Everything we know about selling overnight options on Deribit (BTC/ETH), tested on REAL Tardis marks (Jan–May 2026, ~120 nights, 41,860 ledger trades) and cross-checked on a 3-year model history (2024–2026, 869 nights) to kill regime artifacts. Covers ALL overnight research to date (2024–2026 program + the 2026 deep-dives). Honest report — negatives included.
| Structure (ETH, $100 risk/night, limit fills) | win% | $/night | MaxDD$ | r/DD |
|---|---|---|---|---|
| put +1% naked | 81 | 1.29 | −56 | 2.8 |
| put +2% naked | 89 | 0.98 | −35 | 3.4 |
| put +3% naked ★ | 96 | 0.75 | −17 | 5.2 |
| call +3% naked ★ | 96 | 0.50 | −7 | 8.6 |
| put +3% spread (1-strike cover) | 92 | 1.21 | −99 | 1.5 |
| anything at 5% OTM | — | ~0 | — | — |
Findings: (a) further OTM = monotonically higher win rate AND better r/DD — 3% OTM naked is the sweet spot; (b) spreads ADD $/night via size but WORSEN r/DD (the defined max-loss hits full too often); (c) 5% OTM has no premium left. ⚠️ Execution is everything: at full market-order fills the 3% spread flips from +$69/night to −$220/night (win 21%) — far-OTM spreads only work with patient maker/limit fills at mid. Fill quality is the strategy.
| UTC hour (avg return, bps) | 2026 | 2025 | 2024 |
|---|---|---|---|
| 21:00 | +2.3 | +3.7 | +5.3 |
| 22:00 | +1.6 | +4.7 | +5.7 |
| 23:00 — negative in ALL 3 years | −6.8 | −3.3 | −2.4 |
21:00–22:00 drift up, 23:00 systematically weak (the most robust hour in the whole profile), median night low ~05:00, recovery into 08:00. Practical use: a 00:00 UTC entry sets strikes AFTER the weak 23:00 slot — real 00:00 entries ran 18.2 bps/night at win 81%. Both 21:00 and 00:00 entries work; midnight is slightly cleaner.
Tested closing at 06:00 instead of 08:00 settle, using an UPPER BOUND for the early close (buyback at pure intrinsic, time value for free): early close still loses — +5.54% vs +7.36% total, win 74 vs 77, same DD. Only 38% of nights would have been better; in 3 of the 5 worst nights the damage was already done by 06:00 and price partially RECOVERED into settle. The last hours are also NOT louder (avg |move| 06/07/08h = 39/30/37 bps vs 46 bps at 22:00). Contrast with weeklies: for 7-day sold PUTS, closing Wednesday is a big win (theta mostly earned, tail mostly ahead) — for 0DTE there is nothing left to harvest early.
| Candidate filter | 2026 screen (121 nights) | 3-year check (869 nights) |
|---|---|---|
| DVOL rising vs falling (5d) | 14.1 vs 0.6 bps — looked strong | 15.6 vs 14.0 — gone (mixed by year) |
| US session loud vs quiet | 9.1 vs 0.0 bps | 15.9 vs 13.6 — gone; loud bucket has 2× fatter worst nights |
| This week louder than last | 11.2 vs 0.7 bps | 14.6 vs 15.0 — gone |
| Direction gate from trend signals (EXPANSION/MA bias) | real 2026 0DTE ledger: UNFILTERED +201% (CVaR5 −107) · ALIGNED with signal +86% (CVaR −126, worse tail!) · ANTI +116% — direction carries no usable information, rejected | |
| Watch 21:00→00:00, filter at midnight | quiet 18.0 vs loud 15.2 bps — noise; even loud nights stay positive | |
| Weekday | Sun weak (2.3 bps), Thu/Fri strong — 17 nights/bucket, thin | partially confirmed on 869 model nights: Sun $28/trade + Mon $50 vs Sat $175 / Tue $141 / Wed $133 — Sun/Mon clearly weakest BUT still positive (win 88–90%) → reduce size Sun/Mon, don't skip |
(a) Bounce: after a ≥1% drop, price recovers from the night LOW into 08:00 in 60–71% of nights (all 3 years) — but this is hindsight (you only know the low afterwards). Buying at the first −1% threshold cross earns ~0 (2025/26) to +0.13%/trade (2024) with −19% MaxDD: you catch the knife half the time. (b) End-of-night counter-move after a strong night: tested last-2h/last-1h returns conditional on the 21:00→06:00 move, both assets, 3 years: signs flip across years, magnitudes single-digit bps — nothing tradeable after fees. The only consistent cell (moderate up-nights continuing in 2026) is thin (n=15) and absent in 24/25.
| Study | Result | Detail page |
|---|---|---|
| 0DTE selling matrix (sides × strikes × entry hours, real Tardis) | BTC puts dominated the real window; best variants ATM–3% OTM @20–21h and 00:00 (best win-rate hour) | 0dte_selling · 0dte_master |
| 0DTE on 2024+ model history (869 nights) | pattern stability check across regimes; basis for every long-history cross-check on this page | 0dte_m2_longhist |
| Weekend variant (Fri 21:00 → Sun/Mon, strangle + roll logic) | weekend VRP exists; strangle with tighter IC covers tested — works but weekend nights are already inside the nightly book (Fri–Sun entries: 5.3 bps vs weekday 6.7) | overnight report · weekend-atm |
| Overnight ATM straddle (incl. Fri→Sun expiry) | tested as vol-bet variant; OTM strangle dominates for the income profile | 0dte pages |
| Buying the night (cheap options on 5-min swing signals) | reversion longs lose everywhere; only ATM momentum-scalp WITH the move + TP was positive (thin sample) — selling remains the edge | cheap_options |
| Overnight sleeve in the portfolio | weekly correlation to the futures-stop weekly book: −0.34 (real) — the overnight book is the portfolio's anchor sleeve | portfolio |
| Method | Result | Implication |
|---|---|---|
| Variance Ratio (Lo-MacKinlay), 1h→2h/4h | VR ≈ 0.91–1.04 all years | the night is a random walk — no exploitable trend/reversion rhythm at hourly scale |
| Markov hour-to-hour persistence | P(same direction) 47.5–49.6% night, 46.8–48.3% day | tiny anti-persistence, <2pp — nothing after fees; explains why all timing ideas failed |
| Touch vs settle probabilities (first passage) | −1%: touch 24–31% / settle 17–21% · −2%: 10–11/5–9 · −3%: 3–4% touch / 1–2% settle — stable across 3 years | the key number: touch ≈ 1.5–2× settle → half of all strike-touches come back. This is WHY intra-night stops whipsaw and hold-to-settle wins; and why 3% OTM ≈ 96–99% settle safety |
| Tail concentration | worst 5% of nights = only 14–23% of total movement | night risk is spread out, not one-crash-dominated → supports trading every night, explains why filters add nothing |
| Timing of night low/high | U-shaped: lows 24% at 21–22h + 14% at 08h; highs 20% at 21h + 16% at 08h | extremes sit at the EDGES of the night; the middle (00–06) is calm — matches the 00:00-entry edge |
| Night-to-night vol clustering | autocorr(|move|) = +0.02 to +0.17 | last night barely predicts tonight — another reason conditioning fails |
run_sven_overnight_ideas.py · run_sven_overnight_filters.py ·
run_0dte_selling.py · run_0dte_m2_longhist.py · run_nik_overnight_gate.py ·
log: NIK_ALGOS_LOG.md. Ledger: 41,860 real trades, Tardis Deribit hourly.