Strategies

On-Chain Venue Flow Migration Fade: Short a Hyperliquid Alt Perp, BTC-Beta-Hedged, for a Fixed 5 Days After an Up-Day on Which Hyperliquid's Share of Its Perp Volume (HL vs Binance USD-M) Spikes >=2.5 Sigma (25 HL alts + BTC hedge, Binance volume signal-only, 1-DAY, market-neutral event panel, flat most of the time)

Outcome: Backtesting

HlVenueFlowMigrationFadeBtcHedged

On-Chain Venue Flow Migration Fade: Short a Hyperliquid Alt Perp, BTC-Beta-Hedged, for a Fixed 5 Days After an Up-Day on Which Hyperliquid's Share of Its Perp Volume (HL vs Binance USD-M) Spikes >=2.5 Sigma (25 HL alts…

Strategy report

Hypothesis

Original Research Lead hypothesis. It is a cross-venue, BTC-hedged event panel built on a signal the corpus has not used: where an alt's leverage is being traded. Each UTC day, for each of 25 alts, compute Hyperliquid's share of the coin's perp quote volume, share = HLvol / (HLvol + BinanceUSDMvol), where quote volume = base volume x close on each venue's own 1-DAY bar. Then compute z = (log share - 60-day mean of log share) / 60-day std, using the prior 60 days only. Mechanism: Hyperliquid is the on-chain, KYC-free, high-leverage venue. Its flow is dominated by crypto-native leverage chasers and visible whale/vault positioning, which copy-traders and liquidation hunters then target. When a coin rallies AND its leverage migrates to HL far beyond normal, the rally is being financed by marginal on-chain longs, not by the broad CEX market. That positioning is crowded and transparent (HL positions and liquidation levels are public), so it unwinds over the following days. The idea is not a price-shape fade: the control in the pre-study (up-days of similar size WITHOUT a share spike) shows almost no drift. PRE-STUDY (Research Lead, catalog 1-DAY bars, HL history from each coin's HL listing in 2023-24 to 2026-10-17, gross of costs, entry at the signal day's close; it was used to shape the rule, not to choose names). Up-day with share z >= 2.5: n = 304 name-days on 222 distinct dates, raw HL forward return -1.19% over 3 days and -3.0% over 5 days, median -1.1% at 3 days. BTC-beta-hedged (alt minus 60-day beta x BTC, beta clipped 0.3-2.5), the 5-day return is -2.64% (median -2.43%), negative in every year: 2023 -1.5%, 2024 -4.9%, 2025 -0.9%, 2026 -2.2%. Unconditional hedged 5-day drift is -0.58% overall and -0.78% after up-days (alts lagged BTC over the period), so the incremental effect of the share spike is about -1.9% per 5 days. The z >= 2.0 threshold gives -1.60% hedged (651 events), z >= 3.0 gives -2.41% (117), so the effect is monotone, not a cliff. Non-overlapping per name (5-day lockout): 222 events (~60/yr), hedged 5-day -2.12%. Dropping the best 5% of short outcomes it is still -0.76%. Control: up-days with return z > 1 and NO share spike (z < 1) drifted -0.06% over 3 days (n = 2,770). Low-share days (z <= -2.5) drifted +0.47%, which is consistent with the mechanism. Caveats stated up front: 2023 sample is small (41 events); per-name results are mixed (AVAX/APT/INJ/SUI strongly negative, DOGE/ARB/DOT slightly positive), and all names stay in. Lessons L230 and L237 are addressed: the short arm is measured on these exact names, and the event count inside the full-overlap span is ~220 non-overlapping, well above 30. COST ARITHMETIC: HL taker round trip 0.09% on the alt leg + 0.09% on the beta-sized BTC leg + ~0.05% impact on mid-caps = ~0.25% per hedged pair. Expected gross hedged capture ~2.1% (non-overlapping) minus ~0.8% alt-underperformance drift that should not be counted as edge = ~1.3%, less costs = ~1.0% net per pair, about 4x the round trip. Funding on the short alt leg is usually received after rallies (positive funding) and is a tailwind, credited by the platform's funding ledger. Execution: a decision on the daily close is a market order at that bar's close (24/7 venue: equivalent to the next bar's open, which is the platform's use_equivalent for next_bar_open_fill). The protective stop is an intrabar stop matched by the platform in backtest and paper, not an exchange-resting order.

Fills under-represented buckets: Hyperliquid (13.3% vs >=20% target), cross-venue (10% vs >=15%; Binance volume is a signal-only leg), long-short/market-neutral (24% vs long-only 76%). Avoids BTC as the traded series (BTC is only the hedge). New mechanism (venue flow migration), not a candle pattern, carry, or a re-run of NR7/absorption. Event-triggered and flat most of the time, per L231/L224. The short arm is pre-registered on these exact names per L230; the event count inside the full-overlap span is pre-registered per L237.

Implementation

Daily (1-DAY) event panel across two venues on 20 Hyperliquid alt perps. For each alt, Hyperliquid's share of perp quote volume is HL base volume x HL close divided by (that plus Binance USD-M base volume x Binance close). The share is log-transformed and z-scored against the PRIOR share_lookback (60) days. When z >= share_z_threshold (2.5) on an up-day (HL close above the previous close), the strategy shorts the HL alt and buys the HL BTC perp as a hedge. The alt notional is equity x pair_notional_pct (0.1) x leverage (2). The BTC notional is beta x the alt notional, where beta is the 60-day OLS beta of daily log returns, clipped to 0.3-2.5. Each event keeps its own fixed BTC quantity, and the BTC position is the sum over live events. Both legs close hold_days (5) calendar days after the decision bar. A reduce-only intrabar BUY stop rests alt_stop_pct (20%) above the alt fill price. When it fires, that event's hedge is removed at the next close. At most max_concurrent (6) events run at once, strongest z first. A new event is skipped if gross notional would exceed 90% of equity x leverage (initial-margin fit). Binance bars are used only for the signal. legs_balanced() is checked every bar and flattens the book if an alt leg or the hedge is missing, refused or shrunk. Execution: market orders at the decision bar's close. On a 24/7 venue this is the platform's stated equivalent of a next-bar-open fill.

This implements the hypothesis as specified. The share is HLvol/(HLvol+BinanceUSDMvol) in quote terms, each from that venue's own daily bar. The z-score uses only the prior window, entries need an up-day, and the trade is a short on the HL alt with a BTC hedge sized by the 60-day clipped beta, held a fixed 5 days. The 20% alt stop is intrabar, which the platform supports. It is a raw reduce-only STOP_MARKET on the alt, placed in on_order_filled, because _submit_stop_market only covers the primary. Every parameter name and fixed value from the optimization plan is used verbatim. Leverage 2.0 is used in sizing (alt notional = equity x 0.1 x leverage). Universe: the title says 25 alts, but market_context declares 20, so those 20 are used, all of them, with no selection. BTCUSD.HYPERLIQUID is the primary because it is the hedge leg and the longest-lived anchor (listed 2023-05-12). point_in_time_universe=True, so later-listed alts (NEAR, ADA, TIA, DOT, SEI) join when they list instead of shortening the window, and each coin waits until its own buffers are full. Bar alignment: a coin is evaluated only when its HL and Binance bars carry exactly the primary bar's close timestamp. Two additions are platform constraints, not trading rules. (1) A margin-fit check (gross <= 0.9 x equity x leverage), so neither leg gets capped or refused, which would break the hedge. (2) A fixed BTC quantity per event, so price moves never resize a live hedge. The worst-case loss per event at the alt stop is about 20% x 20% = 4% of equity, against the risk profile limit of 15%.

Verification Results

Optionally add a check so that the summed BTC target notional (`_btc_target() * btc_close`) stays at or below 1.0 x equity, and skip an event that would push it over. This is a sizing guard, not a rule change.

Risk-control pre-screen (L239). Branch (a): the stop is a 20% alt stop on a leg worth 20% of equity, so the loss at the stop is about 4% of equity, above the 2% limit. A gap through the stop could lose more. Branch (b): the exits are signal/time based, but the netted BTC hedge is the sum of beta x alt notional over the live events. The only bound on it is the gross cap `if gross + add > cap` (cap = 0.9 x equity x leverage = 180% of equity). Example: three events with beta near 2 give 60% alt notional plus 120% BTC, which passes the gross cap and leaves one BTC position at about 120% of equity, above the 100% single-position limit. pair_notional_pct, leverage, max_concurrent and the beta clip are all fixed by the research contract, so implementing them is faithful and not a defect here. Risk review may still reject the strategy.

Acceptable as written. A reduce-only BTC trim could be submitted from on_order_filled when the reduce-only BUY stop fills, if tighter hedging is wanted.

When the 20% intrabar alt stop fires, that event's BTC hedge stays open until the next daily close. Step 1 then drops the event and _sync_hedge reduces the hedge. For up to one day the book holds net long BTC. The strategy description discloses this, and it only happens after a +20% alt move.

No change needed.

The hypothesis title says 25 alts, but research_contract.instruments lists 20 HL alts. The code uses exactly those 20 (COINS), with BTC as the hedge and Binance USD-M bars for the signal only. The contract is the specification, so this is not a mismatch. It is noted for the analyst because the pre-study event counts may have come from the larger universe.

The pre-study's -2.6% hedged 5-day figure is gross and entered at the signal close. About 0.6-0.8% of it is the unconditional underperformance of alts against BTC over 2023-26, which is a beta/regime drift and not this signal, so the incremental effect is about -1.9%. Per-name results are mixed: DOGE, ARB and DOT are slightly positive. The pre-study chose the z>=2.5 / 5-day shape on the same history the optimizer will search (z 2.0-3.0, hold 3-7, lookback 30-120). The analyst should compare the result with an up-day-only control at the frozen rule, and check that the profit is not concentrated in 2024 (-4.9% that year against -0.9% in 2025).

Iteration History

Risk-control pre-screen. Branch (a): the stop is a 20% alt stop on a leg worth 20% of equity, so the loss at the stop is about 4% of equity, above the 2% limit. A gap through the stop could lose more. Branch (b): the exits are signal/time based, but the netted BTC hedge is the sum of beta x alt notional over the live events. The only bound on it is the gross cap `if gross + add > cap` (cap = 0.9 x equity x leverage = 180% of equity). Example: three events with beta near 2 give 60% alt notional plus 120% BTC, which passes the gross cap and leaves one BTC position at about 120% of equity, above the 100% single-position limit. pair_notional_pct, leverage, max_concurrent and the beta clip are all fixed by the research contract, so implementing them is faithful and not a defect here. Risk review may still reject the strategy.

Backtest and paper results are hypothetical. Trading involves risk of loss.