Lesson L230 warns that the two sides of an effect can differ, and the literature finds the high-volume (long) side the stronger.

Original Research Lead hypothesis. It is a fresh-universe, fresh-venue test (lesson L219, L228) of the one directional mechanism that has repeatedly been promoted in this factory: momentum with higher-timeframe agreement and an explicit flat-on-disagreement state (SOL 12H/3D, NEAR 4H/1D, ADA 4H/1D and the seven-major 1D/1W basket were promoted with average trades of 0.97-3.5%, while BTC 4H/1D failed with an average trade near zero). None of the six names here has had a confluence test on any venue (prior confluence runs covered BTC, ETH, SOL, BNB, XRP, DOT, AVAX, ADA, LINK, AAVE, ATOM, TRX and NEAR). OKX is under-represented (3.7% of the corpus against a 5% target), and the book is long-short, which the direction quota needs (long-only is 75.6%). Each name is traded INDEPENDENTLY with the same frozen rule inside one strategy, so the panel adds breadth, not a cross-sectional ranking. The instruments are LTCUSDT.OKX (primary), BCHUSDT.OKX, ETCUSDT.OKX, FILUSDT.OKX, XLMUSDT.OKX and UNIUSDT.OKX. Each uses 4-HOUR bars for the signal and execution and 1-DAY bars for the trend confirmation. Bars on all six names start in 2019-12 to 2020-09, about six years of history. DOGE was dropped because its OKX 4H bars only start in 2026. Execution: decisions are taken at the 4H bar close and filled with a market order at that close. On this 24/7 venue that price is the next bar's open (the platform's use_equivalent for next-bar-open fills). The protective stop is an intrabar 3xATR trailing stop matched by the platform's simulated venue, which is the platform's equivalent of a stop and is not hosted on the exchange. Fixed-rule pre-registration measurement: I measured the bare rule on OKX 4H closes (4H momentum over 30 bars, 1D momentum over 20 days, flip/flat on the sign, no trailing stop, 0.10% round-trip charged). LTC came out at -0.15% per trade, BCH +0.14%, ETC +0.40%, FIL +0.42%, XLM +0.26% and UNI +0.07%, with 650-790 trades per name. At 42/30 lookbacks the figures were -0.10%, +0.21%, +0.69%, +0.51%, +0.42% and +0.45%. The long arm carries most of it: the short arm alone averaged about -0.4% to +0.3% net. This crude measurement sets the predictions below. It is not an acceptance test. The ATR trail is expected to cut the losers short and lift the average.

Derived by the Research Lead from "策略：红黑三兵K线形态（裸K）策略回测详解" by 小白菜汤 (fmz) https://www.fmz.com/bbs-topic/9507: From the source: the three-soldiers/three-crows definitions, stop-and-reverse on the opposite pattern, 15m bars, 10% TP, 8% SL and the ladder-TP idea. Chosen by the Research Lead: ETHUSDT.OKX instead of BTCUSDT.BINANCE, numeric wick limits (25%/15% of range), the ladder rule (2.5% steps), 1x equity sizing. Port of the FMZ Quant forum strategy by 小白菜汤 ('策略：红黑三兵K线形态（裸K）策略回测详解', 2022-07-12, https://www.fmz.com/bbs-topic/9507, sections 1, 2 and 4). The source's rule is: a 'red three soldiers' pattern (three consecutive bullish candles, each close above the previous close, each close at or near its high, the third without a long upper wick) closes any short and opens a long; a 'black three crows' pattern (the mirror image with lower wicks) closes any long and opens a short; plus a ladder take-profit, a 10% take-profit and an 8% stop-loss. The author tested it on Binance BTC-USDT 15m bars from 2020-01 to 2022-06 and reports a 77% win rate and 12x return. Those figures are unverified; our backtest with real fees decides. The two attached figures from the post show the pattern on BTC intraday charts (green = up candles on the FMZ chart, prices ~19,400-23,900, July 2022): fig 1 circles three or four rising green candles with tiny upper wicks ending in a large breakout bar after a flat consolidation; fig 2 marks three rising green candles after a pullback low at 22,571, annotated 'three consecutive bull candles, each close above the prior close, the third without a long upper wick', followed by continuation higher. The author notes the pattern works best after a long consolidation. Section 5 (a 4H-trend / 30m-pullback / 5m-breakout discretionary method) gives no exact rules and is NOT part of this test. What I adapted: (1) Instrument and venue: ETHUSDT.OKX USDT perpetual swap (15-MINUTE bars, history 2019-12-25 to now, which covers the author's own 2020-2022 window plus four years out of sample) instead of BTCUSDT.BINANCE. BTC and Binance USD-M are the factory's most over-mined series (Binance is 49% of all experiments), OKX is 3.6%, and the author himself invites testing on other coins. The mechanism is not BTC-specific. (2) The source gives no number for 'close at or near the high'. I chose a baseline of upper wick <= 25% of the bar's high-low range for each of the three bars, and <= 15% for the third bar (the source asks most strongly for the third). Both are tunable. (3) The source names a 'ladder take-profit' (阶梯止盈) without rules. I implement the standard stepped profit lock: each time open profit first reaches a new multiple k of ladder_step_pct (2.5%), the protective stop moves to entry + (k-1) x ladder_step_pct in the trade's favour (so at +5% the stop is at +2.5%, at +7.5% it is at +5%). The fixed 10% take-profit and 8% stop stay as stated. (4) Sizing: the source states none. I use 100% of equity notional at 1x (signal-exit route (b) of the risk policy: a reversal/stop/target exit with a position no larger than equity). The 8% stop loses at most ~8% of equity plus gap slippage, inside the conservative profile's 15% per-trade cap. ETH at 1x is ~75-80% annualized vol, inside the 120% per-position limit. Execution: decisions are taken at the close of the third 15m bar. On a 24/7 venue the 'next bar open' fill is a market order at the signal bar's close (the registry's use_equivalent). Stop, take-profit and ladder stops are intrabar protective stops matched by the platform (supported on every venue), not exchange-resting orders; the source did not ask for exchange hosting. Firing rate (counted on ETHUSDT.OKX 15m 2019-12 to 2026-10 to set the baseline, not to judge edge): at the baseline wick limits the bull pattern fires ~234 times a year and the bear ~182, giving ~180 position flips a year (~1,200 over the history). That is enough for walk-forward windows. The prior Nison three-white-soldiers tests in the corpus (BTC 1D, ADA 4H) died because their strict definition (open inside prior body, long-only, with a regime filter) fired only twice in six years. This port uses the author's looser rule, both directions and 15m bars, so it does not repeat that failure. Fee arithmetic: with ~180 flips/year, the average holding time is about 2 days (shorter when the 8%/10% exits fire). ETH's mean absolute 2-day move is ~4%. OKX round trip is ~0.10% taker. A per-trade edge of 0.25% (min_avg_trade_return_pct) needs only a ~53% directional hit rate on those moves, or a modest win/loss asymmetry from the 10%-target / ladder-lock exits. This is a thin, honest margin: it is 2.5x the round-trip cost, not a large cushion.

Faithful port of the FMZ Quant digest 'Thermostat strategy in practice on the FMZ platform' (恒温器策略在发明者量化平台的实践与应用, by 善, 2019-07-20, https://www.fmz.com/digest-topic/4066, section 'My语言 code'). The source gives complete MyLanguage code; the figures show the strategy arguments (stop-loss price percentage STOPLOSS=1, Bollinger period N=50, Bollinger width N_TMP=2) and an FMZ backtest on Bitfinex BTC_USD 1-hour bars in January 2018 (one-contract sizing; the shown PnL rose then gave back most gains: total +9.9%, max drawdown 27.5%, Sharpe 0.04 - unverified and not relied on). The linked strategy page (fmz.com/strategy/129086) returned HTTP 404 and was NOT read. Mechanism: the Choppy Market Index CMI = |C - C[29]| / (HHV(H,30) - LLV(L,30)) * 100 measures net travel versus range over 30 bars. Below 20 the market is classified as choppy and the system fades short swings with a KD (stochastic) turn inside the oversold/overbought zone; at or above 20 it is trending and the system follows a close outside a 50-bar, 2-std Bollinger band. Exits depend on the regime that was in force when the position was opened. ADAPTATIONS (only venue, instrument, sizing, fees): (1) instrument LINKUSDT.OKX USDT perpetual swap on 1-HOUR bars (same bar size as the source's backtest) instead of Bitfinex BTC_USD, because BTC is the most over-mined series in the corpus (~44-62% share) while OKX is 3.6% of experiments and LINK OKX has 1h history from 2020-03 (~6.5 years); the rules are instrument-agnostic and need both long and short, hence a perp; (2) sizing: the source trades a fixed 1 contract; here each position is a fixed 60% of account equity in notional at 1x (no leverage), chosen so a nearly-always-in-market LINK book (~85% annualized vol at 1x) runs about 50% annualized vol, inside the conservative profile's 60% cap; (3) execution: FMZ fills signals on the bar close; on this 24/7 venue orders are market orders at the signal bar's close (the platform's equivalent of a next-bar-open fill); (4) fees: OKX taker 0.05% per side are charged by the engine. Nothing else is changed: no extra filters, no added stops. Risk-control route (b): signal-based exits (regime-dependent take-profit / conditional stop, and reversal on any opposite entry signal) with capped sizing (60% of equity, never above 100%). Fee arithmetic: LINK 1h bars move ~0.9% per bar; a reversal system on a 50-bar band with a 3% take-profit and 1% conditional stop typically holds ~1-2 days with an absolute swing of 3-6%; if the regime switch adds even a modest directional bias the expected gross capture is ~0.3-0.4% per trade vs a 0.10% round trip (>=3x cost). Expected frequency roughly 3-5 trades a week, ~1,000-1,500 trades over the history.

Original Research Lead hypothesis. It conditions a down-shock liquidity-provision trade on a DIRECT marker of forced selling, as the analyst asked in the OKX stress-day post-mortem (2026-10-18): 'a materially different conditioning variable ... positioning/liquidation data that marks forced selling directly, rather than BTC return z-scores'. The marker is the one-day change in the coin's Binance USD-M perpetual open interest, measured in CONTRACT units (coins), so it is not mechanically linked to the price fall the way notional OI is. A large OI drop on a large down day means leveraged longs were closed: liquidations and margin calls, i.e. price-insensitive selling. Price-insensitive selling is the textbook case where liquidity providers earn a premium (Grossman-Miller 1988; Nagel 2012), and the perp flow is transmitted to SPOT by basis arbitrageurs who buy the perp and sell spot. A down day with RISING OI is a different event: new shorts are opening, which is informed or directional selling, and it should not revert. The trade is placed on BINANCE_SPOT (10.4% of the corpus against a 15% target; long-only and unlevered belongs on spot), using the perp OI only as a signal. There is no perp instrument leg: OI comes from the supplementary feed `open_interest_by_instrument`, which maps each `<SYM>USDT.BINANCE_SPOT` leg to the matching USDT perp's metrics mirror (5-minute Binance Vision OI, downsampled hourly, from 2021-12-01). Universe (point in time): 22 Binance spot alts, all listed on spot before 2021 and with USD-M perp OI history from 2021-12: LINK, ADA, DOGE, XRP, LTC, DOT, ATOM, ETC, BCH, TRX, XLM, FIL, AVAX, UNI, AAVE, SOL, NEAR, ALGO, VET, THETA, SAND, CHZ. A name is eligible at a bar only if it is listed (`is_listed()`), has at least 70 daily bars, and has an OI observation within 2 hours of both the current and the previous daily close; otherwise it is skipped (never a price-only fallback). Bars before 2021-12 produce no trades because there is no OI; the effective test window is about 2021-12 to 2026-10, roughly 4.8 years. Pre-registered event count (counted, not back-tested, on spot 1-DAY closes and metrics OI, 2021-12-03 to 2026-10-17): with z <= -2.0 and an OI drop of at least 8%, there are 310 name-events on 90 distinct UTC dates, about 19 dates per year. Per lesson L238 the evidence unit is the date, not the name-event: 90 dates is the honest sample. Neighbouring settings: z<=-2 with OI drop >=5% gives 419 events on 109 dates; z<=-2.5 with OI drop >=8% gives 193 events on 58 dates. Incremental-expectancy pre-registration (lessons L217 and L214): the comparator is the SAME down-shock rule without the OI condition, which the in-pipeline Binance spot capitulation panels approximate. Prediction: flush events (OI drop >=8%) have a mean 3-day after-cost spot return at least 0.5 percentage points higher than down-shock days where perp OI ROSE. If the flush subset does not beat the rising-OI subset, the OI conditioning adds nothing and the premise is falsified even if the panel is profitable. The developer logs, for every down-shock with z <= z_threshold, the OI change and whether it was traded, so the analyst can compute both subsets from the same run. Execution: decisions on the close of each 1-DAY spot bar (00:00 UTC). A market order at the signal bar's close is the faithful equivalent of a next-bar-open fill on this 24/7 venue (`next_bar_open_fill` is use_equivalent). There are no stops. Risk-control route (b): fixed time exit plus capped sizing (10% of equity per leg, at most 6 legs, so at most 60% gross, unlevered spot). Fee arithmetic: Binance spot round trip is ~0.20% (+ modeled slippage). A <=-2-sigma day on these alts is typically -7% to -12%. The promoted down-shock panels captured 1.0-1.8% per trade over 2-3 days. I expect a gross capture of about 1.3% over 3 days on flush events, so about 1.1% net. That is more than 5x the 0.20% round trip and clears the 0.25% spot floor. Buy-and-hold comparison (lesson L224): an equal-weight buy-and-hold of these 22 spot alts drew down roughly 80-90% in 2022. The panel is flat about 85% of days, with at most 60% gross, so its drawdown should be a fraction of that. I pre-register 25%.

Original Research Lead hypothesis. Mechanism: a sharp daily price drop can come from two different kinds of selling, and open interest tells them apart. If OI falls hard on the down day, longs were closing or being liquidated. That selling is forced and price-insensitive, it is finished once the deleveraging ends, and it leaves a liquidity-provision rebound (the Nagel 2012 'evaporating liquidity' reversal, here identified directly from positioning data rather than from BTC return z-scores). If OI rises on the down day, new shorts opened into the drop. That is discretionary, information-bearing selling, and the drift continues (the OI-confirmed continuation literature). This hypothesis trades the CONTRAST between the two arms on the same date, so the market-wide post-crash beta that sank the three recent OKX down-shock books cancels inside the spread. The 2026-10-18 analyst note on OkxStressDayResidualReversalBtcHedged asked for exactly this: 'positioning/liquidation data that marks forced selling directly'. Difference from BybitOiFlushDownShockAbsorptionPanel (in pipeline): that design is long-only and uses Bybit OI. This one is a dollar-neutral long/short spread whose short arm is defined by OI BUILD, uses the deepest derivatives venue's OI (Binance USD-M) as the signal, and executes on OKX swaps. PRE-STUDY (Research Lead, fixed rule, OKX SPOT daily closes as the price proxy and Binance metrics end-of-day OI, 19 names, 2022-01 to 2026-10). Down-shock z<=-1.0 vs own 30-day vol; flush = OI log-change z<=-1.5 vs its own 30-day std; build = OI change >=0. Results: 120 dates had both arms. The same-date long-flush minus short-build 3-day spread averages +1.38% gross (date-clustered t 2.11). By year: 2022 +3.73%, 2023 +0.37%, 2024 -0.18%, 2025 +1.89%, 2026 YTD -0.96%. A stricter shock (z<=-1.5) gave 60 dates, +2.06%, t 1.97, with all five years positive. The short arm was measured separately on these names (lesson L230): down-shock days with rising OI had a mean 3-day forward return of -0.63% (484 events), against an unconditional -0.28% and +0.78% for the deep-flush long arm. The long arm on its own is weak once clustered by date (+0.17%, t 0.2). The edge is in the spread, not in buying the dip, which is why the design is market-neutral. The incremental value of the OI signal (lesson L217) is pre-registered against the price-only comparator: the same down-shock names unconditioned on OI average -0.34% over 3 days. Event count (lesson L237): every leg has OKX bars and Binance OI from 2021-12, so the overlap is about 4.8 years with roughly 120 qualifying event dates (about 25 a year). Execution: 24/7 venue, so per the capability registry the 'next-day open' entry is a market order at the signal day's close (next_bar_open_fill -> use_equivalent). Binance OI is read point-in-time with supp_as_of / supp_window; the value used is the last 5-minute metrics row KNOWN at the daily close (metrics rows are stamped at the window start, so a row is available one period later).

Original Research Lead hypothesis. It is a dollar-neutral long/short sector spread on two US-listed Select Sector SPDRs, XLF.USEQ (financials) and XLU.USEQ (utilities), signed by a macro feed no strategy in the factory has traded on: the FRED 10-year minus 2-year Treasury spread (series US_2S10S, FRED T10Y2Y, daily from 1976). The platform's macro_supp loader keys each daily value at availability (observation date + 1 day, i.e. after the US close of that date), and the strategy reads it only through supp_as_of('US_2S10S'). Mechanism. Banks borrow short and lend long, so their net interest margin depends on the slope of the yield curve (English, Van den Heuvel & Zakrajsek 2018, J. Monetary Econ.). Utilities are the market's bond-proxy sector: regulated cash flows and high payout. When the curve steepens, especially a bull steepening driven by a falling 2-year yield as the market prices easier policy, the earnings outlook for financials improves relative to utilities. The hypothesis is that sector prices absorb this macro information over weeks rather than at once (slow diffusion of macro and industry information: Hong, Torous & Valkanov 2007, JFE). So a lagged, public curve-slope change predicts the next week's XLF-minus-XLU return. Pre-registration study (Research Lead, pandas, run 2026-10-19 on local USEQ daily bars 1998-12 to 2026-10). Weekly decision at the last session of each week, fill at the next session's open, curve value lagged 2 days (more conservative than the platform's 1-day availability lag), 50/50 legs at 1x gross, no costs. Results: - Baseline (4-week 2s10s change, 3 bp deadband): Sharpe 0.47, about +5.7% a year, 21 of 29 years positive, both halves positive (Sharpe 0.60 before 2012, 0.31 after), 226 position flips (distinct entry dates), +0.70% equity per hold, max drawdown 33.7% at 1x gross. - Neighbourhood (lookback 2/3/4/6 weeks x deadband 0/3/6 bp): Sharpe 0.33 to 0.47 in all 12 cells, so the surface is flat rather than jagged. - Other legs: KRE vs XLU 0.43; XLF vs SPY 0.46, so the effect is specific to financials and not only a utilities short. - Longer signal lag: 5-day lag 0.35, 9-day lag 0.31, a gradual decay. Controls and nulls, pre-registered so the analyst can weigh them: - (i) Price-only control. The sign of the XLF/XLU spread's own 5/10/20-session log return on the same weekly clock gave Sharpe -0.29 / -0.13 / -0.13. The curve signal is therefore not spread momentum in disguise. Incremental expectancy attributed to the external signal (lesson L217) is about +0.6 Sharpe over the price-only arm. - (ii) The 10-year yield LEVEL change has no edge (Sharpe -0.05 to -0.18). - (iii) The 3-month/10-year slope (US_3M10Y) also has NO edge (Sharpe -0.07). This is the honest weak point: the effect sits in the 2-year (policy-expectations) end of the curve, not in the slope as such. A falling 2-year yield alone, used as the sign, gave Sharpe 0.27. - (iv) Correlation of weekly P&L with the promoted NFCI cyclical/defensive spread's signal on the same XLF-XLU pair: 0.04, so the run adds independent information to the portfolio. These numbers come from my own screen of about 20 variants. The analyst should count those looks against the lineage, and the baseline (4 weeks, 3 bp) was the best cell of the grid. Execution (capability registry). USEQ is a session venue. Orders decided on a session's close fill at the next session's open (next_open_fill: supported), with no market-on-close. The week's last trading session comes from the platform session helpers (src.strategies.helpers.sessions / the FactoryStrategy session calendar), never a hard-coded weekday, so holidays and half-days are handled. No stop orders are used. Both legs are traded on USEQ; the curve series is signal-only supplementary data. Cost arithmetic. USEQ round trip is about 0.02-0.05% of notional per leg. A flip closes and reopens two legs (about 0.10% of book notional at worst) about 8 times a year. The expected capture is about 0.70% of leg notional per hold (gross pre-study), well over twice the cost. Risk control route (b): signal-based exit (the weekly signal flip) with capped sizing. Each leg is 35% of equity (0.70 gross, 0 net), no single position above 35% of equity and no leverage. XLF's daily std is about 1.7%, so one leg contributes about 0.35 x 1.7% x sqrt(252) = 9.5% annualized vol, far below the 120% per-position cap. The strategy targets about 10-12% realized annual vol. No stop-loss is added, because a per-leg stop would break the dollar neutrality of a regime spread. The worst single-leg trade in the pre-study at this size stays well under the 15%-of-equity per-trade limit; the analyst should verify this from the trades.