EthBybitMomentumBurstSecondWaveLong

Momentum-Burst Second-Wave Continuation on ETHUSD.BYBIT (4H): after a >=1.8-sigma 80-hour price burst, buy the next open and exit on a FIXED 36-hour clock - an event-triggered, flat-92%-of-the-time long that harvests squeezed shorts and lagged copy/trend flow, not an always-in trend filter

A SINGLE-INSTRUMENT, SINGLE-LEG, SINGLE-VENUE, EVENT-TRIGGERED long on the Bybit linear ETH perpetual (catalog id ETHUSD.BYBIT, 4-HOUR bars, verified span 2019-01-25 -> 2026-09-19, 16,767 bars). Flat ~92% of the time; holds only in the 36 hours after a rare standardised momentum burst. No volume, funding, OI, liquidation, sentiment or cross-venue input. Mechanism (stripped): 'When a perp's trailing multi-day return reaches an extreme relative to its OWN recent distribution of trailing returns, the following ~1.5 days carry a large positive drift, so buy the burst and exit on a clock.' Payers: shorts forced to cover, lagged copy/bot/CTA flow, and the original parent order still completing. Measured: 15 Hyperliquid perps 8H 2022-2026, z>2 -> +1.78% mean / +0.55% median 2-day forward vs +0.20% unconditional, t=10.6, monotone in z; non-overlapping delayed-entry sim +1.54%/trade, 14/15 positive. Chosen cell (ETHUSD.BYBIT 4H, z>=1.8, 9-bar hold, 7.6y): n=149, mean +1.14%, median +0.60%, win 54%, PF 2.02, t=3.12, positive in all 8 calendar years. The fade and the short leg were measured and rejected.

Implements the hypothesis literally and minimally. (1) '>=1.8-sigma 80-hour price burst' → burst_bars=20 four-hour bars for the trailing return, standardised against the instrument's OWN recent distribution of the same trailing return (z_lookback prior observations, running sum/sum-of-squares over a bounded deque so per-bar cost is O(1) and independent of backtest span). calculate_signal returns that raw z every bar (continuous, varies bar to bar, verified 240/240 unique values on all six synthetic scenarios); the 1.8 threshold lives in should_enter, in matching z units. (2) 'buy the next open' → a market entry at the trigger bar's close, which is that bar's next open. (3) 'exit on a FIXED 36-hour clock' → should_exit compares the current bar's timestamp against the position's ts_opened (mirrored into the scalar self._entry_ts_ns so a paper restart cannot re-anchor it) with a half-bar tolerance, giving exactly 9 bars. Nothing else can exit — no stop or target was added because adding one would test a different mechanism. (4) 'the fade and the short leg were measured and rejected' → long-only, so short entries are never generated. Venue stays BYBIT as declared (Layer-1.5 consistency and the hypothesis's own measurement cell); leverage is left at 1.0 and sizing therefore does not reference it beyond the explicit config.leverage multiplier, so nothing is inert. One honest divergence to flag for the Analyst: replaying this exact construction over the full 16,767-bar ETHUSD.BYBIT 4H history gives n=158, mean +0.82%/trade net of 0.11% round-trip taker fees, median +0.30%, win 53.8%, PF 1.58, t=2.03 — clearly above the 0.15% futures per-trade floor and positive across the 120–360 z_lookback plateau and the 10–40 burst_bars range, but weaker than the pre-registered +1.14%/PF 2.02/t=3.12, and the pre-registered 'positive in all 8 calendar years' does not reproduce (2019–2024 all positive, 2025 −0.52%/trade and 2026 −0.46%/trade). I implemented what was asked rather than tuning z_lookback until the claim reproduced; the recent two-year sag is real and should be scored against the pre-registered prediction rather than optimised away.

suspected backtest artifact: the daily MTM series repeatedly prints identical large loss/gain pairs despite a single-instrument, low-exposure book, so the Sharpe and risk metrics cannot safely be optimized.

Event-triggered, long-only second-wave continuation on the Bybit ETH perpetual (ETHUSD.BYBIT, 4H bars). Every bar it computes the trailing 20-bar (80-hour) simple return and standardises it against the prior 240 observations of that same trailing return (current observation excluded) — that z-score is the continuous signal. When the burst reaches +1.8 sigma the strategy buys at the trigger bar's close (== the next bar's open) and exits on a FIXED 36-hour calendar clock (9 four-hour bars). No stop, no take-profit, no trailing exit, no signal exit — the hypothesis prices a fixed-horizon forward drift so the clock is the only exit. Position sizing is 50% of account equity in notional at 1x leverage, so gross exposure never exceeds equity. It holds roughly 9 bars per event and is flat ~92% of the time; it is not an always-in trend filter. Price-only: no volume, funding, OI, liquidation, sentiment or cross-venue input, and no short leg.

Leave the clock exit as-is for this test. If the optimizer later raises position_fraction toward its 1.0 bound, the Risk Officer should treat the per-event tail as the binding constraint, since there is no stop to cap it.

Fixed-clock exit with no stop-loss and 50% of equity in notional. This is a faithful implementation of the hypothesis ('exit on a FIXED 36-hour clock ... no stop, no target'), so it is NOT a defect and adding a stop would test a different mechanism -- but it means a single 36h window is uncapped: the sandbox's largest_loss (-$8,438 against ~$37.5k average notional) is a ~22% single-trade adverse excursion. Acceptable only because the book is flat 91% of the time (exposure_pct 9.0) and leverage is 1.0.

No code change needed for the backtest. Before any paper/live promotion, confirm which contract the ETHUSD.BYBIT bars were actually collected from and either route to ETHUSDT or convert sizing to USD contracts -- equal-ETH sizing on an inverse contract would misstate notional and PnL convexity.

Live/paper parity note on the instrument, not a backtest defect. The catalog models ETHUSD.BYBIT as a CurrencyPair (base ETH / quote USD, size_precision 8, size in ETH), so the engine prices and sizes it LINEARLY and position_size()'s qty = notional/price is correct for the backtest. On the real venue, however, Bybit's 'ETHUSD' ticker is the INVERSE (coin-margined) perpetual whose quantity is denominated in USD contracts; the linear USDT perp is ETHUSDT. The hypothesis calls ETHUSD.BYBIT 'the Bybit linear ETH perpetual'.

No change required; flag only so the analyst reads clustered events correctly in the trade ledger.

should_exit() and should_enter() are mutually exclusive within a bar in the base template (position present -> exit path only), so the exit bar cannot also re-enter; the earliest re-entry is one bar (4h) after the clock exit. The hypothesis does not specify re-entry behaviour, so this is only noted for interpreting overlapping-burst events -- clustered z>=1.8 bars produce a 9-bar-on / 1-bar-off cadence rather than a continuous hold.

Three things the analyst should weigh at the real backtest, none of which block here. (1) The developer's own full-history replay (n=158) gives +0.82%/trade, PF 1.58, t=2.03 vs the pre-registered +1.14%, PF 2.02, t=3.12, and the pre-registered 'positive in all 8 calendar years' does NOT reproduce (2025 -0.52%/trade, 2026 -0.46%/trade). Score that against success_criteria rather than letting the optimizer tune z_lookback until the claim reproduces. (2) The event sample is far less independent than n=149/158 suggests: the signal is an OVERLAPPING 20-bar trailing return standardised against 240 overlapping observations of itself, so one macro burst can emit several consecutive z>=1.8 bars and the effective number of independent events is materially below the trade count -- the t-stats in the hypothesis are optimistic. (3) The clock-only exit with 50% of equity in notional and no stop produced a largest_loss of -$8,438 on ~$37.5k average notional (~-22% adverse excursion inside one 36h window) in the sandbox; drawdown is benign at 8.9% only because exposure is 9%. Sandbox ledger is internally coherent (0.926%/trade x 140 x ~0.375 deployed-equity ratio ~= the +55.6% total return; PF 1.63 > 1 with positive return), so the per-trade economics are usable.

152 trades over 7.2 years with positive total return and 0.859% average trade return.

Trades match the intended long-only burst-continuation mechanism.

The daily return series contains repeated identical large negative/positive pairs inconsistent with the reported single-leg exposure, indicating a suspected MTM/accounting artifact.

Sharpe is only 0.297, its CI spans zero, and 2025–2026 are negative; optimization would fit contaminated metrics.

EthBybitMomentumBurstSecondWaveLong-7f8827d4f6

The strategy tested a long-only, event-triggered continuation trade on ETHUSD.BYBIT, entering after a high z-score 80-hour momentum burst and exiting after 36 hours. It produced positive total return and trade-level profitability, but the daily MTM series contained repeated identical large loss/gain pairs inconsistent with the intended exposure. It was abandoned at backtest review because the suspected artifact made the risk metrics unsafe to optimize.

Validate backtest accounting and daily mark-to-market data before relying on trade-level results or proceeding to optimization.

The backtest-review verdict was abandon because the daily MTM series repeatedly showed identical large loss/gain pairs inconsistent with the reported single-leg exposure, suggesting an accounting artifact. Later stages were not reached.

It tried to buy ETHUSD.BYBIT after an extreme standardized 80-hour upward price burst and exit after a fixed 36-hour holding period.

The backtest produced a 47.07% total return over 152 long trades, with a 0.297 Sharpe ratio, 1.420 profit factor, 51.97% win rate, and 0.859% average trade return. Maximum drawdown was 12.55%.