Strategies

Bybit Two-Day Idiosyncratic Down-Shock Rebound, BTC-Hedged: long a Bybit alt perp + short equal-dollar BTCUSDT.BYBIT after its 2-day move lags BTC by <=-2.5 sigma, fixed 3-day clock (29 pre-2022 Bybit USDT perps, 1-DAY, market-neutral event panel, flat most of the time)

Outcome: Paper Trading

BybitTwoDayIdioDownShockReboundBtcHedged

Outcome Summary

This was an original Research Lead hypothesis: a BTC-hedged, market-neutral panel that buys Bybit alts after an idiosyncratic 2-day drop of at least 2.5 sigma relative to BTC and holds for a fixed few days. It passed backtest review with +26.3% over 185 trades but weak risk-adjusted numbers (Sharpe 0.45), and profit was concentrated in 2022 and 2024. Optimization moved to a stricter threshold (2.84) and a 4-day hold, which gave PF 1.93, a 6.3% drawdown and a positive but non-significant holdout (Sharpe 0.95, t 1.28). The analyst promoted it after one iteration and it is now in paper trading. Its statistical significance is still inconclusive, with survivorship bias, funding carry and regime clustering noted as open risks.

Strategy report

Hypothesis

Original Research Lead hypothesis (not seeded). 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 in that coin's perp, a large holder exiting, or a rumour. Liquidity providers who absorb that inventory ask for a premium, and they are paid as the price partly recovers relative to the market over the next few days. The design takes the long side of that absorption, which is the side that has worked in this factory (L230), and defines the trigger as a BTC residual with a BTC hedge leg, as L256 requires. It is distinct from the promoted Bybit volume-share capitulation (1H, 24h, needs a venue-share condition), the Bybit weekend pump reversal (the SHORT side of upside residual jumps, Mondays only), the in-pipeline Bybit attention-shock continuation (UP residual days, follows the move) and the OI-flush absorption (unhedged, raw-return trigger). EVENT, per alt, at each completed Bybit 1-DAY bar (bar timestamp = close, 00:00 UTC): W_alt = ln(close_t / close_{t-2}), W_btc the same for BTCUSDT.BYBIT, X = W_alt - W_btc. Scale it by the alt's 60-day daily log-return std measured up to the bar two days before the signal (excluding the shock days), times sqrt(2): zx = X / (sd60 * sqrt(2)). If zx <= -2.5, open a market-neutral pair: LONG the alt perp, SHORT an equal-dollar BTCUSDT.BYBIT hedge (hedge ratio fixed at 1.0 dollar per dollar, pre-registered, as in the pre-study). Orders are market orders at the signal bar's close; on this 24/7 venue that price is the next bar's open, which is the platform's faithful equivalent of a next-open fill. Exit both legs on a fixed 3-day clock. PRE-STUDY (Research Lead, decoded Bybit 1-DAY catalog bars, 2020-10 to 2026-10, the 29 names in instruments, non-overlapping per name): zx<=-2.5 with a 3-day hold fired 233 alt legs on 93 distinct event dates (~16 event dates a year). Mean hedged (long alt minus BTC) 3-day forward log return was +3.31% per leg, median +2.33%, gross of costs. By year (legs, mean): 2021 7 legs -5.2%, 2022 40 legs +4.7%, 2023 51 legs -0.1%, 2024 66 legs +4.5%, 2025 33 legs +9.8%, 2026 35 legs +0.2%. So 2021 and 2023 were flat or negative and 2026 is flat; the edge is concentrated in a few years, which I state up front. Neighbouring settings: zx<=-2.0 gives 457 legs on 178 dates at +1.92% (3-day); a 2-day hold at -2.5 gives +2.41%; a 5-day hold +2.56%. The pre-study also tested the mirror of the promoted weekend pump reversal (buy WEEKEND residual dumps, Monday-stamped signals): it does NOT rebound (-0.30% hedged at zx<=-1.5, -1.03% at -2.0, 2-day). Weekend dumps are therefore not excluded by the rule (that would be selection on this sample), but Monday-stamped events are reported as a separate diagnostic. COST ARITHMETIC: each event pays two Bybit round trips (alt + equal-dollar BTC) at ~0.11% each = ~0.22% of the alt-leg notional; against a gross +3.3% (median +2.3%) the capture is ~10x costs, and even a 3x shrinkage to ~1.1% clears the 2x-cost bar. EVIDENCE COUNT (L237/L238): 93 distinct event dates, not 233 legs, is the effective sample. UNIVERSE: fixed list of 29 Bybit USDT perps all with continuous Bybit bars from before 2022-01-01; at each signal a name is eligible only if it has >= 90 days of Bybit bars at that bar (point-in-time via is_listed()/listed_since()). Residual survivorship: Bybit serves no bars for closed contracts, so coins delisted from Bybit before today are missing; this biases the alt leg upward and is named as a risk. Venue: BYBIT is 3.4% of the corpus against a >=5% target; the design is long-short market-neutral (direction bucket 25% long-short).

Implementation

Market-neutral event panel on Bybit USDT perps using daily (1-DAY) bars. At each daily close t, each of 20 alt perps listed before 2022 is checked. A name is eligible only after at least 90 days of bars, judged as of that date with is_listed/listed_since. The score is zx = (ln(C_t/C_{t-2}) - ln(BTC_t/BTC_{t-2})) / (sd * sqrt(2)), where sd is the sample std of the alt's daily log returns over the vol_lookback (60) bars ending at t-2, so the two shock days are left out. zx <= -shock_z (2.5) opens an event: a market BUY of the alt at 10% of equity and a market SELL of an equal-dollar BTCUSDT.BYBIT hedge (hedge_ratio 1.0). The most negative zx goes first, with at most 5 events open and one event per alt at a time. The BTC short is one netted position equal to the sum of the open events' own hedge quantities, tracked per event in a saved scalar string. Each event exits after exactly hold_days (3) daily bars: the alt is sold and that event's BTC quantity is bought back. There is no stop or target. legs_balanced() is checked every bar, and any refused, shrunk or orphaned leg flattens all legs. Execution substitution (next_bar_open_fill = use_equivalent): orders are market orders at the signal bar's close, which on this 24/7 venue is the next bar's open price. Events stamped on a Monday (moves measured over the weekend) are counted in entry_diag['monday_stamped_events'] as a diagnostic only.

Implements the hypothesis rules as written, using the parameter names from the optimization plan. Tunable: shock_z, hold_days, vol_lookback. Fixed: leverage, hedge_ratio, max_concurrent, leg_notional_pct, min_listing_days, move_window_days. The universe is the 20 instruments in the hypothesis's instruments list. The description mentions 29, but the declared list (which Layer 1.5 checks) has 20, and all 20 have Bybit 1-DAY bars in the catalog. The primary BTCUSDT.BYBIT is the hedge and market factor; point_in_time_universe=True. Risk route (b): time exit with capped sizing, each leg 10% of equity and at most 100% gross. At leverage 1.0 the initial-margin check cannot fit a full 5-event book (100% gross) plus fees. Gross is therefore capped at gross_cap_frac=0.95 x equity: only a 5th concurrent event is shrunk slightly, so the BTC hedge is never refused for margin (a refused hedge would flatten the book via legs_balanced). The sd window needs at least 75% of its daily returns to come from consecutive days (min_obs_frac); otherwise the event is skipped as 'sd missing'. An alt bar is used only when it is stamped exactly at the primary decision bar, so there is no look-ahead. The structure reuses the mechanics of BybitIdioAttentionShockContinuationBtcHedged, which passed verification: a netted BTC hedge and a saved scalar book.

Verification Results

Keep it as disclosed, or state in the description that the 5th slot is shrunk to fit margin.

gross_cap_frac=0.95 is not in the frozen contract. The contract allows 5 events x (10% alt + 10% BTC) = 100% gross. With the 95% cap, a 5th concurrent event gets about 7.5% alt notional instead of 10%, not 'slightly' less. The developer disclosed the reason (initial margin at leverage 1.0 plus fees) and it only applies when the book is fully loaded, so it is not a structural mismatch.

Optional: allow re-entry on the exit bar to match 'non-overlapping per name' exactly.

An alt whose event exits on bar t cannot open a new event on that same bar (the 'exited' set excludes it). The hypothesis only forbids re-entry while an event is open, so this is a small extra restriction. It removes a few back-to-back events.

Check the backtest's order_failures count. If flatten-alls show up often, scope the failure check to each event's own legs.

legs_balanced() looks for leg failures since the earliest open ts_opened, and that includes the netted BTC short. The BTC short can stay open across many overlapping events, so one refused order early in that stretch can flatten the whole book on a later bar. This is conservative (it flattens, it never leaves the book unhedged), so it is not a correctness defect.

None needed. The analyst should scale the expected event count to 20 names.

The hypothesis text says 29 names, but research_contract.instruments (the frozen contract) lists 20. The config faithfully trades those 20 and says so. The pre-study event counts (233 legs, 93 dates) were measured on 29 names, so expect fewer events.

In the pre-study the edge sits in 2022, 2024 and 2025 (2021, 2023 and 2026 were flat or negative), and the effective sample is about 93 event dates. Walk-forward windows may disagree and HAC significance may come out inconclusive. Funding on the long-alt/short-BTC pair is not modelled in the signal. Survivorship (no bars for delisted Bybit contracts) biases the alt leg upward.

Backtest Review

Sharpe: 0.45
Total return: 26.26%
Max drawdown: 10.83%
Trades: 185
Win rate: 55.1%
Profit factor: 1.67

The trades match the hypothesis. All 85 alt legs are LONG and all 35 BTCUSDT legs are SHORT. Each hedge is sized equal-dollar to the alts it covers: 1 event gives about a $9.7k short, 2 events about $20.0k, 3 events about $29.2k. Every hold is exactly 3 days. Entries are relative lags against BTC, including dates when BTC was rallying (for example 2024-02-29), which is what the residual trigger defines.

The economics are well above the Bybit floor. Mean return is 1.69% per trade against a 0.15% review floor, profit factor is 1.50 and total return is +14.3%. P&L reconciles: the alt legs made +$18.6k, the BTC hedge lost -$4.3k, and net funding added +$1.9k.

Risk is contained. Maximum drawdown is 10.8%, under the 20% pre-registered limit, and gross exposure is capped at 95% of equity. 120 closed positions on 49 distinct event dates is enough for trade-count-aware windows.

Risk-adjusted performance is weak. Sharpe is 0.33 (CI -0.65 to 1.10), Calmar is 0.25 and the longest drawdown lasted 521 days. The profit is concentrated in two years: 2022 made +7.4% and 2024 made +13.6%, while 2021, 2023 and 2025 were negative. This matches the regime concentration both the pre-study and QA's note flagged.

The largest single win, +$7.3k (2022-05, LUNA period), is about half of the net profit. The worst alt leg was -35.5% on LINK in 2021-05.

There are fewer events than the pre-study found: 49 dates in about 4 years versus about 16 a year claimed. Part of the reason is that the contract lists 20 names, not 29. The effective sample is thin, so a significance result of inconclusive is the likely outcome.

Analysis

No hard failures and no soft failures in numeric_gates. The only inconclusive item is 'not enough trades to validate': 2 of 5 walk-forward windows reached 20 trades, and 3 were required.

Economics of the final parameters (shock_z 2.84, hold_days 4) replayed over all history: 113 closed positions, mean net trade return 2.23% of notional against the 0.22% two-leg Bybit floor (about 10x), profit factor 1.93, win rate 0.575 with payoff 1.43, max drawdown 6.3% against the pre-registered 20%, Calmar 0.58. The baseline (shock_z 2.5, hold 3) gave similar results: 185 trades, 1.75% per trade, PF 1.67, max drawdown 10.8%.

Holdout (2025-07-05 to 2026-10-29, 481 days, 38 trades, not underpowered) with the final parameters: Sharpe 0.95, max drawdown 3.4%, t = 1.28 (hurdle 2.0), so not significant but positive. It is consistent with walk-forward out-of-sample (z = +0.58), and the ratio is above the 0.7 floor.

Walk-forward out-of-sample: 4 of 5 windows positive (0.45, 0.40, -1.29, 0.66, 0.51). Each window re-chose shock_z in the 2.42-2.92 range and hold_days of 4-5, consistently above the baseline. Parameter stability is moderate overall and nothing is unstable.

Sensitivity shows no cliffs. The shock_z x hold_days heatmap rises smoothly toward stricter thresholds and holds of 3-4 days, and hedge_ratio, vol_lookback, max_concurrent and min_listing_days are all flat within one Sharpe standard error (0.43).

The trades implement the hypothesis: 123 long alt legs with equal-dollar BTC shorts, exits after hold_days, and no leftover open positions at the end of the backtest (end_unrealized 0%).

The same mechanism has promoted siblings that use a long-alt / short-BTC residual absorption design: the Hyperliquid down-shock rebound (PF 1.08) and the Bybit volume-share capitulation (PF 1.26). This version clears the fee floor by a much wider margin than either.

Risk control, measured from the trade ledger: worst realised loss per trade is 2.92% of equity (a BTC hedge leg entered 2022-05-12), and the largest position is 47.5% of equity (BTC hedge, 2025-10-11), under the 100% cap. Route (b) applies, a time exit with capped sizing, with no violations. Leverage is 1.0 and the minimum liquidation buffer is 22.8 daily moves.

Pooled walk-forward out-of-sample Sharpe is only 0.235 (t = 0.48) over 1,425 days, so the edge is not statistically established. The sample is about 113 events on far fewer distinct dates.

The edge comes in clusters: 2022 carries 54.9% of the profit, just under the 60% floor. May 2022 (+10.6%), March 2024 (+8.1%) and October 2025 (+6.6%) dominate. 2021 and 2023 were negative, as the hypothesis said in advance. Window 2 (2023) out-of-sample Sharpe was -1.29.

Funding credited a net +4,518 USD, about 17% of the baseline profit. Part of the result is therefore long-alt/short-BTC carry rather than relative price recovery, and that carry can turn negative in euphoric regimes.

Survivorship: Bybit serves no bars for delisted contracts, so alts that died are missing from the long leg. This biases the result upward.

Missed pre-registered predictions: 113 trades against the predicted ≥200, and +25.7% return against the predicted ≥40%. The predictions met were profit factor, win rate, drawdown and per-trade return. These misses are evidence, not floors.

The signal-research IC diagnostic rejected the forecast (IC 0.028, t = 0.6). That diagnostic depends on position state and is not a gate, but it does not add support.

One order was shrunk for margin: a BTC hedge leg in June 2023 was capped at 1.32 instead of the requested 1.47 BTC. This left a small temporary hedge shortfall, which does not change the result materially.

Deflated Sharpe was not evaluated because it is missing from the backtest metrics. Treat it as unknown, not as passed.

inconclusive (t 1.28 vs 2.0; consistent_with_oos true)

positive, significant

unevaluated

advisory only

inconclusive (t 0.48)

HAC t >= 2

This is a low-confidence paper observation, not evidence for real money. It is not statistically proven: pooled out-of-sample t = 0.48 and holdout t = 1.28. The economics are strong and robust across parameters, though: 2.23% per trade against a 0.22% floor, profit factor 1.93, max drawdown 6.3%, no cliffs, and a positive holdout consistent with walk-forward. The verdict judges the Research Lead's own operationalization of the idea, and promoted siblings use the same long-alt / short-BTC absorption mechanism. Trade the final parameters: shock_z 2.84, hold_days 4, 10% per leg, at most 5 concurrent events, leverage 1.0. Use a reduced paper allocation because events cluster on a few dates. Kill criteria, set now so paper results can falsify the idea: (1) after at least 25 paper events, or 18 months, whichever comes first, mean net trade return must be ≥ 0.6% of alt-leg notional and profit factor ≥ 1.2, otherwise retire; (2) retire immediately if paper drawdown exceeds 12% (about twice the backtest maximum); (3) report net PnL split into price and funding components; if funding is more than half of the paper PnL, re-review, because the premise is relative price recovery, not carry. Expect long flat periods, about 1.5 events a month on average.

Risk Review

Risk control holds on route (b): a fixed holding-time exit (hold_days = 4) with capped sizing at 1x leverage. Measured from the trades: the largest position is 47.3% of equity (the netted BTC hedge short, 2023-10-24), under the 100% cap. The worst realised trade loss is 2.92% of equity (BTC hedge leg entered 2022-05-12, -6.2% of notional on a 47% position). The worst adverse excursion is 7.45% of equity. The minimum liquidation buffer is 46 daily moves against the 8 the conservative profile requires. There is no stop-loss and none is needed on this route.

Drawdown profile is acceptable. Max drawdown over the development period is 6.32% (bootstrap CI 3.5-13.1%) at 1x, well inside the hypothesis's pre-registered 20%. Realized volatility is 6.8% a year, inside the profile's 5-60% band. The longest drawdown lasted 783 days because the strategy is flat most of the time, so expect long flat stretches rather than a sudden cliff.

Correlation with the deployed book. Five paper strategies already trade the same long-alt / short-BTC absorption mechanism: HlIdioDownShockReboundBtcHedged, BybitVolumeShareCapitulationBtcHedgedRebound, SpotLedIdioAltSelloffReboundBtcHedged, BybitFlushAbsorptionPanel and HyperliquidDailyDownShockAbsorptionPanel. Their events will cluster on the same alt-crash dates (e.g. May 2022, which alone made 13% here). The diversification this adds is small, and in a crash the book's combined long-alt exposure is larger than any one member's budget suggests. Three of those siblings currently show a 'diverged' decision replay, though none is in breach.

The edge is thin and concentrated. The final parameters were not statistically accepted: there were not enough trades to validate (2 of 5 windows reached 20 trades). Pooled out-of-sample Sharpe is 0.24 (t = 0.48). 2022 carries 55% of the profit, and 2021, 2023 and 2025 lost money. The strategy missed two of its own pre-registered predictions: 113 trades against ≥200 and 25.7% return against ≥40%. These numbers support low-confidence paper observation at a small allocation, not any real-money claim.

Part of the profit is funding carry (long alt / short BTC), not relative price recovery, and that carry can turn negative in euphoric regimes. Survivorship also biases the long leg upward, because Bybit serves no bars for delisted contracts.

The book can flatten itself on any leg failure. legs_balanced() checks failures since the earliest open position, and that includes the long-lived netted BTC short, so one refused order can close every leg. This errs on the safe side (the book is never left unhedged). One BTC hedge order was shrunk for margin in June 2023 (1.32 instead of 1.47 BTC) because gross exposure was capped at 95% of equity. Trade frequency is low (about 75 closed positions over 5 years of development data, around 16 event dates a year), so fees are not a concern but evidence builds slowly.

Market impact: each alt leg is 10% of equity, there are at most 5 concurrent events, and gross exposure stays at or below 95% of equity. On liquid Bybit USDT perps (including the smaller AXS, SAND and MANA) this is negligible at platform capital. Measured impact cost is 2.4% of gross, against a 50% limit.

Live book: there are no members in breach and no open execution items. Many members are still collecting data, and several replays show 'diverged'. None of this blocks adding a paper strategy.

Paper only, with the persistent label 'low confidence: not statistically proven'. This is not real-money evidence.

Count this strategy and the five deployed long-alt/short-BTC absorption siblings as one correlated cluster in capital allocation. Do not size it as an independent sleeve.

Keep leverage at 1.0, leg_notional_pct at 0.10, max_concurrent at 5 and gross_cap_frac at 0.95 as backtested.

Outcome Summary

The edge came in clusters (2022 made 54.9% of profit, with 2021, 2023 and 2025 weak or negative), and about 17% of baseline profit was long-alt/short-BTC funding carry. Paper monitoring should separate relative-price recovery from carry and check the result across regimes.

The analyst recommended promotion: there were no hard or soft gate failures, the per-trade economics were about 10x the 0.22% two-leg fee floor, 4 of 5 walk-forward windows were positive, and the holdout was positive and consistent with OOS. The run is now in paper trading. Its significance is inconclusive (too few trades for validation windows, deflated Sharpe not evaluated), so the result is not statistically proven.

When a Bybit alt perp's 2-day move lagged BTC by at least 2.5 sigma, the strategy bought the alt and shorted an equal-dollar amount of BTCUSDT.BYBIT, then closed both legs on a fixed clock. The idea was to earn a rebound premium for absorbing forced, coin-specific selling.

The full-history baseline (shock_z 2.5, hold 3 days) returned +26.3% over 185 trades, averaging 1.75% per trade, with Sharpe 0.45 (CI -0.37 to 1.12) and a 10.8% max drawdown. The final parameters (shock_z 2.84, hold 4 days) made 113 trades averaging 2.23% with PF 1.93 and a 6.3% drawdown. Pooled walk-forward OOS Sharpe was only 0.235 (t 0.48), and holdout Sharpe was 0.95 (t 1.28, below the 2.0 hurdle).

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