Absorbed-Flush Rule, Fourth Frozen Breadth Replication on a Name-Disjoint 18-Alt Bybit Perp Panel: buy at the close of a down day whose range was >=2x its 60-day return std but that closed in the upper 40% of its range, fixed 3-day clock, 10% per leg, max 6 concurrent (BYBIT 1-DAY, long-only event panel, flat most of the time)

Hypothesis

This is an all-fixed plan on purpose. The rule is frozen from two promotions (L262: keep range multiple, close location and clock fixed, use a universe with no shared names, pre-register Sharpe and drawdown). A parameter search would turn a breadth test into selection. The pre-registered numbers: 121 distinct event dates and 297 name-events before the concurrency cap; Sharpe at least 0.5; drawdown at most 22%. The drawdown matches what the mechanism needs at 6 x 10% correlated legs; the spot replication's 22.7% shows the scale. Bybit is chosen to fill its quota gap, and it is the venue where the rule first worked, now on names never used. That isolates universe breadth. BTC, ETH and Binance USD-M are not used, so this does not add to their concentration.

Implementation

Long-only event panel on 18 Bybit USDT perps (SOL primary + 17 extras, all 1-DAY bars, none shared with the earlier Bybit/HL/spot panels). At each daily close, for every listed leg whose bar closes at that instant: sd60 = sample std of the prior 60 daily log returns (today excluded); range_pct = (high-low)/prev close; close_loc = (close-low)/(high-low) (skipped if high==low). Signal = close < prev close AND range_pct >= 2.0*sd60 AND close_loc >= 0.60, leg flat, fewer than 6 positions open; extra signals beyond free slots are ranked by range_pct/sd60. Entry: market BUY at the signal bar's close (24/7 equivalent of the next bar's open), notional = 10% of account equity, leverage 1. Exit: market SELL at the close of the 3rd daily bar after entry (from position.ts_opened). No stop/target.

Frozen absorbed-flush rule (L262) implemented exactly with the plan's fixed parameter names and values; no tunables. Execution: next_bar_open_fill is use_equivalent, so entries are market orders at the signal close. Point-in-time: point_in_time_universe=true and candidates are drawn from universe_at() at each bar; a leg needs sd_lookback+2 = 62 of its own bars (60 prior returns excluding today plus today), which satisfies the >=61-bar listing warm-up. Extra legs only contribute when their newest bar is stamped exactly at the primary's decision instant (no look-ahead). Risk control route (b): fixed 3-day time exit with sizing capped at 10% of equity per leg, max 6 legs = 60% gross, unlevered; well inside the conservative risk profile. The code is the promoted Hyperliquid replication's implementation with parameter names aligned to this plan and close_loc threshold inclusive (>=0.60) as specified.

Verification Results

Optional: make the earliest-listed name the primary, or confirm that SOL's first Bybit bar comes no later than the other legs' listings plus 60 bars.

SOL is the primary leg, and the base template's barrier only processes a timestamp when a primary bar arrives. Its min_bars_required is 62, so no leg can trade until SOL has 62 bars of its own. If SOLUSDT.BYBIT was listed after some extras (e.g. XRP or LINK from 2020), those legs' early flush events (part of the 2021 pre-study sample) are never evaluated. This shortens the sample but does not bias it toward any outcome.

None required.

When a leg's 3-day exit falls on the same close, its slot is freed for new entries on that bar (a due position is not counted as active). That leg itself stays in 'held' and cannot re-enter on the same bar. This reading of 'fewer than 6 positions open' is reasonable (the exit and the entry fill at the same close) and is noted only for transparency.

Optional: keep rolling sums of the returns if the full-history runtime becomes a problem.

The sandbox ran for 186 s and was flagged runtime_review_required. _features recomputes the 60-bar standard deviation for all 18 legs on every bar, which is O(18*60) per bar. That is a cost concern, not a correctness one.

The evidence rests on 121 distinct event dates, with legs firing together on market-wide flush days. The pre-study return is right-skewed (mean +1.72%, median +0.37%), 2021 and 2026 were negative, and the panel contains only today's Bybit survivors, which biases a long-only rebound upward. The analyst should judge on event dates rather than trade count, and check how much of the result comes from the few largest rebounds (GALA, FLOW, LRC, CRO).

Backtest Review

Sharpe

0.65

Total return

46.47%

Max drawdown

11.13%

Trades

228

Win rate

51.8%

Profit factor

1.58

The code follows the frozen rule. All 198 trades are long. Every hold is exactly 3d 0h, and every position is 10% of equity (leverage 0.1 each). At most 6 trades open on the same date, which matches max_concurrent.

Economics on the development period (2021-12 to 2025-10) clear the bar comfortably: 198 closed trades, an average trade of +2.42% of notional against the Bybit floor of about 0.15%, profit factor 1.91, Sharpe 0.87 (95% CI -0.04 to 1.75), +56.6% total return, CAGR 12.3% and drawdown 10.6%. That is inside the pre-registered 22% limit and close to the 958bf541 sibling (PF 1.72, Sharpe 0.82, DD 10.2%).

Every calendar year with trades is positive: 2022 +0.75%, 2023 +5.9%, 2024 +30.5%, 2025 +12.5%. The strategy lost little in stress windows: LUNA +7.0%, FTX -1.2%, SVB -0.1%.

The plan is all-fixed. Optimization here means one evaluation of a frozen rule (walk-forward plus holdout), not a parameter search, so it adds evidence without a selection penalty.

The trades fall on 90 distinct entry dates, below the 121 dates in the pre-study, because of the cap and the development window. That is above 's floor of about 60 event dates, but the sample is thin and correlated. The real N is the dates, not the 198 trades.

Profit is concentrated. The 3 best entry dates make about $31k of the $56.6k total (about 55%), and the top 10 trades make $34.1k. Only 48 of the 90 dates are positive. Without the 3 best dates the result is still positive (about +$25.6k), but QA's right-skew concern is partly borne out: the result depends on a few cluster days, and 2024 alone makes 57% of the profit.

Five of the 18 names lose money: IMX -$4.5k, KSM -$3.3k, ENS -$2.0k, SOL -$0.7k and MASK -$0.5k. GALA, FLOW, LRC and CRO, the names QA flagged, earn $33.6k between them, together with XRP's $9.6k.

The panel holds only Bybit's surviving contracts, so a long-only rebound is biased upward. The holdout and paper trading have to carry the forward test.

The worst trade lost $2,759, about 2-3% of equity. That fits the route (b) time-exit plus capped-size control: no position is above 10% of equity and gross exposure is at most 60%.

Analysis

This is a frozen-rule breadth replication with an all-fixed plan and one OOS variant, so the hurdle is not inflated by search. All 6 pre-registered checks that could be scored were met: 228 trades vs at least 180, PF 1.58 vs 1.3, average trade 1.81% vs 0.6%, return +46.5% vs 40%, drawdown 11.1% vs at most 22%, win rate 0.52 vs 0.48. The OOS Sharpe prediction (at least 0.5) is not scored in the brief; the pooled OOS Sharpe was 0.96.

The economics clear costs by a wide margin. The average trade is 1.81% of notional against the 0.22% floor (BYBIT round trip about 0.11%). Commission is 2.4% of gross and impact 4.4%. With fees, spread and impact raised 50% (about +$2.6k on $5.2k of costs plus $0.8k funding), net is still about +$43.9k, roughly 1.7% per trade.

Pooled walk-forward OOS over 6 disjoint windows: annual Sharpe 0.96, HAC t = 1.81 against a hurdle of 2.0 (N_oos = 1, so the deflation adds nothing; DSR/PSR 0.965 vs 0.977 needed). That is just short of significance, not adverse. 5 of 6 windows were positive.

Every parameter except close_loc_min sits on a wide plateau. The close_loc_min cliff is only on the tighter side (0.75 and above, which removes the very events the rule is about); looser values are flat or better. None of this was tuned.

Risk is small and inside the rules: unlevered, about 11% of equity deployed per trade (46.47 / (228 x 1.812) = 0.113), largest position 10.1% of equity, worst realised loss 2.6% of equity (ZEC, 2025-11-09), worst adverse excursion 5.7%, peak gross 61%. Beta 0.03, positive alpha, and the strategy makes money in both the stressed and normal volatility thirds.

On the full history the mechanism broadly reproduces the two promoted siblings on a universe with no shared names (Bybit 13 names: PF 1.72, Sharpe 0.82; Hyperliquid 15 names: PF 1.70, Sharpe 0.73; this panel: PF 1.58, Sharpe 0.65).

The holdout is negative: 2025-10-30 to 2026-11-02, 30 trades on 18 event dates, -$10.2k, Sharpe -0.91, t = -0.87. It is statistically inconclusive, not adverse, and consistent with OOS (z = -1.19 > -1.645). It matches the pre-study's negative 2026 and points to possible decay in the current regime.

Profit is extremely concentrated, as feared at the backtest review: the 108 entry dates hold $46.5k of profit. The best 3 dates (2024-03-09, 2024-08-06, 2024-02-29) carry $31.0k, or 67%; removing them leaves +$15.4k. The best 5 dates carry 99% (+$0.5k left). 2024 alone is 66% of the profit, which is a soft-floor failure. The effective sample is a handful of market-wide flush days.

The edge is right-skewed: median trade +0.26% vs mean +1.81%. 13 of 18 names made money. KSM, IMX, ENS, SOL and MASK lost a combined -$15.4k.

Survivorship: the panel is today's Bybit survivors, which biases a long rebound upward. The figures above are an upper bound.

Bybit is the same venue as the first promotion, so this tests breadth across names, not across venues. On paper it is highly correlated with the two absorbed-flush siblings already running, so it adds little diversification.

t=-0.87, inconclusive; consistent_with_oos true (z=-1.19)

positive and significant

unevaluated in backtest metrics; the pooled OOS DSR is 0.965, short of the 0.977 needed

advisory only

inconclusive, N_oos=1

soft, waived

Risk Review

Risk control route (b) holds: a fixed 3-day time exit with capped, unlevered sizing. Each leg is 10% of equity x 1.0 leverage, with at most 6 concurrent legs (60% gross). Measured on the backtest: largest position 10.1% of equity (limit 100%), worst realised trade loss 2.01% of equity (IMX, entered 2024-08-02, -20.1% of notional; the profile allows 15%), worst adverse excursion 5.7% of equity, peak margin utilisation about 0.61. Realised volatility is 14.5% against the conservative profile's 5-60% band, and peak single-position volatility is 17.7% against a 120% cap. No stop-loss is needed on this route.

Tail and drawdown concentration. Development max drawdown is 10.6% (11.1% on the gate's full evidence), well inside the pre-registered 22% at 1x. The drawdown bootstrap upper bound is 25.5%, though, and the longest drawdown lasted 362 days. Returns are strongly right-skewed (skew 2.1, kurtosis 33) and depend on a handful of market-wide flush days: March 2024 alone returned +15.5%. Events cluster, so up to 6 correlated alt longs open on the same day. With 60% gross, a market-wide crash continuing after a flush day can hit several legs at once (Aug 2024 yen-carry window: 7.9% intra-window drawdown).

Correlation with the deployed book. Two absorbed-flush siblings are already in paper (BybitFlushAbsorptionPanel and HyperliquidAbsorbedFlushReboundPanel), along with several other capitulation and down-shock rebound panels (e.g. BybitSpotAltPanelCapitulationAbsorption4H, HyperliquidDailyDownShockAbsorptionPanel). The names do not overlap, but all of them fire on the same market-wide flush days, so this strategy adds long-alt exposure on the same dates rather than diversifying. The live book shows no members in breach and no open execution items. Every sibling is still collecting with 0 decisions, so their forward correlation has not been observed yet.

The holdout (2025-10-30 to 2026-11-02, 368 days) was inconclusive and its point estimate negative per the gate. Pooled walk-forward out-of-sample results also fell short of significance (HAC t 1.81 vs the 2.0 hurdle). This is labelled low-confidence paper evidence only. Survivorship bias (today's Bybit survivors) means the backtest figures are an upper bound.

Market impact: about $20k per leg on $200k equity, on liquid Bybit USDT perps. Measured impact is about 4.4% of gross. That is acceptable, but thinner names (LRC, FLOW, MASK, DASH) on flush days deserve monitoring in paper. Trade frequency of about 198 trades over 3.9 years at about 1.8% per trade net is fine; commission is a small share of gross.

Paper only, with the low-confidence label shown. This approval is not real-money evidence.

Treat this panel and the two deployed absorbed-flush siblings as one risk sleeve for capital allocation: on a shared flush day, the combined new long-alt notional across the three must not exceed what any one of them may deploy at its own cap.

Keep leverage at 1.0, position_pct at 0.10 and max_concurrent at 6. Any increase needs a new risk review.

Outcome Summary

The rule's profit rests on a few market-wide flush days (the best 3 entry dates made 67% of profit, the best 5 made 99%, and 2024 alone made 66%), so the real sample is the number of event dates, not trades; forward and paper results should be judged on that basis, with survivorship bias and possible decay in 2026 in mind.

The analyst recommended promotion: every scorable pre-registered check passed and costs were cleared by a wide margin, while OOS significance fell short and the holdout result (t = -0.87) was inconclusive, not adverse. The strategy then reached paper trading.

A frozen breadth replication of the absorbed-flush rule on 18 Bybit USDT perps that share no names with the earlier panels: buy at the close of a down day whose range was at least 2x its 60-day return std but which closed in the upper 40% of that range, hold for exactly 3 days, 10% of equity per leg, at most 6 positions open at once, long-only and unlevered.

Over the full history (2021-12 to 2026-11) it made 228 trades for +46.5% total return, Sharpe 0.65, profit factor 1.58, an average trade of +1.81% of notional and an 11.1% max drawdown. Pooled walk-forward OOS Sharpe was 0.96 (HAC t = 1.81 against a 2.0 hurdle), but the holdout (2025-10-30 to 2026-11-02, 30 trades) lost $10.2k with Sharpe -0.91.