Strategies

Cross-Venue Spot Idiosyncratic Down-Shock Rebound, Perp-Hedged: buy a post-2022 Bybit SPOT alt at the close of a day whose 2-day move lagged BTC by <=-2 sigma, short equal-dollar BTCUSDT.BYBIT perp, fixed 3-day clock (14 names disjoint from the promoted pre-2022 Bybit perp panel, 1-DAY, market-neutral event panel, max 2 new pairs per date)

Outcome: Paper Trading

BybitSpotIdioDownShockReboundPerpHedged

Outcome Summary

This strategy tested whether the promoted Bybit perp down-shock rebound template carried over to 14 post-2022 Bybit spot alts hedged with a BTC perp short. The first backtest showed per-event economics far above the ~0.31% two-venue cost, though the backtest review found 2024 carried about 85% of return. After optimization the final parameters (shock_z 2.15, hold 2 days) gave PF 1.82 and positive but non-significant OOS (t 1.41) and holdout (t 1.14), and the largest year carried 45% of PnL. The analyst recommended promotion on iteration 1 and the strategy is now in paper trading as low-confidence research, with survivorship bias, low capital deployment and signs of 2026 decay noted as open risks.

Strategy report

Hypothesis

Original Research Lead hypothesis (not seeded). It tests the long-side idiosyncratic down-shock template (lesson L266) on a different instrument type and a name-disjoint universe. The cash leg is Bybit SPOT and the hedge is the Bybit USDT perpetual, so the book spans two venues (BYBIT_SPOT + BYBIT). The template was promoted on iteration 1 on 29 pre-2022 Bybit USDT perps (185 trades, PF 1.67). Here every alt was listed on Bybit after 2022-01-01, so the universe shares no names with that panel. MECHANISM: when one alt falls far more than BTC over two days without a market-wide move, the excess drop is mostly forced or impatient idiosyncratic selling: a liquidation cluster, an unlock or large-holder exit, or a rumour. Liquidity providers who absorb that inventory demand a premium, and they are paid as the price partly recovers relative to the market over the next few days. Holding the alt in SPOT means the long leg is never liquidated and pays no funding. The BTC perp short removes market beta, as lessons L256 and L267 require for shock panels. EVENT, per alt, at each completed 1-DAY bar (bar timestamp = close, 00:00 UTC): W_alt = ln(close_t / close_{t-2}) on <ALT>USDT.BYBIT_SPOT; W_btc is the same quantity on BTCUSDT.BYBIT; X = W_alt - W_btc. Scale X by the alt's 60-day daily log-return std, measured up to the bar two days before the signal so the shock days are excluded, times sqrt(2): zx = X / (sd60 * sqrt(2)). If zx <= -shock_z (baseline 2.0), open a pair: LONG the alt on BYBIT_SPOT and SHORT an equal-dollar BTCUSDT.BYBIT hedge (hedge_ratio 1.0, pre-registered). Orders are market orders at the signal bar's close. On these 24/7 venues that price is the next bar's open, which is the platform's faithful equivalent of a next-open fill; I name the substitution here. Both legs exit on a fixed 3-day clock. PRE-STUDY (Research Lead, decoded catalog 1-DAY closes: alt from BYBIT_SPOT, BTC from the BYBIT perp, 2022-03 to 2026-10, at least 90 days of bars per name, one open event per name): at zx <= -2.0 with a 3-day hold, 161 alt legs fired on 77 distinct event dates. The mean hedged forward 3-day log return (alt minus BTC) was +2.30% per leg, median +1.33%, hit rate 61%, gross of costs. By year (legs, mean): 2022 3 / +5.1%, 2023 20 / +1.7%, 2024 60 / +2.5%, 2025 52 / +3.6%, 2026 26 / -0.75%. I state up front that 2026 is negative. At zx <= -2.5 the effect is larger (+4.31% mean, 70 legs) but there are only 31 distinct dates, below the ~60-date evidence floor in L271, so the baseline is 2.0. A 5-day hold decays to +0.27% mean (median -0.72%), so the edge is a short rebound and the clock must stay short. That is the main reason hold_days is capped at 4. COST ARITHMETIC: each event pays a spot round trip on the alt (~0.20%) plus a Bybit perp round trip on the equal-dollar BTC hedge (~0.11%), about 0.31% of the alt-leg notional. Against a gross +2.3% (median +1.3%) the capture is about 4-7x costs; a 2x shrinkage to ~1.15% still clears the 2x-cost bar. EVIDENCE COUNT (L238/L271): the effective sample is 77 distinct event dates, not 161 legs. CONCURRENCY (L267): at most 2 new pairs per date (when more names qualify, take the most negative zx) and at most 5 open pairs. Each alt leg is 10% of equity, so same-date gross long is at most 20%, total long at most 50%, and gross book at most 100% with the hedge. Drawdown expected near 10-15%. UNIVERSE (fixed list, point in time): APT, OP, ARB, SUI, SEI, INJ, TIA, PEPE, WLD, JUP, ORDI, LDO, ENA, GMX as <ALT>USDT.BYBIT_SPOT. A name becomes eligible only after 90 days of BYBIT_SPOT bars at that bar (is_listed()/listed_since(), parameters.point_in_time_universe = true). All 14 are still listed. Bybit serves no bars for closed pairs, so delisted spot names cannot be included. That biases the alt leg upward and is named as a risk. Corpus: BYBIT_SPOT is 0.1% of experiments; cross-venue is 10% against a 15% target; long-short is 25%. Distinct from in-pipeline items: BybitSpotAltPanelCapitulationAbsorption4H (unhedged, 4H, raw capitulation), BybitDisjointAbsorbedFlush* (perp, range/close-location trigger), BybitIdioAttentionShockContinuation (UP residual days, follows the move).

Fills the most under-represented buckets at once: BYBIT_SPOT (0.1%), cross-venue (10% vs a 15% target) and long-short market-neutral (25%). It uses the one shock template that keeps promoting on first iteration (the residual-defined, BTC-hedged long down-shock, L256/L266), on a name-disjoint universe. The pre-registered event floor is met: 77 distinct dates at the baseline, which L271 asks for (about 60). Avoids the families flagged by L250/L255/L270 (indicator states, two-name daily spreads) and OKX transfers (L242). Trade-count note: min_trades counts both legs (about 161 alt plus about 161 hedge positions); the avg trade prediction is per position, so the zero-mean hedge legs dilute it. venue_type is futures because the primary instrument and the short hedge are the BTCUSDT.BYBIT perp; the long legs are BYBIT_SPOT cash instruments.

Implementation

Two-venue, market-neutral event panel on 1-DAY bars. The primary instrument is BTCUSDT.BYBIT (the USDT perp): it is both the market factor and the short hedge. The 14 post-2022 Bybit SPOT alts (APT, OP, ARB, SUI, SEI, INJ, TIA, PEPE, WLD, JUP, ORDI, LDO, ENA, GMX) are the long legs. At each daily close the strategy computes, for every alt, zx = [ln(C_t/C_{t-2}) - ln(B_t/B_{t-2})] / (sd60 * sqrt(2)). sd60 is the alt's daily log-return std over the 60 bars ending at t-2, with at least 75% of those bars present. An alt is eligible only after 90 days of spot bars and listing (is_listed/listed_since, point-in-time). When zx <= -shock_z, the strategy buys the alt on BYBIT_SPOT at 10% of equity and shorts BTCUSDT.BYBIT for hedge_ratio x the alt notional actually sent. Candidates are taken most-negative-zx first, at most 2 new pairs per date and 5 open pairs. Both legs exit after exactly hold_days daily bars. Market orders at the signal bar's close are the platform's 24/7 equivalent of a next-bar-open fill (capability next_bar_open_fill = use_equivalent).

Follows the promoted BybitTwoDayIdioDownShockReboundBtcHedged template. The long leg moves to BYBIT_SPOT, plus the per-date cap and the hedge-follows-fill rules. (1) An alt contributes only when its last bar is stamped exactly at the primary bar's timestamp (L139 alignment), and both ends of the 2-day move must sit on the daily grid. (2) If the spot BUY is refused at submission (no new in-flight entry recorded), that pair gets no hedge and no book row. Otherwise the hedge is sized from the quantity actually sent after the base caps. On later bars, if the filled notional (qty x avg_px_open) differs from the ledger by more than 5%, the pair's hedge quantity is rescaled. (3) The BTC short is one NETTING position equal to the sum of a per-pair ledger, kept as a persisted scalar string. The hedge opens from flat via should_enter and closes to flat via should_exit; resizes go through one direct order per bar. (4) legs_balanced() over the open alts plus the BTC perp runs every bar; on False, every leg is flattened. A delisted or flat alt leaves the book and its hedge quantity is bought back. (5) Risk route (b): fixed time exit with capped sizing. Each alt is 10% of equity, each hedge at most 10%, the net BTC short at most 50%, and gross at most 1x; no position is near 100% of equity. Sizing reads get_equity_mtm() of the primary (perp) account. Each venue is a separate pre-funded account, so 10% of perp equity is about 5% of the summed two-venue capital. Leverage is 1.0 and not used in sizing, as the plan fixes. fill_tolerance (0.05) is an implementation constant, not a trading rule. Survivorship: Bybit serves no bars for delisted spot pairs, so the fixed list holds only survivors, as the hypothesis states.

Verification Results

No change is needed for measurement. The analyst should read the size as 10% of one venue account.

Sizing reads get_equity_mtm() of the PRIMARY (BYBIT perp) account: `notional = equity * self.leg_notional_pct`. The runner funds BYBIT and BYBIT_SPOT each with their own starting balance, so each alt leg is 10% of the perp account (about 5% of combined capital), not 10% of total equity. The structure (10% per leg, 2/date, 5 open, 1:1 hedge) matches the contract; only the reference equity is per-account. The developer discloses this.

For paper, skip rows whose alt has an open/in-flight entry order (check cache.orders_open) before dropping them.

Book integrity drops any row whose alt has no open position, then buys its hedge back. In the backtest, market orders fill in-event, so this is correct. In paper, a spot BUY still in flight on the next bar would be read as 'never filled' and its hedge would be unwound. legs_balanced() returns True while a market order is in flight, but the live_rows filter runs before it.

Acceptable as is. Check order_failures in the full backtest report; if it is nonzero, consider passing the specific pair's legs only.

legs_balanced() measures leg failures from the earliest ts_opened among the open legs. The netted BTC short can stay open across many overlapping pairs, so a single refused or shrunk hedge-resize order flattens the whole book on the next bar, including pairs opened earlier. This is conservative and keeps the hedge honest, but in a busy period it could flatten good pairs early.

Optionally cache per-alt log returns incrementally.

Runtime: the sandbox took 142 s for 2,305 daily bars (runtime_review_required). Each bar recomputes the 60-return std for all 14 alts by slicing the buffer. That costs about 14x61 operations per bar, which is cheap, so most of the time is probably engine and multi-venue overhead. This is a resource concern, not a correctness one.

The pre-study's 2026 cell is negative (26 legs, -0.75%). The 14-name list holds only survivors, which biases the long leg upward. The effective sample is about 77 event dates, mostly from 2024-2025. Also note that the hedge legs (BTC perp) are counted as trades, so avg_trade_return_pct is diluted roughly 2x relative to the per-event alt-minus-BTC capture. Judge the economics per event (alt leg PnL + its hedge) against the ~0.31% two-venue round trip. Per lesson L211, the deployed size is about 10% of the perp account's equity per alt leg. Each venue starts with its own balance, so that is about 5% of the combined capital. Total return and drawdown will therefore look small relative to the per-event capture.

Backtest Review

Sharpe: 0.53
Total return: 12.20%
Max drawdown: 2.76%
Trades: 175
Win rate: 53.7%
Profit factor: 1.59

The trades implement the stated mechanism. There are 78 alt legs, all LONG on BYBIT_SPOT across all 14 names (GMX 10 down to JUP 2). The hedge is one netted SHORT on BTCUSDT.BYBIT perp, which is why there are 42 hedge positions. Every hold is 3 days, and the sizes match the 10%-of-equity alt leg with an equal-dollar hedge.

Per-event economics clear costs by a wide margin. The alt legs made +$22.7k and the BTC hedge lost -$6.4k (BTC rose over the period), so the pair book netted +$16.3k over 78 events. That is about $209 per event on about $10k of alt notional, roughly +2.1% per event after fees and funding, against a two-venue round trip of about 0.31%. The position-level avg_trade_return_pct is 1.59%, which is diluted by the hedge legs, against a 0.25% BYBIT_SPOT floor.

Profit factor 1.57, win rate 0.56, Sharpe 0.44, max drawdown 2.76%. Total return of 8.15% on the two-venue 200k base reconciles with the +$16.3k ledger.

Only 2 parameters are tunable (shock_z and hold_days) over narrow, pre-registered ranges, so the optimization search is small.

The sample is thin for the validation design. There are 78 events on 56 distinct entry dates, against the roughly 60 that asks for and the roughly 60 validation trades that 3 windows of 20 need. The walk-forward may end as 'not enough trades to validate' or use few, long windows. Even then, that outcome still produces final parameters and a holdout.

Profit is concentrated in one year. By year (alt legs, alt-leg PnL): 2022 3 / +$3.1k, 2023 19 / -$0.8k, 2024 39 / +$19.1k, 2025 (through the cut) 17 / +$1.3k. 2024 carries about 85% of total return (6.89 of 8.15 points), so the soft floor of no more than 60% in one year fails. 2023 was negative, and 2025 is -0.14% at the portfolio level.

The universe holds survivors only, which biases the long leg upward, as the hypothesis itself states.

Capital is barely used: about 10% of one venue's equity per event, so about 5% of combined capital. A 2.8% drawdown therefore says little about risk at a larger allocation. tail_ratio is 0.14 and cvar_95 is 0.32%.

QA's concern (survivors, diluted per-trade figure, thin event count) is partly borne out by the concentration and the 56 dates. It is not borne out on per-event economics, which are well above costs on the development period.

Analysis

No hard failures and no soft failures in numeric_gates. The evidence is inconclusive, not adverse.

Out-of-sample evidence is positive. Pooled outer OOS over 1,432 daily observations: Sharpe 0.63, HAC t 1.41 against a hurdle of 2.0 (N_oos = 1, so DSR adds no correction; p = 0.08). The holdout, 2025-07-08 to 2026-11-02 (482 days, 40 trades, not underpowered), has Sharpe 0.87, t 1.14 against 2.0, a ratio of 1.38 to WF-OOS and z_vs_oos +0.20, so it is consistent with OOS. Neither test is statistically adverse.

Economics clear costs by a wide margin. The final-parameter full-history replay (shock_z 2.15, hold_days 2) has 127 closed positions, PF 1.82, an average of 1.76% of notional per position against the 0.22% floor, win rate 0.54 with a 1.58 payoff ratio, and +9.9% total return. Cost stress: one event costs about 0.31% of alt notional (spot round trip plus perp hedge). Raising fees and impact by 50% takes that to about 0.47%, against a gross capture of about 2%, so expectancy stays well positive. PF is far above the thin-PF zone.

The trades implement the hypothesis, as confirmed at backtest review: long BYBIT_SPOT alts on a fixed clock with a netted BTC perp short, and no order failures. End-unrealized is 0%.

Risk is small. Max drawdown is 2.96%, worst realised loss is 0.77% of equity per trade, the largest position is 19.3% of equity (the netted BTC hedge), and peak margin use is 0.40. The risk_control route is signal/time exit, with no violations. Calmar is 0.51.

The result replicates the promoted pre-2022 Bybit perp down-shock panel (PF 1.67, 1.75% per trade) on a name-disjoint, post-2022 SPOT universe. That is independent corroboration of the same mechanism.

Sensitivity is clean: no cliffs, and every parameter stays inside the one-SE noise band (shock_z from 1.5 to 3.0 gives 0.27-0.53). 4 of 5 measured years are positive, and the largest year carries 45% of PnL.

Significance is not established at either scope, so this is low-confidence paper only and not proof of an edge.

Signs of decay. Walk-forward window 4 (2024-09 to 2025-07) had OOS Sharpe -0.03, and the final parameters on it gave -0.24 (partly in-sample). The pre-study's 2026 legs averaged -0.75%, and the 2026 calendar year in the initial backtest is -0.43%. The holdout's positive Sharpe leans on 2025-H2 (October 2025 +2.6%).

Parameter stability is unstable for hold_days: walk-forward windows chose 4, 3, 2, 4 and 2, and the final value is 2 against the pre-registered baseline of 3. shock_z is moderate (2.09-2.33). Sensitivity puts hold 2 and hold 3 within noise (0.48 against 0.44), so the deployed clock is a coin-flip choice inside a plateau, not a sharp optimum.

Capital is barely deployed. Implied deployment is total_return / (trades x avg_trade_return_pct) = 9.89 / (127 x 1.764), about 4.4% of equity per position. The hedge positions, whose return is near zero, dilute the average, and each alt leg is 10%. Treat the 2.96% drawdown and 2.8% annual volatility as figures at 10% legs, not as proof of low risk. At this sizing CAGR is only about 1.5%.

Pre-registered predictions were missed on trade count (127 against 250) and return (9.9% against 40%). The miss comes from sizing and per-date caps, not from per-trade edge: 1.76% per trade against a predicted 0.5% or more.

Survivorship: only currently listed Bybit spot names are in the universe, so the alt leg is biased upward. The bias size is unmeasured.

The signal-research diagnostic shows IC 0.012 (t 0.28). That is an execution-path diagnostic, and the state-dependent forecast is noted as unreliable, but it does not support the edge.

inconclusive; consistent_with_oos z=+0.20

unevaluated in backtest metrics; the registered pooled-OOS DSR/PSR is 0.92 against 0.977

advisory only

inconclusive, not adverse

PAPER ONLY, labelled 'low confidence: not statistically proven'. Pooled OOS t is 1.41 and holdout t is 1.14, both against 2.0. Neither is adverse, and the holdout is consistent with OOS. Trade the final parameters (shock_z 2.15, hold_days 2, all other values as frozen). Keep the 10% alt leg as the research unit, at reduced allocation. Disclosed risk figures: worst realised loss 0.77% of equity per trade, largest position 19.3% of equity, implied deployment about 4.4% of equity per position, max drawdown 2.96% at this sizing. These figures are not comparable with drawdown floors for a fully deployed book. Pre-registered paper kill criteria: (1) Re-evaluate after 30 closed alt-leg pairs, about 9-12 months at the observed rate of roughly 3 a month. Retire if the mean hedged net return per pair is below +0.5% of alt notional, or if PF is below 1.1 on that sample. (2) Retire early if paper drawdown exceeds 6% at the 10%-leg sizing, which is about 2x the backtest maximum. (3) Cost check: if realised round-trip cost per pair (spot fees plus spread plus BTC hedge, funding included) averages above 0.5% of alt notional, re-score the backtest expectancy with that cost before continuing. (4) The decay watch is explicit. Window 4 and calendar 2026 were flat to negative, so a first 15 pairs with a negative mean should trigger an early review, not a wait for 30. This sits close to the promoted pre-2022 Bybit perp down-shock panel and is likely correlated with it on market-wide alt-selloff dates. The portfolio manager should treat the two as one mechanism for allocation.

Risk Review

Correlation with the deployed book. This is the same idiosyncratic down-shock rebound mechanism as several paper members: BybitTwoDayIdioDownShockReboundBtcHedged, HlIdioDownShockReboundBtcHedged, SpotLedIdioAltSelloffReboundBtcHedged and AsxGoldMinerIdioDownShockRebound/TsxMinerIdioDownShockRebound (other markets), plus the Bybit spot and perp capitulation/flush panels. The universe has no names in common with the pre-2022 perp panel, but the events cluster on the same market-wide alt sell-off dates. Each also carries a short BTC perp hedge, so the book's net BTC short adds up across these strategies. Every member is still collecting (no breaches, no open items), so none of this correlation has been measured live yet.

Signs of decay. The last development walk-forward window (2024-09 to 2025-07) is negative: OOS Sharpe -0.03, and the final parameters give -0.24. The 2025 development year is -0.65%, and the pre-study's 2026 cell is negative (-0.75% per leg). The holdout is called consistent but remains inconclusive. Significance is not established in pooled OOS (t 1.41 against a 2.0 hurdle) or in the holdout, so this is low-confidence paper only.

Paper fill-timing hazard (QA, line ~214). The book-integrity step drops any row whose alt has no open position and buys back that row's BTC hedge. In paper, a spot BUY that is still in flight on the next daily bar would be read as never filled, and its hedge would be unwound. The result is an unhedged alt long. A related point: one refused hedge resize trips legs_balanced() and flattens every open pair. Neither occurs in the backtest, where market orders fill within the bar.

Sizing references only the perp account. leg_notional_pct (10%) is applied to get_equity_mtm() of the BYBIT perp account, but the alt legs trade from a separately funded BYBIT_SPOT account. Each alt leg is therefore about 5% of combined capital. In paper and live, the two accounts' equity can drift apart, and the alt's size would follow the perp wallet rather than the spot wallet that pays for it.

Risk control route (b) holds: a fixed hold_days clock exit (87 closed trades show it firing) with capped sizing. The largest measured position is 19.3% of equity, the netted BTC hedge (limit 100%). The worst realised trade loss is 0.77% of equity and the worst adverse excursion 1.71%, both at 1x leverage. The liquidation buffer is about 64 daily moves against the 8 required. Having no stop-loss is not an issue on this route.

The capital is barely used. Realized volatility is 2.8%/yr, below the conservative profile's 5% band floor, and peak gross exposure is low. The 2.96% max drawdown against the 15% pre-registered limit reflects this small sizing, not proof of low risk per event. The return and trade-count predictions were missed (127 trades vs 250 predicted; 9.9% return vs 40%) because of sizing and the per-date caps. Per-trade edge met its prediction (1.76% vs 0.5%).

Parameter stability: the walk-forward windows chose hold_days 4, 3, 2, 4 and 2 (unstable), and the final value of 2 is below the pre-registered baseline of 3. Sensitivity puts hold 2 and hold 3 inside the noise band, so this is a choice within a plateau, not a cliff. shock_z is moderately stable (2.09-2.33). Both parameters are in safe ranges.

Survivorship bias: all 14 alts are currently listed Bybit spot names (Bybit serves no bars for delisted pairs), so the long leg is biased upward by an unmeasured amount. Market impact is small: legs are about 5% of capital, at most 2 new pairs per day, on liquid spot alts and the BTC perp.

Soft floors: none are in play, so no waivers are needed. The advisory in-sample Sharpe floor (1.5) fired; it is advisory and non-blocking. deflated_sharpe is unevaluated in the backtest metrics and is treated as unknown, not as passed.

Paper only, with the low-confidence label (not statistically proven). This review is not evidence for real money.

Before paper, or as the first paper fix: the book-integrity step must skip rows whose alt has an open or in-flight entry order (cache.orders_open) rather than dropping them and unwinding their hedge.

Size the alt leg from the BYBIT_SPOT account's equity (or the combined equity), not the perp account's, so that paper and live sizing follow the wallet that pays for the leg.

The Portfolio Manager should treat this as one sleeve with the other idiosyncratic down-shock rebound strategies already in paper, not an independent allocation. It should watch the combined net BTC perp short across those hedged strategies.

Outcome Summary

The idiosyncratic down-shock rebound held on a name-disjoint spot universe, but significance needs more distinct event dates; watch for decay (walk-forward window 4 and 2026 were negative) and an unstable hold_days choice, and size up before reading the 2.96% drawdown as low risk.

Backtest review sent it to optimization and the analyst recommended promotion after one iteration: no hard or soft gate failures, economics well above costs, and OOS/holdout positive but not statistically significant. It is now in the paper stage as low-confidence, not statistically proven.

Buy a post-2022 Bybit SPOT alt (14 names that share none with the promoted pre-2022 perp panel) after its 2-day move lagged BTC by at least 2 sigma, short an equal-dollar BTCUSDT.BYBIT perp as a hedge, and exit on a fixed short clock, capturing the rebound after idiosyncratic forced selling.

The initial full-history backtest returned 12.2% (Sharpe 0.53, PF 1.59, 175 closed positions, max drawdown 2.76%, avg 1.68% of notional per trade). The final-parameter replay (shock_z 2.15, hold_days 2) returned +9.9% on 127 positions with PF 1.82 and 1.76% per position; pooled OOS Sharpe was 0.63 (HAC t 1.41) and holdout Sharpe 0.87 (t 1.14, 40 trades), both below the 2.0 significance hurdle.

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