Verification

Verification

What was checked, against what, over which range, and what the checks cannot reach. A calendar can be wrong in a way a moving average cannot, so this page states its evidence rather than its confidence.

The oracles

Layer	Checked against
Civil conversions	Python datetime, date.toordinal
Easter and Good Friday	dateutil
Zone rules	zoneinfo, the IANA time zone database
Exchange calendars	exchange_calendars (XNYS, XLON, CMES, XTKS, XEUR, XHKG, XASX, XTSE, XSHG, XBOM, XSES)
Day counts	The ISDA 2006 published worked examples
VIX settlement	The published Cboe calendars
Pattern formatter	strftime plus pinned Java-convention vectors

Civil core

Round-trips against date.toordinal over 1970 to 2200: 4,716 of 4,716 exact. Every day-of-week from 2000 to 2100 brute-forced, along with the adjusters at nth +1 to +5 and -1 to -3. Easter and Good Friday match dateutil across 1970 to 2200.

Zone engine

Brute force rather than sampling: every transition instant the IANA database contains, in every modelled zone, probed at one second before, at, and one second after. 1,106 transition instants matched exactly across 1970 to 2100, with 16,416 gap and overlap probes corroborated against IANA's own fold=0 and fold=1 answers.

Exchange calendars

Every calendar was diffed date by date against its reference over the range below, and every one reaches zero differences apart from the three deliberate divergences named further down.

Calendar	Reference	Diffed over	Sessions
NYSE	XNYS	1976-01-01 to 2035-12-31	7,794
LSE	XLON	1976-01-01 to 2035-12-31	7,833
CME	CMES	2000-01-01 to 2035-12-31	7,995
JPX	XTKS	2006-08-14 to 2027-08-13
EUREX	XEUR	2006-08-14 to 2027-08-13
HKEX	XHKG	2000-01-01 to 2049-12-31	12,336
ASX	XASX	1995-01-01 to 2049-12-31	13,923
TSX	XTSE	1995-01-01 to 2049-12-31	13,811
SSE	XSHG	2005-12-26 to 2026-12-31	5,101
BSE	XBOM	2006-01-01 to 2026-12-31
SGX	XSES	2006-01-01 to 2026-12-31

Lunch breaks were checked as instants, not just as flags. The Hong Kong break matches XHKG to the millisecond on all 12,336 sessions, including the 99 half days. The Tokyo 11:30 to 12:30 break matches XTKS on all 5,135 sessions from 2006 to 2027. Shanghai's 11:30 to 13:00 break matches XSHG on all 5,101.

The rules themselves, not a restatement of them

This is the check worth understanding, because until it existed the suite had a hole.

Every other harness diffs a hand-written Python mirror against an exchange calendar. That checks the mirror. It does not check the Pine: inserting a spurious holiday into the library's own ASX rule chain left all eight harnesses green, because nothing read the Pine's rule bodies.

The fix transpiles the holiday-name chains straight out of the Pine source and executes them. All 25 transpiled functions now run against all eight oracles and every one diffs to zero: NYSE and LSE over 15,652 weekdays each, CME 9,391, JPX and EUREX 5,480 each, HKEX 13,045, ASX and TSX 14,349 each. Replaying the ASX mutation now fails, with 55 mismatches, and fails on ASX alone.

Sessions, expiries and day counts

5,276 US sessions are all exactly 390 minutes, or 210 on half days, across every daylight-saving transition from 2015 to 2035. Over 1,004 sessions, is_last_bar flags exactly the closing minute bar and neither neighbour.

The VIX rule reproduces all 72 published Cboe settlement dates from 2021 to 2026, including all four Tuesday exceptions, pinned as literals.

30/360 and 30E/360 reproduce 14 published ISDA worked examples, both as year fractions and as raw integer numerators.

Three deliberate divergences

Each is pinned in both directions, so the suite fails if the divergence grows or disappears.

1. CME session close. The library models Globex equity index closing at 16:00 Chicago; CMES labels it 17:00. Opens must match to the millisecond and closes must be exactly one hour earlier.
2. CME, 11 to 14 September 2001. exchange_calendars lists those four days as ordinary sessions. CME was down from the 11th and the equity-index complex did not resume until the 17th. The library closes them and exempts them from the diff. This is a gap in the reference, not an error here.
3. JPX close, 5 November to 30 December 2024. The library models the TSE close move at year granularity, so those sessions close 30 minutes early. Every divergent session is walked and pinned rather than sampled.

What the checks cannot reach

Stated because a verification page that lists only its successes is not evidence.

• The FOMC table has no external oracle available offline. It is checked structurally instead. Those checks would catch a mistyped digit, a duplicated line or a missing meeting. They would not catch a meeting that moved by one day and stayed a Tuesday. The retrieval date and source URL are pinned in the source for re-diffing.
• Pine's copy-on-assignment and UDT aliasing are not exercised. No Python harness can exercise them. See Value Semantics for the behaviour you have to hold in your head instead.
• period_between has no independent oracle. It is verified by its defining round-trip property, with OpenJDK's LocalDate.until transcribed alongside for comparison. Of 4,550 date pairs the two disagree on 18, and in all 18 java.time fails its own documented round-trip promise.
• Session instants in daylight-saving zones hold only to 2037. Past the 32-bit boundary the reference misapplies daylight saving, so there is nothing to diff against. Hong Kong, on a fixed offset, runs clean to 2049.
• NYSE and LSE rules run unchecked before 1976, and JPX from 2000 to 2005, because the references do not reach there.
• Unscheduled closures are recorded for the past only. No table predicts the next hurricane.
• VIX past the last published Cboe calendar is projection. Cboe can move any future date by circular, which no rule predicts.

---

Previous: Scope and Limitations  ·  Next: Performance and Limits  ·  See also: Versioning and Data Currency