Options Research Dashboard

Top 5 — Real-Limit Total (16 weeks)

Best 5 setups by Real-Limit Total return. Click a row to jump to its detail view. Limit @ Mid is the most realistic execution model (sit @ midprice, get filled when spread allows).

#	Setup	Lim Total ⓘ	Lim CAGR ⓘ	Lim Sharpe ⓘ	Lim Max DD ⓘ	Lim Win ⓘ	Mod1 Total ⓘ	Mod2 Total ⓘ	Exec Total ⓘ
#1	BTC Overnight — Mon-Thu weekday — Naked (no hedge) — +1 strike (further OTM)	+6.62%	+5.30%	1.07	-2.32%	58%	+12.79%	—	-19.58%
#2	ETH 5/7 — Closer to Strike	+6.45%	+23.24%	10.69	-0.16%	94%	+24.72%	+10.05%	-29.65%
#3	ETH Overnight — Fri→Sun weekend — 5% wings hedge — +1 strike (further OTM)	+6.32%	+20.83%	2.22	-0.67%	53%	-1.05%	—	-37.09%
#4	BTC 3/5 — 3x Hedge	+5.98%	+20.97%	2.25	-1.92%	69%	+15.67%	-1.37%	-6.21%
#5	ETH Overnight — Mon-Thu weekday — Naked (no hedge) — ATM (1st strike above)	+5.93%	+4.77%	1.19	-2.78%	69%	+21.25%	—	-39.55%

Model 2 — Tardis-Calibrated Vol Surface

Model 2 is a Tardis-calibrated empirical implied-volatility surface that replaces the simple smile-bump formula used in Model 1. It enables a more realistic pricing baseline by reading IV directly from real Deribit observations.

How it works

1. Source data: 1.5M Tardis options observations on BTC + 1.1M on ETH (hourly snapshots over Jan 23 – May 23, 2026). Each row contains the actual mark IV at a given (timestamp, strike, expiry, side).
2. Binning: All observations are placed in a 3D grid of moneyness × time-to-expiry × DVOL. For each cell the median IV is taken (filters outliers; requires ≥ 5 obs per cell). Separate surfaces for BTC and ETH because IV regimes differ.
3. Lookup: At any (spot, K, T, DVOL, side) the surface returns the median IV from the matching cell; if the cell is empty it falls back to the nearest non-empty neighbor; final fallback is DVOL itself.
4. Pricing: The looked-up IV is plugged into Black-76 to get the model premium. Same machinery as Model 1, only the IV source changes.

Accuracy comparison vs Real-Mark

Below: total PnL over the 16 weekly cycles for each setup. Err = Model − Real-Mark. |Δ| = how much closer Model 2 is to Real-Mark than Model 1 (positive = improvement, negative = Model 2 over-corrects). Sorted by improvement descending.

Key findings

• BTC: Mean |error| from 7.5pp → 3.4pp (55% reduction). Largest improvements on far-OTM puts and ATM short-dated options where Model 1 had structural biases.
• ETH: same fix works after building separate ETH surface. Initial attempt failed because we queried BTC-trained surface for ETH options (different IV regime → 60-90pp errors). Now per-underlying caches.
• Signed bias: Mod 1 was +7.06pp average (systematically over-pricing premium → backtest looks too good). Mod 2 is +0.91pp (essentially unbiased relative to real Deribit marks).
• Caveats: Model 2 only valid within the Tardis training window (Jan-May 2026, DVOL 37-90%). Outside that range it falls back to DVOL + neighbor-bin extrapolation, which can over-correct.

Source code: vol_surface.py (builder + lookup), vol_surface_diagnose.py (validation plots in analysis/).

About this dashboard

This is a research dashboard for systematic BTC and ETH options strategies. It compares a portfolio of weekly-cycle option structures against 4 months of real Deribit option data (Tardis Jan 23 – May 23, 2026, 16 weekly cycles, ~2.6 million hourly option observations).

Strategy families tested

• Outer Ratio (S3+1) — sell weekly OTM put + call, buy daily long-only hedges. 15 setups across OTM-distance (4%–8%), hedge ratio (1× / 2× / 3×) and hedge frequency (daily / 3-day).
• Inner Ratio — 3-leg static put: SELL inner + BUY 2× midpoint + SELL outer. Magic cap below outer strike (intrinsic exposure constant when K_mid = (K_inner + K_outer) / 2). 6 setups across various distances.
• Hybrid — Inner Ratio put + bull call credit spread (5-leg). Adds upside premium with defined max loss.
• Overnight Straddle — sell ATM-ish straddle every 21:00 UTC, optional 5–7% wings. Tests weekend VRP harvest. (Backtested separately, 12 variants × 82 nightly cycles, see JSON ledgers.)

Pricing series shown on every card

• Model 1 — theoretical price from Black-76 with DVOL + smile bump (0.07 × |moneyness|^0.7 / √T).
• Model 2 — empirical vol surface calibrated directly to Tardis (1.5M observations binned by moneyness × T × DVOL). Mean |IV error| 3.5pp vs Model 1's 7pp.
• Real-Mark — Deribit posted mark price (theoretical mid, no spread cost).
• Real-Exec — worst-case real execution. Sells at bid, buys at ask.
• Real-Limit — limit order placed at midprice with realistic fill probability (90% if spread ≤ 5%, 60% if ≤ 15%, 30% if ≤ 40%, 0% above; unfilled legs contribute zero PnL).

Methodology notes

• All strategies use cash-settled European options on Deribit (option of choice for BTC).
• Empty orderbook fallback: min-tick assumption (0.0001 BTC ≈ $7.50 per contract).
• Nearest-strike fallback when target K is not listed (Tardis exposes the actual fill strike).
• Hedge cost vs payoff separated: every hedge leg is decomposed into cost-when-worthless (calm week burn) and payoff-when-winning (crash-week recovery). See Hedge Comparison page.
• Per-cycle trade ledger shows every leg with bid/ask/mark/fill prices, intrinsic at expiry, and per-series PnL. Available in each strategy's detail view.

Built with Python + matplotlib + Tardis Deribit historical data. Static HTML, no backend. Hover any column header marked with ⓘ for an explanation.