Research note 008: a full year of carry, bear grind to bull ramp

Date: 2026-10-14 · Status: twelve consecutive months, hourly bars, stated assumptions. The window-picking antidote to notes 004–007: all of 2023, January to December, unconditional hold plus the p90/p10 timed rule from note 007 — same legs, fees, and fills throughout, so the year decides between them instead of any quarter.

Question

Notes 004–007 chose their windows (bull quarter, bear quarter, both halves). Does the carry thesis survive a whole year it did not get to choose — and does timing earn its turnover over twelve months?

Dataset and method

• Binance public archive, 36 slices pinned by SHA-256 in research/scripts/carry_008.py: 1,095 funding payments plus 8,759 paired hourly bars. Pairing is by timestamp, not position: March carries one perp-only bar (the 2023-03-24 spot halt), dropped and disclosed — the script asserts at most a handful of unpairable bars and reports the count (1).
• Two provenance upgrades over notes 004–007, both forced by findings: pins cover the extracted CSV bytes, not the zip container (the CDN re-packages identical CSVs into byte-different zips — observed on one file across three downloads with identical content), and every file asserts its calendar row count beside its hash (spot files carry no header while perp/funding do — found the hard way, documented in code).
• Unconditional hold over the full year ($20 in fees). Timed p90/p10 over the same bars with January doubling as the 720h warmup. Same $10k legs, 5 bps taker, fills at closes, bronze labeling throughout.

Results

Funding by month ($): Jan 80.00 · Feb 69.91 · Mar 52.44 · Apr 45.13 · May 51.29 · Jun 60.29 · Jul 56.28 · Aug 53.89 · Sep 9.74 · Oct 37.03 · Nov 101.84 · Dec 168.72. September — 33 of 90 payments negative — is the year's dead month; November–December, zero negatives, is where the year was made. Spot ran $16,529.67 → $42,283.58 across the window.

Reading

Three facts that end the window debate. First, the unconditional hold compounds across regimes: eleven of twelve months paid, September cost little, and the year cleared 3.7% after all stated costs — the carry thesis survives contact with a year nobody chose. Second, timing helped drift again (+$158 vs −$20) and lost again (−$505 of gap): twelve careful turns still cost more than the selectivity they bought, at $20 a turn. Third, the fee mechanism is now visible across four studies: turnover is the only variable that ever decides the sign. The trilogy's moral holds at twelve months: timing is real edge on drift and negative edge after turnover, and no band has yet been both selective and cheap.

What would strengthen or overturn this: maker fills where the book supports them (the fee floor remains the entire game); sizing by conviction; borrowing-inclusive inverse positions for negative-funding regimes; and the live-captured version, where fills stop being assumed.

Reproduce

python3 research/scripts/carry_008.py --data-dir /tmp/carry008

Thirty-six CSV content hashes plus calendar row counts pinned in the script, refused on any mismatch — verified by a from-scratch rerun with identical figures.