Hypotheses

Test whether releasing a binding Hyperliquid open-interest constraint allows a large venue-specific ICP price dislocation to unwind. Trade synchronized 4-HOUR ICP perpetual bars on Hyperliquid and Bybit, with equal base quantities and opposing sides. This is an original event-conditioned convergence hypothesis, not funding carry or ordinary continuous spread reversion. The documented endpoint reports currently capped coins, but does not establish predictive returns: [Hyperliquid perpetual API](https://hyperliquid.gitbook.io/hyperliquid-docs/for-developers/api/info-endpoint/perpetuals). Market orders execute at completed signal-bar closes, the registry's faithful equivalent of next-bar-open execution on continuous venues. Both instruments support netted positions; separate opposing legs require neither hedge mode nor atomic execution. Historical testing is conditional on acquiring complete successful cap snapshots: local hl_oi_cap storage records positive membership only, and its collector conflates failed requests with empty responses. Do not infer a release from missing rows. Use all common audited history once the required event data exist; do not reconstruct historical caps from prices or current snapshots.

Test whether a large relative repricing between payment-network tokens partially unwinds after its first observable reversal. Trade synchronized 4-HOUR XRPUSD.HYPERLIQUID and XLMUSDT.OKX bars, shorting the relative winner and buying the laggard. This is cross-asset relative value, not guaranteed arbitrage. Market orders execute at the signal bar close, the supported equivalent to next-bar-open execution on continuous venues. Individual protective stops trigger intrabar through platform matching. Include actual funding and execution costs on both legs. All capture and firing-rate estimates are unmeasured research predictions.

Test whether persistently expensive one-sided perpetual positioning that cannot extend price unwinds after price breaks against the funding-paying side. Trade three independent Hyperliquid perpetuals using 12-HOUR OHLCV bars and settled hourly funding. This is a directional positioning hypothesis, not funding carry. Market orders execute at the completed signal-bar close, explicitly using the registry's continuous-market equivalent for next-bar-open execution. Protective stop-market orders trigger intrabar through platform matching. Use each instrument's actual overlapping bar/funding history. Predict 10–25 qualifying episodes per instrument-year; this is unmeasured. No preliminary performance test was run.

Retain the SOLUSD.HYPERLIQUID 4-HOUR activity-surge contrarian signal and fixed 72-hour exit, but reduce position_fraction from 0.90 to 0.45. Freeze the completed optimizer's surge_ratio_min at 2.0380667881633885. Risk control route (b): time-based exit, leverage 1.0, maximum entry exposure 45% of equity, one net position and no pyramiding. Market orders execute at the signal bar close. Retain the enforced 20% drawdown limit. This revises the excessive-risk premise; it does not establish fresh predictive evidence or authorize paper deployment.

Replace unconditional beta amplification with a conditional hypothesis: trade SOL in BTC's trend direction only when trailing paired returns show positive, stable coupling. Use synchronized 4-HOUR bars; BTC is signal-only. Freeze the previously selected trend parameters and reduce risk_pct from 0.02 to 0.015 and max_notional_frac from 0.60 to 0.45. Market orders execute at the signal bar close, the supported continuous-market equivalent of next-bar-open execution. Protective stops trigger intrabar in the platform. Preserve all prior attempts and holdout exposure in lineage accounting; historical results remain development evidence, and validation requires genuinely new forward observations.

Derived by the Research Lead from the supplied digest of 小草's FMZ article: test hourly high-activity continuation with modest exposure. Implement the analyst's specific sizing revision while preserving the registered long-only contract. Change position_fraction from 1.4910757559263004 to 0.20; retain fill_buffer=0.08, giving reference entry notional/equity of 0.20/1.08=18.5185%. Use risk-control route (b): a four-hour time exit with capped unlevered sizing. Market orders execute at signal-bar close, the capability registry's equivalent to next-bar-open execution on a continuous venue. CLV is a price-location proxy, not observed aggressor flow. Reduced sizing does not establish positive OOS expectancy or compliant drawdown. Preserve all prior searches, result looks, OOS windows and holdout exposures.

Long SANDUSD.HYPERLIQUID and short an equal base quantity of SANDUSDT.BYBIT using synchronized 4-HOUR bars, holding up to 60 days. Test persistent venue-specific funding differences rather than directional forecasting or price-gap convergence. Execute market orders at signal-bar closes, the registry's equivalent of next-bar-open execution on continuous venues. The funding snapshot contains 678 overlapping days from 2024-12-04 through 2026-10-12, with an unconditional Bybit-minus-Hyperliquid daily funding difference of +0.009197%. This descriptive statistic is not a strategy backtest and does not establish profitable conditional carry. Hyperliquid settles hourly; negative funding pays longs ([funding documentation](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/funding)). Use the current runner's funding-accrual ledger exactly once; its implementation supersedes the stale data-specification warning that funding is not credited. Risk control route (b): capped unlevered sizing with funding, basis and time exits. Fee sources: [Hyperliquid base-tier fees](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/fees) and [Bybit VIP-0 derivatives fees](https://www.bybit.global/en/help-center/article/Trading-Fee-Structure).

Test whether unusually unstable Hyperliquid funding relative to the same coin on Binance predicts weaker subsequent returns. Trade INJ and ICP Hyperliquid perpetuals; matching Binance perpetuals supply funding references only. Use synchronized daily bars and weekly decisions. Market orders execute at signal-bar closes, the registry's equivalent of next-bar-open execution on continuous markets. Protective stops trigger intrabar through platform matching. Official documentation establishes settlement mechanics and fees, not this predictive hypothesis: [Funding](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/funding), [Fees](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/fees). Coverage checks found Hyperliquid INJ funding from 2023-05-12 and ICP from 2025-11-08, with older Binance reference histories. Evaluate all joint available history after eligibility and warm-up.

Derived by the Research Lead from Imran Lakha's Deribit Insights article: test whether ETH volatility accelerating relative to BTC creates underpriced ETH convexity. The article discusses ETH/BTC option-volatility spreads in relation to realized spreads; it supplies no executable rules or profitability evidence. Source: [Options Dynamics: Reading the Signals in BTC and ETH](https://insights.deribit.com/industry/options-dynamics-reading-the-signals-in-btc-and-eth/). Buy an ETH weekly reverse iron condor: long the nearer-ATM call and put, short further-OTM call and put, all at one expiry. Use ETH-PERPETUAL.DERIBIT's own completed 4-HOUR bars as the untraded primary clock and BTCUSD.HYPERLIQUID synchronized 4-HOUR bars as an untraded reference. Extend RollingOptionsStrategy directly and use its historical ETH chain expansion; distinguish same-kind legs by strike, rather than a single C/P side map. Persist every resolved concrete option instrument ID in extra_instruments and the hypothesis execution manifest. The listed instruments are fixed clock/reference instruments; the options_universe specification below identifies the required dynamically resolved traded legs. No fixed expired option is a trading anchor. Ordinary continuous-market orders execute at signal-bar close, the registry's equivalent of next-bar-open execution. No atomic combo, independent opposing positions in one instrument, selectable same-bar policy, exchange-hosted protection, or inverse-perp sizing is required. Request up to 730 days and report actual synchronized chain coverage; historical underlying prices do not extend the option evidence span. Predict 15-25 qualifying weekly structures per covered year, not a daily scalping strategy.

Derived by the Research Lead from John Ehlers' Super Bandpass trading rules published by ProRealCode: https://www.prorealcode.com/prorealtime-indicators/super-bandpass-filter-john-ehlers/. Preserve the published filter recurrence and symmetric RMS-crossing entries and exits. The Research Lead selects COMP and TRX perpetuals, 4-HOUR bars, initialization and conservative sizing. Market orders execute at the completed signal bar's close, the registry's continuous-market equivalent of next-bar-open execution. Risk control route (b): signal exits with fixed 5% equity notional per instrument, maximum 10% gross exposure, leverage 1.0 and no pyramiding. Predict 30–70 completed trades per instrument annually, generally held for several days. Hyperliquid base-tier taker fees are 0.045% per fill, matching the engine and [official fee schedule](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/fees). Capture and frequency estimates are untested predictions.

Derived by the Research Lead from [Quantuple's butterfly payoff explanation](https://quant.stackexchange.com/a/28014): test whether a weekly ETH call butterfly is underpriced relative to an empirical physical terminal-price distribution. The source supplies the bounded +1/-2/+1 call payoff, not a trading strategy or evidence of profitability. Use RollingOptionsStrategy directly, actual historical option contracts, and ETH-PERPETUAL.DERIBIT-4-HOUR-LAST-EXTERNAL as the untraded primary clock. Market orders execute at completed signal-bar closes, the registry's equivalent of next-bar-open execution on continuous markets. Request 1095 days of overlapping history and report actual coverage. Predicted frequency is 10–20 structures annually, without a performance pre-study.

Test equal-base-quantity funding carry using AXSUSD.HYPERLIQUID as the primary short and AXSUSDT.OKX as the long hedge, with synchronized 4-HOUR bars. Normalize past settled funding by actual calendar time rather than assuming identical settlement frequencies. Hyperliquid settles hourly ([official documentation](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/funding)). The funding snapshot contains 268 overlapping days from 2026-01-18 through 2026-10-12, with a mean daily HL-minus-OKX funding difference of +0.072161%. This is descriptive evidence, not a strategy backtest. Enter market orders at synchronized signal-bar closes, the registry's equivalent of next-bar-open execution on continuous venues. Use existing funding-accrual reporting without double-counting cash flows. Risk control route (b): capped, unlevered sizing with funding, basis and time exits. Each leg is capped at 10% of combined portfolio MTM equity, with 20% maximum gross exposure.

Test whether unusually asymmetric ETH option insurance demand subsequently normalizes. Buy the relatively cheap Deribit option wing through a debit vertical and sell the expensive wing through a protected credit vertical, all at one expiry. Hedge the resulting delta with ETHUSD.HYPERLIQUID. Use completed 4-HOUR bars and actual overlapping historical coverage. Extend RollingOptionsStrategy directly; its chain resolver adds every concrete historical option instrument id to extra_instruments and persists those ids. ETH-PERPETUAL.DERIBIT supplies only its own primary bars and is never traded. Continuous-market orders execute at the signal close, the registered equivalent of next-bar-open execution. This is original autonomous research, not a source port. Fee references: [Deribit standard fees](https://support.deribit.com/hc/en-us/articles/25944746248989-Fees) and [Hyperliquid fees](https://hyperliquid.gitbook.io/hyperliquid-docs/trading/fees). Assume no combo discounts.

Derived by the Research Lead from [Deribit Insights Episode 76](https://insights.deribit.com/crypto-options-unplugged/episode-76-crypto-pulls-back-pre-jackson-hole/), which discusses protection buying and retaining hedges during crypto strength: test a monthly ETH collar consisting of a purchased downside put, a sold upside call and an equal-base-quantity long Hyperliquid perpetual. Only the archived article summary was read, not the podcast. These trading rules are my operationalization, not the presenters' strategy. Use RollingOptionsStrategy, a FactoryStrategy subclass, with the supported single-expiry Deribit chain. ETH-PERPETUAL.DERIBIT supplies the untraded primary 4-HOUR clock; ETHUSD.HYPERLIQUID supplies the underlying exposure. Historical option full ids are resolved through rolling_options and injected into extra_instruments before verification. Market orders execute at completed signal-bar closes, the registry's equivalent of next-bar-open execution on continuous venues. Fills are sequential, not atomic. Evaluate all actual joint history; do not substitute the longer underlying history for option coverage.

Reframe the failed premium arbitrage as a directional test of venue-local speculative pressure. Use synchronized 4-HOUR bars for ZECUSD.HYPERLIQUID, traded long or short, and ZECUSDT.BYBIT, signal only. Fade substantial ZEC moves when Hyperliquid's normalized turnover exceeds Bybit's. Remove the absolute price-gap requirement and passive paired entries. Market orders execute at the signal-bar close, the supported equivalent of next-bar-open execution on continuous venues. Protective stops trigger intrabar in the platform. Evaluate all common available history without assuming this listing has 2.3 years. Baselines: move_bars=6, activity_lookback=120, shock_pct=3.0, activity_ratio=1.25, max_hold_bars=18. Predicted frequency is 25–60 trades annually, unmeasured. No performance pre-study was run.

Test whether opposing changes in normalized trading activity between privacy-focused crypto assets predict subsequent relative performance before their relative prices move substantially. Trade XMRUSD.HYPERLIQUID against ZECUSDT.OKX using synchronized 4-HOUR bars, either direction. Execute market orders at the completed signal bars' closes, the supported equivalent to next-bar-open execution on continuous markets. Compare activity within each instrument, never raw volumes between venues. Exact catalog directories exist; filenames indicate common coverage beginning 2025-11-06, subject to continuity and volume-unit validation. This is an original participation-migration hypothesis, not cointegration reversion, funding carry or an established trading edge. Intrabar protective stops are matched by the platform. No preliminary backtest was run.

Derived by the Research Lead from the FMZ Aroon digest as summarized in the original hypothesis: test relative-recency persistence only after a compressed AVAX/NEAR log-price spread breaks its preceding range. The source text was not supplied in this revision; the regime and exit rules are Research Lead choices. Use synchronized 12-HOUR closes. Market orders execute at the signal close, the supported continuous-market equivalent of next-bar-open fills. Risk route (b): signal exits with capped sizing; allocate 10% of mark-to-market equity per leg, 20% aggregate entry exposure, leverage 1.0, and a 25% ongoing per-leg exposure cap. Predict 10–20 paired episodes annually, held up to seven days. Cost arithmetic: a hypothesized mean signed relative return of 2.0% yields approximately 1.0% gross per aggregate entry notional; four 0.05% fills cost approximately 0.10% on that denominator. A further 0.20% allowance for impact and adverse funding leaves 0.70% net, exceeding twice the assumed 0.30% total cost. These are predictions, not measured findings. Fee reference: [OKX futures fee calculation](https://www.okx.com/en-us/help/how-to-calculate-the-contract-transaction-fee). Use actual overlapping history and retain all exposed lineage evidence.

Faithfully implement the paper's contrarian signal on SPY.USEQ daily bars. Baseline return_lookback=1: calculate log(close_t/close_(t-1)), buy after a negative return and sell after a nonnegative return. The latest completed return is the previous interval relative to the forecast period; do not introduce another session of lag. No RSI, crash threshold, stop, target or predictive model is added. Interpret source DFL=1 as unlevered exposure capped at 100% account NAV, with whole shares. Risk route is signal exit plus capped sizing. Hold existing quantity through consecutive negative-return sessions without unnecessary close/reopen turnover. Decisions execute at the next session open, so an entry cannot earn the preceding overnight gap. Search return_lookback over 1–3 sessions around the source's one-day baseline; retain an explicit baseline replication report. Predict approximately 50–90 completed positions per year.

Port the paper's ARIMA(p,0,q)-GARCH(1,1) daily-return forecasting rule to SPY.USEQ 1-DAY. Preserve daily parameter estimation and quarterly order selection. Baselines are training_bars=756, max_order=5 and order_criterion=AIC. Fit Gaussian joint ARMA-GARCH conditional likelihood on log returns with an intercept. Enumerate p and q from zero through max_order, including (0,0). Initialize residual history at zero and conditional variance at training sample variance; require positive omega, nonnegative alpha and beta, alpha+beta<1 and stationary/invertible ARMA coefficients. At the first eligible session and first session of January, April, July and October select orders by the lowest information criterion; ties prefer lower p+q, then lower p. Daily re-estimate the chosen order on the trailing training window. On estimation failure reuse the last valid model and advance its state with observed returns; remain cash if none exists. Implement genuine joint likelihood estimation with sandbox-allowed NumPy; statsmodels and arch imports are unavailable, and an AR-only substitute is not faithful. Interpret source DFL=1 as exposure capped at 100% account NAV, whole shares, no leverage. Risk route is signal exit plus capped sizing, with no added stop. Fills occur at the next session open rather than the paper's decision-price accounting. Daily model refitting is part of the fixed algorithm; optimizer-selected meta-parameters stay frozen.

Faithful port of Crawford Das, Tie and Zhang’s bidirectional GBM threshold policy. Trade WMT and TGT on USEQ daily bars, retaining one completed pair round trip per evaluation window. Adapt the paper’s share units to equal-dollar model units anchored at the first synchronized warm-up close: x1=100*WMT_close/WMT_anchor and x2=100*TGT_close/TGT_anchor. Freeze anchors throughout the window. This denomination adaptation is not an exact reproduction of the historical example. Use split-adjusted prices; dividends and stock-borrow costs are not fully modeled. Observe thresholds at synchronized daily closes and execute market orders at the next session opens. Risk route (b): source signal exits with entry gross exposure capped at 50% of equity, each leg capped at 25%, leverage 1. Preserve the source’s absence of stops, targets and time exits. After liquidation, remain inactive.

Derived by the Research Lead from the supplied Myquant ATR_STOCK port: preserve the daily close > open + 2xATR14 expansion signal, original data-covered A-share pool, 6.8% loss exit, 20%-armed 6.8% trailing profit exit and 22-calendar-day lockout. This is an amended port. Replace source same-close execution with session-supported next-open execution. Explicitly authorize an IOC LIMIT buy with limit = signal close * 1.005 + one instrument price tick, rounded upward to a valid tick. It is eligible at the next session open only; if the opening execution price exceeds the limit, cancel without chasing or resting through the session. A signal close of 10.00 followed by an affordable open of 10.20 intentionally cancels. Lockout starts at the qualifying signal even if quantity is zero or the order cancels. Retain partial fills and cancel the remainder. Reserve cash using the limit price and modeled entry costs without consulting the future open. Daily exit decisions fill at the next session open. Source text was not independently fetched; source attribution and rules come from the supplied brief. CNEQ T+1 and price limits are not modeled; dividends are excluded. Preserve the fixed source pool as a reproduction universe, disclose coverage and survivorship limitations, and subscribe only once to duplicated 601899.CNEQ.

Revise the supplied 14-name OKX daily MAX basket by reducing fixed gross exposure from 1.00× to 0.30× current mark-to-market equity. Preserve the universe, weekly Monday 00:00 UTC rebalance, MAX ranking and 3×14-day ATR protective stops. Allocate at most 0.15× equity long and 0.15× short, equally weighted within each side. Authorize proportional reductions of retained legs to enforce the aggregate budget and matched side notionals. Confirm cancellations, exits and reductions before submitting replacement entries; reserve margin and execution costs across every pending order. Market orders at the signal bar’s close implement the registered next-bar-open equivalent on this continuous venue. Protective stops trigger intrabar through platform matching. Preserve all prior research looks and the exposed holdout’s status.

Event-driven, multi-instrument, long-short basket on HYPERLIQUID daily bars. Mechanism: tokens whose perp is newly listed on Hyperliquid (TGE launches, airdrops, memecoins that just went viral) start life with a small float, a high FDV, airdrop recipients and early backers selling, unlock schedules ahead and a listing-hype premium. Once the perp exists, that supply overhang can be expressed short, and prices decay over the following weeks. The trade: on the entry_day-th UTC daily close after each pre-registered listing date (baseline entry_day = 3), short the new perp with a FIXED 3% of equity and buy the same notional of BTCUSD.HYPERLIQUID as a market hedge. Hold hold_days (baseline 45), then close both. No signal fitting: the event (listing date) is known at the time it happens. UNIVERSE (pre-declared, point-in-time): every Hyperliquid perp whose first daily bar falls between 2024-01-01 and 2026-08-24 (each has at least 48 days of history by now). That is 110 listings, including those later DELISTED (MYRO, JELLY, OMNI, LISTA, CATI, kDOGS, NEIROETH, AI16Z, LAUNCHCOIN...), so survivorship does not flatter the short side. Excluded by a pre-declared rule, not by returns: ticker rebrands or migrations of a token already on HL (RENDER<-RNDR, POL<-MATIC, S<-FTM, SKY<-MKR) and the tokenized-gold RWA PAXG. The listing date of each instrument is passed as the FIXED parameter map `listing_dates` in optimization_plan.fixed (symbol -> first HL daily bar date, from the catalog). The strategy MUST use this map, not 'the first bar this backtest saw', because walk-forward windows start mid-history. PRIOR ART: two earlier versions of this idea (Bybit 76 listings; Binance USD-M ~37 listings) died for INFRASTRUCTURE reasons (sandbox timeout and data_unavailable) and never produced a measured backtest. Neither is evidence about the idea. This version moves to Hyperliquid (11.5% venue share vs a >=20% target) where full daily history of delisted perps is in the catalog. RL PRE-CHECK (rough, from catalog daily closes, no costs, no stop, not a backtest; first-bar dates are an approximation of listing): across about 100 listings from 2024-03, a short from the day-3 close held 45 days returned a mean of +20.6% (median +35%) unhedged and +22.4% BTC-hedged, with 78% of legs winning. By listing year: 2024 +20%, 2025 +26%, 2026 +8% (only 5 legs, weaker; crowding risk). Hold 30 days: +12% mean; hold 14 days: -2.4% mean. The decay needs weeks, and squeezes dominate the first two weeks. Right tail: half of all legs see a >=+30% adverse excursion inside 45 days, and stops at +40% or +60% cut the mean to about +8-9% (squeeze-then-dump). The risk is therefore controlled by SMALL fixed sizing, not a tight stop. Funding: Hyperliquid funding received or paid by the short over days 4-48 averaged -0.8% (median +1.35%, 10th percentile -9.3%; 2026 listings -10.8% on average, the crowding signal). This motivates the funding-squeeze gate. FEE ARITHMETIC: expected per-alt-leg capture is about 15-20% after funding vs Hyperliquid round-trip taker cost of about 0.09% per leg (0.18% for both legs): over 50x the cost. EXECUTION: 24/7 venue, so orders decided at a daily close fill at that bar's close (the platform's equivalent of next-bar open). The disaster stop is an intrabar protective stop (reduce-only stop-market, matched by the platform inside the bar), not an exchange-resting order. Expected cadence: about 40 listings a year, about 5-6 concurrent legs on average, at most 15. Implementation notes: daily bars only, O(1) work per bar, funding via supp_as_of/supp_window only (never full-history rescans), because a prior 76-instrument version timed out in the sandbox. An instrument with no bars yet (before listing) or after delisting must be skipped gracefully. If an open leg's instrument stops printing bars (delisted), close it at the last available bar.

Mechanism: the 08:30 ET slot carries most US tier-1 macro releases: NFP, CPI, PPI, retail sales, GDP and jobless claims. These move real yields and the dollar, and gold reprices on them more directly than any crypto asset does. The first 15 minutes are dealer and algorithmic repricing. Real-money macro and CTA positioning (ETF flows, the 10:00 ET London PM fix and the US session) then continues to push in the same direction for the rest of the New York morning and afternoon. The test asks whether that slower follow-through is tradable in spot gold. No hard-coded event calendar is needed: an event is any weekday whose 08:30-08:45 ET 15-minute bar (stamped 08:45 ET, because bars are close-stamped) has a high-low range of at least k times the median range of the same slot over the previous N weekdays. A release that moves gold detects itself; quiet release days and no-release days stay below the threshold. DST is handled by converting to America/New_York with the platform session calendar. Direction comes from close location in the event bar. Close in the top fraction q of the range means long; close in the bottom q means short; a close in the middle means no trade (ambiguous reaction). Execution: SPOTMETAL is a session venue, so the decision on the 08:45 ET bar close fills at the next bar's open (next_open_fill, supported) with a market order. The exit is a market order at a fixed New York time (exit_time_et), always at or before 16:00 ET. The position therefore never crosses the 17:00 NY overnight-financing cut and pays no financing. Protective stop: an intrabar reduce-only stop-market at the event bar's opposite extreme (below the event-bar low for longs, above the high for shorts), matched inside the bar by the platform; no exchange-resting order is required. At most one trade per day. Novelty against prior art: the crypto versions failed. 'US Macro-Release-Hour Repricing Continuation' on the AAVE USD-M perp (fee_edge, -0.24%/trade) and the two ETH macro-event continuation tests (fee_edge/overfit) traded an asset with no first-order link to US rates, paid ~0.10% round trip, and used a calendar of only 3 event types. This test moves the mechanism to the asset whose price is directly a function of US real rates and the dollar. Costs are ~0.015% round trip (6-7x cheaper), and the data-slot detector is calendar-free. Existing spot-gold hypotheses in the corpus are Asian-range / London-open breakouts, an Asia-session drift, a previous-day-level OCO breakout and daily macro-trend ideas; none conditions on the US 08:30 ET release slot. Diversity: SPOTMETAL holds 0.2% of experiments; the strategy is long/short and short-horizon (15m), against a corpus that is 52.7% BINANCE and 77.8% long-only. Fee arithmetic: in the 2003-05 to 2007-07 catalog slice (the Dukascopy backfill is still progressing, so this is all the history currently on disk), event days at k=2 with a decisive close-location show a mean absolute 08:45->16:00 ET move of ~0.8%. If continuation holds 55-58% of the time with a symmetric payoff, the signed capture is ~0.12-0.20% per trade, 8-13x the ~0.015% round-trip spread. Disclosure: that same small early slice (58 events, 2003-2007) gave a NEGATIVE signed mean at k=2 (-0.29%). Direction is pre-registered as continuation from the economic argument and is fixed, not tuned. If full history confirms fading, the hypothesis is falsified, not flipped. Firing rate: ~14 events/yr at k=2 in the early slice, more in the higher-volatility post-2008 regime. Expect 250-450 trades over 2003-2026 once backfilled. Risk control route (a): position size = risk_pct x equity / stop distance, where stop distance = |entry - opposite event-bar extreme|. This loses 0.75% of equity at the stop, with notional capped at max_leverage x equity (fixed 4x, below the 20x venue cap).