Research note 009: the tape keeps venue hours, and batches on both venues
Date: 2026-10-14 · Status: one full week, two tapes, stated sessions. Answers the two follow-ups notes 001–003 deferred that archive data can answer: burst statistics split by session, and same-millisecond batching by venue. The third (simultaneity in the fill model) stays open — it needs the engine, not a week of bars.
Question
When does the market actually trade — and does the millisecond-batching note 003 found on spot also happen on perp? One week of Binance BTCUSDT aggTrades (2024-01-15..21, Monday to Sunday), spot and USDT-perp side by side: ~15.1M prints, streamed and never held. Sessions are UTC thirds (Asia 00–08, EU 08–16, US 16–24), stated so the buckets can be argued with.
Dataset and method
- Spot: 7,553,902 prints, $10.06B notional. Perp: 6,524,257 prints,
$66.57B notional. Fourteen daily slices pinned by extracted-CSV SHA-256
in
research/scripts/sessions_009.py(stdlib only, one command). - Timestamps asserted non-decreasing per file; pairing is per-print event
time, never wall clock. Spot files carry no header, perp files do;
spot booleans read
True, perptrue— the script normalizes both, and a 100%-taker-buy reading during writing caught the case difference before it reached this note.
Results
| Measure | Spot week | Perp week |
|---|---|---|
| Prints | 7,553,902 | 6,524,257 |
| Taker-buy share | 40.3% | 50.1% |
| Same-millisecond prints | 45.5% (889,556 runs) | 49.6% (614,429 runs) |
| Prints/sec, median / max | 4 / 3,091 | 5 / 2,256 |
| Weekend share of prints | 15.7% (of 28.6% of time) | 11.3% |
Prints and volume by session (share of venue week):
| Session (UTC) | Spot prints | Spot volume | Spot taker-buy | Perp prints | Perp volume | Perp taker-buy |
|---|---|---|---|---|---|---|
| Asia 00–08 | 21.8% | 21.1% | 43.9% | 20.3% | 18.9% | 50.2% |
| EU 08–16 | 36.3% | 35.2% | 41.3% | 35.6% | 35.9% | 50.2% |
| US 16–24 | 41.9% | 43.7% | 37.6% | 44.2% | 45.2% | 49.9% |
Reading
Three facts, each with a consequence. First, the week has a timezone: the US third carries 42–44% of prints on both tapes while Asia carries ~20%. Any feature averaged over "the day" mixes two different markets; session tags belong on every print-level feature the way venue tags already do. Second, spot shows session directionality and perp does not: spot taker-buy share slides 43.9% → 37.6% across the day while perp sits at 50% in all three sessions. Direction lives in spot flow; perp flow is balanced size passing through — consistent with hedging and basis traffic rather than directional betting, and with note 002's dust (equal-weighted signals overweight the noisiest, least directional flow). Third, batching is a venue property, not a spot property: ~half of all prints share their millisecond on both tapes, so replay order inside the millisecond is arrival order on perp exactly as on spot. The simultaneity question for the fill model just got twice as urgent — and the remaining venue in it is Bybit, which publishes no archive to check against.
What would strengthen or overturn this: the same week on a volatility breakout (does the US share stretch further?); Bybit batch rates from a live capture, since no archive exists; taker-buy persistence (does an imbalanced hour predict the next one, per session?); and the fill model with explicit same-millisecond simultaneity, backtested against the print-by-print default.
Reproduce
python3 research/scripts/sessions_009.py --data-dir /tmp/sess009
Fourteen CSV content hashes pinned in the script, refused on any mismatch — verified by a from-scratch rerun with identical figures.