Frozen Daily Donor-Pattern Transfer to Unseen Hyperliquid Alts

Derived by the Research Lead from Wood, Kessler, Roberts and Zohren's X-Trend paper (https://arxiv.org/html/2310.10500v2): test whether volatility-normalized daily patterns and their subsequent returns transfer from Binance donor assets to unseen Hyperliquid assets. This is our deliberately simpler nonparametric operationalization of donor-to-target contextual transfer, not a reproduction of X-Trend or its learned representations. The source motivates transfer and context similarity; every numerical rule, feature, weighting formula, label horizon, execution choice and risk rule below is our operationalization. Nothing in this lineage was coded or tested; the revision resolves implementation scope, not negative performance evidence. Implement in one FactoryStrategy file with NumPy and ordinary in-memory arrays. There is no GP segmentation, neural network, gradient computation, decoder training, external model, network access, model-file access, seed search or training-validation optimization inside the strategy. INSTRUMENTS AND CLOCK. Trade LINKUSD.HYPERLIQUID as primary and AVAXUSD.HYPERLIQUID and ATOMUSD.HYPERLIQUID as additional targets. ADAUSDT.BINANCE, DOTUSDT.BINANCE, LTCUSDT.BINANCE and BCHUSDT.BINANCE are donor-only subscriptions, never traded. Subscribe all seven to 1-DAY-LAST-EXTERNAL bars, timestamped at UTC close in nanoseconds. This is a fixed named transfer case study, not a liquidity-ranked or survivor-selected universe. Check listing eligibility point in time for each traded target. The donor pool excludes LINK, AVAX and ATOM on every venue. Target prices are used only for features, volatility, sizing and exits: never for fitting, labels, donor selection or model selection. The two supported venues require only ordinary market orders and one netted position per target. Submit at a completed daily signal close, the registry's stated continuous-market equivalent to next-bar-open execution. No hedge mode, exchange-hosted stop, selectable same-bar policy or executor framework is needed. CAUSAL FEATURES. Write r_t=C_t/C_(t-1)-1. sigma_t=max(0.005, population_std(r_(t-59),...,r_t)); annual volatility v_t=sigma_t*sqrt(365). At each date form f_t with five coordinates (C_t/C_(t-h)-1)/(sigma_t*sqrt(h)) for h in [1,5,21,63,126], clipping each coordinate to [-10,10]. The 15-dimensional pattern x_t concatenates f_(t-10), f_(t-5), f_t in that order. Require at least 147 consecutive daily closes for a valid pattern; require uninterrupted daily closes for every lookback and label window. No interpolation, full-history scaling, fitted ticker embedding or target-label input. Any nonfinite feature invalidates the pattern. FROZEN DONOR LIBRARY. Maintain causal donor histories from received bars, never read future catalog arrays. At the first common donor close after all four donors have at least 730 consecutive daily closes, designate that timestamp T0. Wait until all four donor bars at T0 have arrived. Freeze a library once, using only their last 730 closes ending at T0. For each donor, enumerate candidate endpoints backward from T0 minus 5 days in steps of 5 days, retaining those with valid x_u inside that 730-close window and complete next-five-day labels ending no later than T0. For each retained endpoint store x_u and q_u=clip((C_(u+5)/C_u-1)/(sigma_u*sqrt(5)),-5,5). Label and predictor normalization both use sigma_u, not future volatility. This deterministic five-day stride avoids overlapping label intervals within a donor; cross-donor dependence still exists. Sort rows by donor order ADA,DOT,LTC,BCH and ascending endpoint. Require at least 100 rows per donor; otherwise remain flat with an insufficient_donor_library diagnostic, never substitute another signal. Freeze the row set, features and labels for the entire evaluation/deployment instance; do not append new rows. Earliest target decision must be strictly later than T0. Each historical walk-forward, validation or holdout instance resets all state, builds its own library solely from its permitted pre-boundary warm-up history, and stays flat if that history is insufficient. Never reuse a later-cutoff library in an earlier window. With 730 pre-boundary daily bars supplied, the library is ready at the boundary; missing warm-up is an evidence/availability diagnostic. No target-return or donor validation search is performed by strategy code. SIMILARITY TRANSFER. For each target's valid current x, compute d_j=mean((x-x_j)^2) for every frozen donor row. Within EACH donor select the 20 smallest d_j, breaking ties by earlier endpoint; no target rows participate. For those 20 rows set a_j=exp(-(d_j-min_selected_d)/(2*temperature^2)), then normalize a within the donor. The predicted normalized five-day return for a donor is m_d=sum(a_j*q_j)/sum(a_j). Equal-weight donors: m=(m_ADA+m_DOT+m_LTC+m_BCH)/4. The target's predicted raw five-day return is mu=m*sigma_target*sqrt(5). Do not replace similarity weights with target momentum or fit coefficients on targets. Subtracting the minimum distance is a numerical stabilization, not a changed distance rule. All math is float64. temperature is the sole tunable, initial 1.0, permitted range 0.5..2.0. All other settings are fixed. A nonfinite prediction blocks entries and requests exit of existing exposure on its next tradable bar. ENTRY AND EXIT. Evaluate once per target per UTC daily close after the frozen library is ready, independently of other targets' arrival order. Cache processing timestamps to prevent duplicate orders; use on_extra_bar for extra targets and donors, and the base primary hooks for LINK without overriding base on_bar/on_start dispatch. While flat and with no pending order, enter long when mu>=0.006 or short when mu<=-0.006; otherwise remain flat. Entry threshold 0.006 is in decimal return, meaning 0.60%. Record actual entry fill timestamp and quantity. Hold the filled quantity unchanged until five daily intervals have elapsed from that fill, then close the entire position at that target's next completed daily close. Exit earlier at a valid daily close if sign(mu) is strictly opposite to the open position; zero mu does not trigger an early exit. Never resize an open position, pyramid, or maintain simultaneous opposing legs. Any exit takes precedence over entry. Do not enter again on the exit-decision date or while closure is pending; the next daily close after confirmed closure is the earliest new entry. No signal-change reversal is opened in the exit callback. There is no protective stop: this uses risk route (b), signal/time exit with capped unlevered sizing. SIZING AND MISSING DATA. At entry use current get_equity_mtm(). Desired absolute account-equity weight is min(1/3,0.05/v_target), giving at most 5% ex-ante annualized standalone volatility contribution per target and at most 15% in the conservative sum across three targets at entry. Config leverage=1.0. Before submitting, further cap the new notional by max(0,equity-sum absolute current marked notionals of all three open targets minus pending entry notionals). Maximum new single-target exposure is one-third equity and maximum new aggregate gross exposure is 100% equity; market moves may cause subsequent drift because quantity stays fixed. No leverage or volatility guarantee is asserted after price gaps. Round quantities down to instrument size precision; apply native minimum quantity/notional rules and an additional $10 minimum. Skip undersized entries; close full residual positions using the platform close helper. Do not submit duplicate orders while pending. Missing target bars block its new orders and leave other targets independent; if a target returns after a timestamp gap greater than one day, close any existing position on its returning bar and do not enter that date. Reset that target's contiguous feature warm-up after a gap. Missing donor bars invalidate crossing feature/label windows; after library freeze donor outages do not change the frozen model. If there is no tradable bar, no exit fill is invented. ECONOMICS AND EVALUATION. Planning hypothesis, not measured evidence: a five-day 3% directional move with 20% capture gives 0.60% gross per completed trade. Hyperliquid round-trip taker fees 0.09% plus a 2bp per-side slippage floor (0.04% round trip) total 0.13%; 0.60/0.13=4.62 times baseline cost and 0.60%-0.13%=0.47% before additional impact. Actual impact can consume this margin; the mu threshold is a model forecast, not proof of realized capture. Predict average net notional trade return of 0.30%, pooled OOS Sharpe 0.5, net total return 5%, and 60 closed target positions across available eligible history, with 15% maximum account drawdown. No funding income is assumed; actual execution costs are applied once by the platform. Keep slippage_bps_per_side=2 in config parameters; do not deduct synthetic fees in strategy returns. Report per-target trades, holding durations, costs and aggregate results. No auxiliary decoder comparison is required: this revision tests the simpler transfer model on its own and makes no claim to reproduce neural-model results. Do not run a quick performance screen to decide whether to implement. Sparse history or no qualifying forecasts is inconclusive evidence, not permission to weaken the rules.