Strategies

ETH Deribit Long-Convexity Directional Trend Capture, Long-Short via Cheap Options (DERIBIT ~30-delta single-leg call/put, ~30 DTE, entered on a SLOW daily-trend sign ONLY when option IV is NOT rich, biweekly roll, defined-risk premium spend)

Outcome: Abandoned

EthDeribitTrendLongConvexityCheapIvRoll

Outcome Summary

EthDeribitTrendLongConvexityCheapIvRoll correctly implemented its thesis — buying cheap ~30-delta ETH options in the direction of a slow trend only when IV was not rich, giving defined-risk convexity that is the opposite sign of the factory's short-vol graveyard. The realized numbers were even mildly favorable: +$450 average trade, 1.35 profit factor, +5.4% total return, with Deribit's ~0.06% round-trip fee trivial next to premium moves. But over the entire available option window it fired only 12 times, producing a Sharpe of 0.389 whose confidence interval straddled zero, and the biweekly-plus-double-vol-gate cadence would top out near ~40-45 trades across all history. On its first iteration the analyst abandoned it at backtest review as unmeasurable — a well-built mechanism that simply cannot trade often enough to prove it has an edge.

Strategy report

Hypothesis

A LONG-only-premium (never short vol), single-leg, one-position-at-a-time DIRECTIONAL options strategy on DERIBIT ETH options... [fills options 0-6% bucket, long_short, and cross-venue; avoids short-vol / skew-fade / perp-momentum-gate graveyards]

Implements the hypothesis literally: LONG-only premium (leg_sides BUY for both kinds, never a short-vol leg), single-leg, one-position-at-a-time, ~30-delta (offset +-0.07) ~28-DTE Deribit ETH contracts from the framework's rolling weekly chain, entered on the SIGN of a slow daily trend and gated by an 'IV not rich' filter (DVOL iv_rank <= 0.65 and IV/RV <= 1.45), with a biweekly cadence enforced calendar-wise (min_days_between_entries measured from the bar timestamps of the previous entry, plus max_hold_days = 14). Structure follows RollingOptionsStrategy exactly (chain + schedule injected by the framework); there is no underlying smoke path and no price-only fallback: if the chain is absent or DVOL is missing/stale the strategy trades nothing. The four required hooks are all live — calculate_signal returns the continuous trend gap every bar, should_enter thresholds it (side is always BUY because direction is expressed by the contract KIND), should_exit is the trend-flip exit, and position_size converts a premium budget into contracts. Venue is DERIBIT (options, leverage 1.0, 0.03% per side) as the hypothesis specifies; per-bar work is O(1) (memoized metrics, a schedule cursor, bounded buffers) and lookback_days=730 keeps the chain at ~208 legs, a size that has completed the Layer-3 options sandbox before. One reservation, stated once: buying premium is structurally short the (positive) crypto VRP, so the whole edge rests on the IV-not-rich gate plus trend persistence — I implemented it as asked.

data_wall: unmeasurable / insufficient statistical power: the strategy produced only 12 option trades over the full available ETH-Deribit option history (legs span ~2024-08 to 2026-09, ~2 years), with Sharpe 0.389 on a CI [-1.46, 1.46] that straddles zero and PSR 0.53 — non-diagnostic. This is NOT a fee/expectancy failure (Deribit RT ~0.06% is trivial vs premium moves, realized expectancy +$450/trade, PF 1.35); the problem is that the mechanism is structurally too slow to validate. Biweekly cadence gated further by a double vol filter extrapolates to only ~40-45 trades over ALL of history, so 2 hours of 3-phase optimization would only curve-fit ~12 trades of noise (walk-forward windows holding ~4 trades each, holdout a handful). The trades correctly implement the hypothesis (7 calls/5 puts by trend sign, ~0.69 IV, ~28 DTE, positive delta/vega, negative theta) so this is not a broken mechanism, but the sample cannot statistically support a promote decision and cannot be enlarged by tuning. QA's structural concern (long premium = short the persistently positive crypto VRP, edge hinges entirely on the IV-not-rich gate + trend extending past strike) is unresolvable on 12 observations, and the book already went negative in 2026. abandon_class = unmeasurable.

Implementation

Directional LONG-PREMIUM option roll on Deribit ETH weeklies. Every ~2 weeks, when a slow daily trend on ETH-PERPETUAL (20d vs 80d SMA gap) is decisively signed AND implied vol is NOT rich (Deribit DVOL at/below the 65th percentile of its trailing year AND <= 1.45x 20-day realized vol), the strategy BUYS a single ~30-delta (+-7% moneyness) ~28-DTE option: a call in an uptrend, a put in a downtrend. It never sells an option, so maximum loss per cycle is the premium paid (sized to 3% of equity) while the payoff is convex if the trend extends. One position at a time; exits on a 14-day max hold (the biweekly roll), the 7-day pre-expiry roll buffer, a trend flip against the leg, a -55% premium stop or a +120% premium take-profit. The underlying perp is never traded — it only supplies the trend clock and the realized-vol reference.

Verification Results

Verify the base RollingOptionsStrategy/FactoryStrategy warmup path still functions with the override, or rename to an internal helper (e.g. _min_bars) if the shadow is unintended.

min_bars_required(self) overrides a framework method (flagged by static analysis as base_shadow). The base template may rely on its own warmup gate; the override returns trend_slow_days+5 which is a sane, larger warmup so behavior is not broken, but confirm the base does not depend on the original.

Consider deriving the leg's quote currency from the instrument definition rather than a 1.0 magnitude threshold, so sizing is robust if a leg ever quotes in USD near/above 1.0.

position_size infers option premium denomination by magnitude (premium_px < 1.0 -> treat as coin premium and multiply by spot; else treat as USD). This heuristic is reasonable for ~30-delta ETH options (coin premium typically 0.02-0.08 ETH) but is a soft assumption; a mis-classified quote would mis-size the debit.

Long-premium single-leg buying is structurally SHORT the (persistently positive) crypto VRP: the whole edge rests on the 'IV-not-rich' gate (DVOL iv_rank<=0.65 AND IV<=1.45x RV) plus the trend actually extending past the strike before theta bleeds the debit. The developer acknowledges this. Sandbox is mildly favorable (total_return +5.4%, PF 1.35) but on only 12 trades with a Sharpe CI of [-1.46, 1.46] straddling zero, so it is non-diagnostic. Also note trade cadence: ~12 trades over 640 sandbox days extrapolates to only ~40-45 over full ETH-Deribit/DVOL history (biweekly cadence further cut by the double vol gate), so the full backtest may be near the low end of statistical power. Advisory only — carry to the analyst for the real full-history read; do not block.

Backtest Review

Sharpe: 0.39
Total return: 5.40%
Max drawdown: 6.86%
Trades: 12
Win rate: 33.3%
Profit factor: 1.35

Long-premium single-leg convexity — the correct sign vs the short-vol graveyard (max loss = defined debit), and the trades match the hypothesis: 7 calls in uptrends / 5 puts in downtrends, ~0.69 avg IV at entry, ~28 DTE, ~0.99 moneyness, net_entry_delta and vega positive, theta negative (paying for convexity as intended).

Deribit fee (~0.06% RT) is trivial vs premium moves, so this is NOT a fee-floor casualty; realized expectancy is positive (+$450/trade, PF 1.35, +5.4% total).

Only 12 option trades produced over the entire available ETH-Deribit option window (legs exist ~2024-08 -> 2026-09, ~2 years). Sharpe 0.389 with CI [-1.46, 1.46] straddling zero and PSR 0.53 — statistically non-diagnostic.

The cadence is structurally rare: biweekly roll further cut by the double vol gate (DVOL rank<=0.65 AND IV<=1.45x RV). Over full history this extrapolates to only ~40-45 trades at best, so the mechanism cannot generate a validatable sample even after optimization.

Optimizing 15+ parameters against 12 trades would curve-fit noise, not find an edge. Walk-forward windows would hold ~4 trades each and the holdout a handful.

Directionally the book already turned negative in 2026 (-4.9%), and QA's structural concern (long premium is short the persistently positive crypto VRP, edge rests entirely on the IV-not-rich gate + trend extending past strike) cannot be resolved on 12 observations.

Outcome Summary

A mechanism whose entry cadence is inherently rare cannot generate enough trades to validate an edge no matter how favorable the per-trade economics, so trade frequency must be checked against statistical power before committing to a slow-signal options structure.

The analyst abandoned it at the pre-optimization backtest-review gate: not a fee or expectancy failure, but a data wall — the biweekly cadence plus double vol gate is structurally too slow (extrapolating to only ~40-45 trades over all history), so 12 trades cannot statistically support a promote decision and optimizing 15+ parameters against them would only curve-fit noise.

A long-premium, single-leg ~30-delta ETH Deribit options roll that bought a call in a slow daily uptrend or a put in a downtrend, but only when implied vol was not rich (DVOL rank <=0.65 and IV <=1.45x realized), to capture directional convexity with defined risk (max loss = premium paid).

Over ~2 years of available ETH-Deribit option history it produced just 12 option trades (7 calls / 5 puts) with a positive expectancy of +$450/trade, profit factor 1.35, and +5.4% total return, but a Sharpe of only 0.389 on a CI of [-1.46, 1.46] and PSR 0.53. Win rate was 33% and max drawdown 6.9%.

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