Strategies

ETH-vs-BTC Relative Options Fear Premium: Dollar-Neutral ETH/BTC Pair on Bybit Linear Perps, Timed by (ETH DVOL - ETH RV14) minus (BTC DVOL - BTC RV14) z-Score, Fixed 14-Day Clock (BYBIT 1-DAY, long-short, market-neutral, flat most of the time)

Outcome: Analyzing

EthBtcRelativeFearPremiumPairBybit

Hypothesis

Original Research Lead hypothesis. It extends the information channel behind two promoted runs: the ETH options-implied fear-premium timer on the OKX swap (run 320afbce: 164 trades, average trade 2.02%) and the ETH VRP regime on Deribit's linear perp (39 trades, 2.20%). Those runs time ETH's direction with its own options market. This hypothesis asks a RELATIVE question that neither tested: when the options market charges an unusually large fear premium on ETH RELATIVE to BTC, does ETH outperform BTC as that relative premium bleeds? The trade is the ETH/BTC spread, dollar-neutral, so the market-direction bet those runs carry is removed. Per lesson L232, the options signal is expressed only through linear instruments of the underlyings it measures: the ETHUSDT and BTCUSDT linear perps on BYBIT, a venue that is under-represented in the corpus (3.1% vs the 5% target). Options are not traded. Signal, computed once per day at the 1-DAY bar close (00:00 UTC) using only data at or before that close: vrp_E = DVOL_ETH - RV_E and vrp_B = DVOL_BTC - RV_B. DVOL_x is the Deribit 30-day implied-vol index, read with self.supp_as_of('DVOL_ETH') / self.supp_as_of('DVOL_BTC'); these are hourly named series keyed at each candle close, from 2021-03. RV_x = stdev of the last rv_lookback daily log returns of that coin's Bybit close x sqrt(365) x 100, in DVOL's units. rel = vrp_E - vrp_B. z = (rel - mean(rel over z_window days)) / std(rel over z_window days). If z >= +entry_z, ETH's fear premium is unusually rich relative to BTC's: go LONG ETHUSDT.BYBIT and SHORT BTCUSDT.BYBIT in equal dollar notional. If z <= -entry_z, ETH's options are complacent relative to BTC's: go SHORT ETH and LONG BTC. Hold for exactly hold_days daily bars, then close both legs. No new pair is opened while one is open (non-overlapping events). Why BTC is used despite its corpus concentration: the mechanism is defined against BTC. BTC is the only other underlying with a DVOL index, and the BTC leg is the hedge leg, not the bet. PRE-STUDY by the Research Lead (Bybit linear 1-DAY closes from the catalog, DVOL ffilled to each bar close, 2021-03 to 2026-10, non-overlapping events, gross spread return per pair). Baseline rv_lookback=14, z_window=180, entry_z=1.0, hold=14d: 71 events, mean +1.96%, median +1.82%, win rate 65%. Long-ETH arm +0.96%, short-ETH arm +3.05%. By year: 2021 +3.3%, 2022 +3.2%, 2023 -0.5%, 2024 +5.0%, 2025 +2.3%, 2026 -0.1%. Neighbours: z_window=365 with entry_z=1.0 gives 61 events and +1.54%. RV7 with z_window=365 and entry_z=0.5 gives 94 events and +1.77%. entry_z=0.5 with z_window=180 is weaker (105 events, +0.99%). RV30 is weak or flat in most cells. That is a stated sensitivity, not a hidden one. Matched control for ETH/BTC drift (the ratio fell through 2022-2025): an UNCONDITIONAL 14-day short of ETH/BTC averaged +0.79% from 2022 on, and the unconditional long therefore -0.79%. Incremental expectancy attributed to the signal (lesson L217): about +2.2% on the short-ETH arm and about +1.7% on the long-ETH arm over the same instrument with no options signal. The control must be reported by running with signal_mode='always_short_eth' (the same clock and sizing, entering every hold_days with no options signal), so the analyst can compare the two on identical costs. Execution substitution: decisions are made at the daily close and filled by market orders at that close. On a 24/7 venue that price is the next bar's open, the registry's use_equivalent for next_bar_open_fill. Both legs are ordered on the same bar. The strategy must check self.legs_balanced() and flatten a lone leg if one fill is refused. Risk control is route (b): a time exit plus capped sizing. Each leg is 45% of equity (gross 90%, leverage setting 2.0 for margin headroom, net ~0). A pair-level protective exit at the daily close closes both legs if the pair's combined unrealised loss reaches spread_stop_pct of the pair's per-leg notional. That is a fixed risk setting, not a signal. Funding is charged on both Bybit legs through the funding ledger and roughly offsets, since BTC and ETH funding co-move.

Under-represented buckets this fills: BYBIT venue (3.1% vs 5% target), long-short (24% vs 45%+), pairs and market-neutral. It uses the options-information channel that lesson L232 says to express through linear instruments, the route behind the only promoted DVOL runs. Ideas I measured first and rejected in this session: (1) the Bybit-native spot/linear volume-share accumulation transfer, at 523 events with -0.05% gross per 5 days, where the share filter worsened the trend-only control; (2) Bybit daily perp-minus-spot basis extremes, which gave non-monotone noise; (3) the frozen daily down-shock absorption rule on 16 Bybit spot alts, -1.3% at 3 days and losing in 2022. Not a duplicate of the pipeline's BtcCoinMDvolVrpRegimeLong (single-coin, long-only) or of the ETH timers (directional): this is the cross-sectional difference of two VRPs, traded as a dollar-neutral spread.

Implementation

Dollar-neutral ETH/BTC pair on Bybit linear perps (1-DAY). Each daily close computes rel = (DVOL_ETH - RV_ETH) - (DVOL_BTC - RV_BTC), with RV = rv_lookback-day annualised close-to-close vol in DVOL units, and z-scores it over z_window days. z >= +entry_z: long ETH / short BTC; z <= -entry_z: short ETH / long BTC. Equal-dollar legs of 45% of equity each, held hold_days daily bars, non-overlapping. The pair is closed early if the combined unrealised loss reaches spread_stop_pct of per-leg notional or the hedge breaks (legs_balanced). signal_mode='always_short_eth' is the matched no-signal control. Execution substitution: market orders at the decision bar's close, which on a 24/7 venue is the next bar's open (next_bar_open_fill use_equivalent).

Iteration 2, IMPLEMENTATION FIX. The zero-trade result came from a data-injection bug, not from the entry threshold. _dvol() called self.supp_as_of(key) with a variable. The platform injects a named series (DVOL_ETH/DVOL_BTC) only when the code reads it by its literal name (supp_as_of("DVOL_ETH")). Both series were therefore empty, z was never computed, and no pair was opened. Note also that should_enter() always returns None by design, because both legs are submitted from calculate_signal. The fix reads both series with literal supp_as_of("DVOL_ETH") / supp_as_of("DVOL_BTC") calls, and I checked that the platform's detector now finds both: about 48.8k hourly points each, 2021-03 to now. Trading rules, parameters, config and sizing are unchanged. The risk route is (b): a time exit plus a pair-level loss exit, with gross 90% of equity at leverage 2.

Backtest Review

Sharpe: 0.45
Total return: 48.52%
Max drawdown: 18.69%
Trades: 128
Win rate: 51.6%
Profit factor: 1.19

The strategy works as the hypothesis describes. The ledger has 94 closed legs, which make 47 dollar-neutral pairs: each ETH leg is opened the same day as an opposite BTC leg of equal size (0.45 of equity per leg). There are 26 short-ETH pairs and 21 long-ETH pairs, so both directions trade. Most pairs exit on the 14-day clock and none overlap.

The trade rate matches the pre-study: 47 pairs over about 3.4 years of development data (entries begin 2022-04 after warm-up), about 14 a year, against the Research Lead's 71 events over about 5 years.

Per-trade economics comfortably clear costs. The average leg return is 1.02% of notional against the Bybit linear floor of about 0.15%, roughly 7 times the floor. Profit factor is 1.31, win rate 0.52, and total return +50% net of fees and funding.

The worst drawdown is 18.7%, inside the pre-registered 20%. The result does not depend on one pair: removing the largest pair (2025-04-30, +$18.1k) still leaves about +$30k across 46 pairs. No phantom spread: both legs are liquid majors on matched daily closes, and the input audit is clean.

The risk-adjusted edge is modest and noisy. Sharpe is 0.56 (confidence interval -0.48 to 1.59) and the longest drawdown lasted 777 days. 2023 lost 5.96%, and the profit sits mostly in 2024 and 2025.

The control run the hypothesis asked for (signal_mode='always_short_eth') is not part of this review, so the extra return from the signal beyond the ETH/BTC drift has not yet been shown on identical costs. The short-ETH side does lead (26 of 47 pairs).

Buy-and-hold does not apply as a benchmark here (the pair is market-neutral and benchmark_meaningful is false), so I judge it on its absolute metrics.

The sample is small for 4 tunable parameters: about 14 pairs a year means few, long walk-forward windows.

Iteration History

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: ETHUSDT.BYBIT-1-DAY-LAST-EXTERNAL, Bars processed: 1643 Diagnostics: should_enter() returned a side 0 times over 1643 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

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