Calendar Spreads on Futures Signals — the BTC 28/14 Timing Edge

Question (Sven): instead of taking a futures trade on every algo signal, enter a calendar spread (buy long-tenor / sell short-tenor, ATM). Tested all directions and combinations on Nik's algos (MA_CND_CROSS_H4, EXPANSION_H4, STOCH_H4, RISE_FALL_H1, MA_CND_CROSS_H1), BTC+ETH, tenors 14/7, 28/7, 28/14, M2 model 2021–2026 + real Deribit marks 2026. Six iterations, honest report — most combinations are dead; one specific configuration carries a real, statistically significant timing edge.

TL;DR

1. The road there: what is dead

2. The find: per-trade fairness on BTC 28/14, 2024–2026 only

Totals flatter the signals because they trade more often. Per-trade comparison, identical structure and evaluation (buy 28d / sell 14d ATM, hold 13.5d, M2 marks), entries 2024-01 onward:

By-year per trade (MA_CND_CROSS, signal-direction side): 2024 +0.63 · 2025 +0.42 · 2026 +0.37 — the edge does not live in one regime. First options structure in this whole program where signal timing measurably adds value (credit spreads, naked options, long options on signals all failed this test).

3. Channel: direction or timing? — pure timing

2×2 decomposition per signal direction × calendar side (MA_CND_CROSS, per trade, 24–26):

Reading: which side you pick is irrelevant (SAME == OPP), and signal entries beat blind on BOTH sides by ~2×. The information in the signal is when, not where (mechanism in section 5c). Consequence: trade it as a double calendar (both sides, direction-neutral) — that is also what survives execution best.

4. Real Deribit 2026 + execution costs (the decisive table)

Real Tardis quotes Jan 23 – May 22, 2026. Entry: buy ~28d / sell ~14d ATM at the nearest real expiries; exit 0.5d before short expiry, marked at real quotes. MID = everything at mark; EXEC = crossing the spread on all 4 legs (entry long@ask short@bid, exit long@bid short@ask):

Three results in one table: (a) the blind calendar is NOT tradeable — 4 legs of friction eat the entire carry (vs ~$10 for one future); (b) the MA_CND_CROSS-timed version keeps ~75% of its mid edge and stays clearly positive; (c) signal entries even pay tighter spreads — they happen intraday when books are tight, Friday 08:00 entries sit on Deribit settlement when books are wide. Real marks also confirm the model ranking.

5. Significance: month-cluster bootstrap

13.5d holds overlap (avg ~4 open positions) → trades are serially correlated and a naive t-test lies. Bootstrap resamples calendar MONTHS (signal and blind paired on the same months, 10,000 resamples), statistic = per-trade mean difference signal − blind, M2 2024–26:

MA_CND_CROSS survives even a Bonferroni correction over the full tenor×asset grid it was selected from (~18 cells: 0.0002 × 18 = 0.004).

5b. Robustness sweep: all 5 algos × both assets (the full answer)

Same per-trade test (28/14 double calendar vs blind Fridays, M2 2024–26, month-cluster bootstrap):

Reading: on BTC the edge is broad — 4 of 5 algos positive, the MA family + EXPANSION significant — so it is not a one-cell fluke. On ETH there is no timing edge at all (all p ≥ 0.25): the finding is BTC-specific, consistent with everything else in this program (ETH calendar carry itself was a 2021–23 artifact). RISE_FALL_H1 (the fastest algo) carries no calendar-relevant information on either asset.

5c. Tenor anatomy + mechanism: it's terminal pinning, not vega

Same per-trade test across tenors (BTC MA_CND_CROSS, hold to 0.5d before short expiry):

The edge lives entirely in the 14d short leg / 13.5d hold — 6.5d holds carry zero signal value. And the mechanism is NOT volatility: DVOL paths after signal entries are flat and identical to blind (Δ DVOL +1d/+3d/+7d/+13.5d all ≈ 0.00 vs blind, entry DVOL level identical). What differs is the terminal distribution: |spot move| over 13.5d is SMALLER after signals (5.90% vs 7.29% blind), while the 7d move is the same (5.1 vs 4.9). Two weeks after an MA_CND_CROSS entry, price tends to sit back near the entry level — and a calendar held to short expiry is exactly a bet on spot finishing near the strike. So the honest mechanism: 2-week mean reversion to signal-entry levels, harvested via theta differential, not vega. (42/14 tested even better, but that is one more cell picked after looking — treat as data, not as the trade.)

Cross-check that nails it: a naked SHORT 14d ATM straddle (same terminal bet, no 28d protection leg) shows the same edge even bigger — signal-timed +1.79%/trade vs blind +0.22% (diff +1.57, bootstrap p = 0.0002). But its risk profile shows why the calendar is the vehicle: straddle worst trade −27%, MaxDD −64% of notional, vs the calendar's defined risk (max loss = net debit, ~1–1.5%). The calendar keeps ~⅓ of the pinning harvest for ~2% of the tail risk. (Straddle numbers are mechanism confirmation, NOT a trade proposal.)

6. Honest caveats

7. If traded (sketch, not a go-live)

Sources: run_nik_calendar_signals.py, run_nik_cal_channel.py, run_nik_cal_real26.py, run_nik_cal_boot.py; full iteration log in NIK_ALGOS_LOG.md (iterations 1–6, June 2026 night program).