Trading calendars and sessions
Use an explicit date window and a versioned calendar. Keep session labels separate from observation timestamps. An overnight session's date need not be its opening civil date.
Reproducibility trap
✅ Source-verified 2026-09-14: installed exchange_calendars 4.13.2,
exchange_calendars/exchange_calendar.py, lines 59 and 62, derives default bounds from
pd.Timestamp.now() at import, minus 20 years and plus 1 year. Calendar-specific bounds
can narrow this range. Supply start= and end= and record the package version.
✅ Measured 2026-09-14 by calendars.py, with the library installed:
| Simulated import date | Default XNYS first session | Last session | Sessions |
|---|---|---|---|
| 2026-06-15 | 2006-06-15 | 2027-06-15 | 5282 |
| 2026-06-16 | 2006-06-16 | 2027-06-16 | 5282 |
The second index loses 2006-06-15 and gains 2027-06-16: symmetric difference 2. The test runs imports in child processes so it does not modify a live application's pandas clock. Without the library, the demo labels its weekday-only fallback; it is not an exchange calendar.
import exchange_calendars as xc
cal = xc.get_calendar("XNYS", start="2000-01-01", end="2026-12-31")
Aggregate observations using the supplied schedule
✅ Executed reference implementation, 2026-09-14: schedule_bars(px, schedule, bar_minutes)
accepts a price Series and explicit open, close, optional break_start, break_end
columns. The schedule index is the session label. It uses left-closed/right-open intervals,
starts a new bucket after lunch, retains empty buckets as missing values with count zero,
and rejects observations outside the schedule. Missing observations do not shift boundaries.
Use timezone-aware instants for DST and overnight sessions. Half-days use their actual close.
The minutes output counts nonmissing price observations; it equals trading minutes only
for complete one-minute input. It does not impute missing observations or validate whether
an externally supplied schedule matches an exchange announcement.
🚨 session_bars is a separate synthetic comparison: it counts available rows and compresses
non-trading time, so a bucket can span lunch. It requires complete one-minute observations.
Do not use it to hide holes in a delivered dataset.
✅ Measured 2026-09-14: on the demo's fixed Tokyo-like schedule, wall-clock half-hour resampling
creates 16764 buckets, only 2500 containing observations. At hourly resolution, wall-clock
counts are [60, 60, 30, 30, 60, 60], while compressed trading-time counts are
[60, 60, 60, 60, 60]. The synthetic process has no lunch shock; partial buckets have
0.4768 times the variance of full buckets in this seed. This is a bucket-duration comparison,
not evidence that actual lunch or overnight volatility is zero. The synthetic weekdays and
fixed historical shape are not an up-to-date Tokyo schedule.
Cross-venue holes and library disagreements
✅ Measured 2026-09-14: bundled 2024 closure tables agree with installed
exchange_calendars 4.13.2. XNYS has 252 sessions and XTKS 245; 16 occur only on XNYS and
9 only on XTKS. Reindexing Tokyo onto New York adds holes and drops Tokyo-only observations.
Keep the holes and a venue-open mask rather than treating carried prices as fresh trades.
✅ Measured with pandas_market_calendars 5.4.0 and holidays 0.104, 2026-09-14:
holidays.US(years=2024)disagrees with XNYS on 3 dates. Its federal-holiday list is not the financial calendar;financial_holidays("NYSE", years=2024)agrees in this test.- The HKEX/XHKG libraries disagree on 2024-09-06. Compare against an exchange notice before choosing one; library agreement alone is not exchange verification.
- The installed NSE calendar returns all 261 weekdays for 2027, while XBOM rejects that requested window. This flags incomplete forward coverage; neither result verifies the actual future exchange schedule. NSE and BSE are different venues, not interchangeable.
Run python scripts/calendars.py from this skill directory to reproduce the table and
optional library comparisons. Re-run after a calendar dependency changes. The bundled closures
cover only 2024 and intentionally reject other years; use a maintained schedule in production.