Non-Participation Continuation on Bybit: When BTC Makes a >=2.5-Sigma 4h Move and an Alt Perp Fails to Follow Its Beta by >=1.5 Residual Sigma, Bet the Divergence Widens for 24h (BTC-beta-hedged, 10 Bybit linear perps, 1H, long-short)

Hypothesis

Fills under-represented buckets: BYBIT venue (2.3% vs >=5% target), long-short (15.7% vs <=55% long-only target) and market-neutral. The mechanism claim is new: continuation of non-participation residuals, the opposite sign to the dead residual-reversal family (L113), and it is pre-studied on two venues with a smooth parameter surface. Avoids BTC as a directional instrument (BTC is only the hedge/factor). Universe capped at 10 names per L173, with O(1) incremental statistics to avoid the timeouts.

Implementation

Event-driven BTC-beta-hedged long-short on 10 Bybit linear alt perps (1H). When BTC's 4h log return exceeds btc_sigma_k trailing sigmas and an alt's 4h residual vs its lagged 720h beta runs against BTC's direction by >= resid_sigma_m residual sigmas, trade the alt in the direction of its divergence (short the non-participant of a rally / long the outperformer of a selloff) with an opposite-direction beta*notional BTC hedge, hold hold_hours (calendar, from entry ts) with a 4% pair-P&L stop; max 6 names, one open trade per name. The BTC hedge is one netted position reconciled each bar to the sum of signed per-name hedges.

Direct implementation of the pre-studied residual-continuation rule: beta from 720 1h returns, BTC and residual sigmas from 720 overlapping 4h returns via O(1) running sums (residual variance = var_a - 2b cov + b^2 var_b, so warmup ~725 bars matching the plan's 730), all statistics lagged one bar, entry at bar close after the event. Sizing: alt notional = equity * leg_fraction * leverage (16% at 2x), gross alt+|hedge| capped at equity*leverage. All optimization_plan fixed/tunable names implemented verbatim. Alignment: every leg is 1H; an alt bar contributes only if its ts equals the primary ts (<= primary, same bar), the base barrier waits for same-ts legs. Per-bar cost O(#names), independent of span. Ledger note: each alt leg is its own position/trade; the BTC hedge is netted, so BTC trade count understates per-name hedge events.

Verification Results

Set per-name state only after confirming the order was submitted (e.g. compare _entry_diag counters or check orders_open/positions for the alt), or reconcile the hedge from actual open alt positions rather than intended ones.

Per-name trade state (_ent_ts, _hedge_qty, ...) is set unconditionally after _submit_entry_instrument, which can silently drop the order (no market / entry in flight / CASH guard). If the alt entry is dropped, the BTC hedge for that name is still reconciled in. The book then holds a naked BTC beta position for up to hold_hours, and the time-stop/pair-stop bookkeeping refers to an alt position that does not exist. This is rare in backtest because market orders fill in-event, but it matters in paper/live.

Optionally skip re-entry on the bar a name was exited, or carry the position forward with a refreshed entry timestamp when the side is unchanged.

A name that times out is cleared in the exit loop and can be re-entered on the same bar if a new event is active, which sends close_position plus a new market entry on the same instrument within one bar. Under NETTING this nets correctly, but it adds an avoidable fee round trip when the new side matches the old one.

When judging fee viability, have the analyst use the equity-based return per event (total net PnL / number of alt events / alt notional), not avg_trade_return_pct.

Ledger (L133): the BTC hedge is a single netted position, so hedge P&L and hedge fees are not attributed to the per-alt trades. total_trades (685) counts alt legs only. avg_trade_return_pct therefore mixes unhedged alt-leg returns with a separately booked BTC position, so it is not the per-pair residual return the hypothesis measures. The equity curve (Sharpe/return) is the correct measure of the pair book.

Optionally require ts - prev_ts == 1h before pushing a 1h return.

1h returns (btc_r1, alt r1) are taken versus the previous seen close without a timestamp check, so after a data gap a multi-hour return enters the 720-bar beta window as if it were 1h. The 4h window does check its span exactly. The effect on beta is negligible given the reported gap counts (BTC 0, LINK 1).

Sandbox, unoptimized, 834 days: -27.9% return, PF 0.87, avg_trade_return_pct -0.032% on 685 closed alt positions. The ~0.45% gross residual from the pre-study must cover about 0.11% alt round trip plus about beta x 0.11% on the BTC hedge (Bybit taker 0.055%/side) plus impact ($14.5k folded). That is roughly 0.25% or more per event, so net margin is thin. The pre-study also had negative years (2021/25), and its naive t of 4.2 overstates significance because events cluster on the same BTC shock. PF is 0.87, not <0.70, so this is not a polarity red flag. First checks: the per-year breakdown, and whether the 4% pair stop is cutting the continuation tail.

Backtest Review

Sharpe

-0.10

Total return

-18.77%

Max drawdown

52.03%

Trades

1407

Win rate

49.5%

Profit factor

0.98

Implementation looks faithful: point-in-time lagged beta and sigma, next-bar entry, netted BTC hedge, and the 1407 trades are split across both long and short alt legs as the hypothesis describes.

There is a gross edge before costs. Net PnL is about -19k, commission is 56.1k and impact is 59.8k, so gross is about +97k, or roughly +0.30% of notional per trade. That is in the same direction as the pre-study's +0.45% but smaller.

Negative expectancy on a large sample: 1407 closed trades, avg_trade_return_pct -0.059%, PF 0.975, Sharpe -0.105 with a CI of [-0.89, 0.71], PSR 0.45. Lesson 141 applies: a negative Sharpe on more than 100 trades is a completed measurement.

Fee floor comparison: -0.059% against a single-leg Bybit floor of 0.15%. Each pair trades two legs (alt plus a beta-scaled BTC hedge), so the leg-scaled floor (L127) is about 0.25-0.35%. Commission alone is about 0.17% per trade, which leaves at most about 0.13% net even with zero impact. That is still below the floor.

Modeled impact is 146% of gross PnL and capacity is only $46.6k, so the edge exists only at toy scale.

The executed signal has IC 0.028 (t=1.2) and the signal-research diagnostic returned rejected_no_information.

Max drawdown is 52% and the longest drawdown lasted 903 days. Annual returns were 2022 +8.3%, 2023 +27.4%, 2024 -14.7%, 2025 -31.1%, so the recent years are strongly negative.

Events cluster on shared BTC shock dates (L185), so the effective sample is much smaller than 1407.

benchmark_meaningful is false, so no buy-and-hold comparison applies to this market-neutral book.

Outcome Summary

Check bar coverage for every leg over the exact backtest window before a multi-instrument strategy is coded; the hypothesis said LINK had only 1 gap, but the backtest found 5039 missing bars, so the gap check and the backtest window did not agree.

The full backtest failed 3 times with data_unavailable: LINKUSDT.BYBIT-1-HOUR-LAST-EXTERNAL had 5039 missing expected bars, including the requested boundaries. The pipeline marked this as a data failure, not a verdict on the strategy, so backtest review, optimization, analyst and risk stages were never reached.

A BTC-beta-hedged long-short trade on 10 Bybit linear alt perps (1H): when BTC makes a >=2.5-sigma 4h move and an alt fails to follow its trailing beta by >=1.5 residual sigma, bet the divergence keeps widening for 24h (residual continuation after non-participation).

No backtest results exist (the backtest report is empty and no optimization ran). The only performance evidence is the hypothesis pre-study: +0.447% average BTC-hedged 24h residual on 1457 Bybit events (naive t 4.2, overstated by event clustering) and +0.395% on 1834 Binance USD-M events.