Should holidays be excluded from statistics, or are they expected traffic? #202

New issue

Closed

opened 2026-10-02 14:57:29 +00:00 by gabogg · 1 comment

gabogg commented

2026-10-02 14:57:29 +00:00

Owner

Question

Should Holiday Days (and Event Days) be excluded from statistics at all?

#161 treats them as atypical: they are left out of the Usual Weekday Baseline, multiplier learning, uncertainty and maturity, and (decided 2026-10-02) the counter-spike reference. The working assumption is that a holiday skews the numbers.

The maintainer questions that framing. Holidays are no secret: they happen every year on known dates and reliably move traffic. If an effect is expected and recurring, is it really a skew, or is it part of the facility's normal pattern that statistics should reflect?

This needs discussion before any change. It's deliberately not scheduled yet.

What to settle

Language. Is "atypical" or "skewed" the right description for a known, recurring holiday? Or should CONTEXT.md describe holidays as a different kind of normal, with their own expected pattern?
Baselines. The Usual Weekday Baseline answers "what does an ordinary Tuesday look like?" If holidays stay out of it, should there be a separate holiday baseline (same holiday last year, or holidays in general), so a holiday is compared with its own kind instead of with nothing?
Period totals and KPIs. Week, month and year totals already include holidays. Should comparisons such as month over month or year over year call out holiday-count differences between the periods (e.g. "this month had 2 more holidays")? Or should they normalise for them?
Investor view. For analytics shown to investors, is it more honest to show holidays in the trend with a marker, or to remove them?
Learning and integrity. These are separate from statistics:
- multiplier learning;
- the counter-spike check;
- uncertainty and maturity.
Their exclusion protects the calibration model rather than describing traffic, so they may keep excluding holidays whatever is decided for statistics. Confirm that split.
Event Days. Do the same answers apply to Event Days? An event is less predictable than a calendar holiday.

Context

#161 (holiday and event context) and its implementation PR #178 currently exclude Holiday and Event Days from baselines and learning.
Maintainer decisions recorded on #161 (2026-10-02):
- existing holidays stay holidays;
- the spike check skips them and excludes them from its reference.
CONTEXT.md: Holiday, Event Day, Usual Weekday Baseline, Schedule Exception.

Outcome

A recorded decision (comment here, then CONTEXT.md and possibly an ADR) on how holidays and events are treated in each of: baselines, period KPIs and comparisons, investor presentation, learning, and integrity checks. Follow-up issues come from that decision.

## Question Should Holiday Days (and Event Days) be excluded from statistics at all? #161 treats them as atypical: they are left out of the Usual Weekday Baseline, multiplier learning, uncertainty and maturity, and (decided 2026-10-02) the counter-spike reference. The working assumption is that a holiday *skews* the numbers. The maintainer questions that framing. Holidays are no secret: they happen every year on known dates and reliably move traffic. If an effect is expected and recurring, is it really a skew, or is it part of the facility's normal pattern that statistics should reflect? **This needs discussion before any change. It's deliberately not scheduled yet.** ## What to settle 1. **Language.** Is "atypical" or "skewed" the right description for a known, recurring holiday? Or should CONTEXT.md describe holidays as a *different kind of normal*, with their own expected pattern? 2. **Baselines.** The Usual Weekday Baseline answers "what does an ordinary Tuesday look like?" If holidays stay out of it, should there be a separate **holiday baseline** (same holiday last year, or holidays in general), so a holiday is compared with its own kind instead of with nothing? 3. **Period totals and KPIs.** Week, month and year totals already include holidays. Should comparisons such as month over month or year over year call out holiday-count differences between the periods (e.g. "this month had 2 more holidays")? Or should they normalise for them? 4. **Investor view.** For analytics shown to investors, is it more honest to show holidays in the trend with a marker, or to remove them? 5. **Learning and integrity.** These are separate from statistics: - multiplier learning; - the counter-spike check; - uncertainty and maturity. Their exclusion protects the calibration model rather than describing traffic, so they may keep excluding holidays whatever is decided for statistics. Confirm that split. 6. **Event Days.** Do the same answers apply to Event Days? An event is less predictable than a calendar holiday. ## Context - #161 (holiday and event context) and its implementation PR #178 currently exclude Holiday and Event Days from baselines and learning. - Maintainer decisions recorded on #161 (2026-10-02): - existing holidays stay holidays; - the spike check skips them and excludes them from its reference. - CONTEXT.md: Holiday, Event Day, Usual Weekday Baseline, Schedule Exception. ## Outcome A recorded decision (comment here, then CONTEXT.md and possibly an ADR) on how holidays and events are treated in each of: baselines, period KPIs and comparisons, investor presentation, learning, and integrity checks. Follow-up issues come from that decision.

gabogg added this to the Historical passenger-flow backfill milestone

2026-10-02 14:57:29 +00:00

gabogg added the

needs-triage

label

2026-10-02 14:57:29 +00:00

gabogg referenced this issue

2026-10-02 14:57:43 +00:00

feat(statistics): holiday and event context for investor analytics #161

gabogg referenced this issue

2026-10-02 15:41:47 +00:00

feat(schedule): movable holidays defined by recurrence rules #204

gabogg referenced this issue

2026-10-02 15:41:49 +00:00

feat(statistics): holiday baseline — compare a holiday with the same holiday in earlier years #205

gabogg referenced this issue

2026-10-02 15:41:52 +00:00

feat(statistics): show holiday and event counts on period comparisons #206

gabogg commented

2026-10-02 15:42:51 +00:00

Author

Owner

Triage decision, 2026-10-02 (maintainer grilling session)

The current design already treats holidays as expected traffic. Holidays are not excluded from statistics. Closing this issue: the question is answered, and its follow-ups are filed separately.

What is actually excluded.
- Holidays and Event Days stay in every total, KPI, daily series and investor view, with their marker and name.
- They are kept out only of the ordinary-day references: the Usual Weekday Baseline (as a source), multiplier learning, variance and maturity, and the counter-spike reference.
- They are still compared against the baseline (e.g. "Christmas: 2.3× a usual Tuesday"). The traffic is never hidden; it just doesn't define what an ordinary day is.
Language. A Holiday or Event Day is expected, but not ordinary. It differs from the Usual Weekday Baseline by design, and that difference is information, not error. Avoid "skewed", "anomalous" and "outlier"; "outlier" belongs to sensor faults. The glossary changes go in a docs PR.
Holiday Baseline (compare with the same holiday in earlier years): a future feature, outside this milestone. See #205. It depends on movable-holiday rules (#204) for Carnival, Easter and similar holidays.
Different holiday counts in period comparisons: annotate the counts with no normalization. A future feature, outside this milestone: #206.
Learning and integrity stay separate from statistics. Multiplier learning, variance, maturity and the spike check keep excluding holidays and events, whatever statistics later shows. They protect the counting model and don't describe traffic.
Event Days get the same treatment: in totals with a marker, out of the ordinary-day references. They don't get a "same event last year" reference.

🤖 Generated with Claude Code

## Triage decision, 2026-10-02 (maintainer grilling session) **The current design already treats holidays as expected traffic.** Holidays are not excluded from statistics. Closing this issue: the question is answered, and its follow-ups are filed separately. 1. **What is actually excluded.** - Holidays and Event Days stay in **every total, KPI, daily series and investor view**, with their marker and name. - They are kept out only of the **ordinary-day references**: the Usual Weekday Baseline (as a source), multiplier learning, variance and maturity, and the counter-spike reference. - They are still compared *against* the baseline (e.g. "Christmas: 2.3× a usual Tuesday"). The traffic is never hidden; it just doesn't define what an ordinary day is. 2. **Language.** A Holiday or Event Day is **expected, but not ordinary**. It differs from the Usual Weekday Baseline by design, and that difference is information, not error. Avoid "skewed", "anomalous" and "outlier"; "outlier" belongs to sensor faults. The glossary changes go in a docs PR. 3. **Holiday Baseline** (compare with the same holiday in earlier years): a future feature, outside this milestone. See #205. It depends on movable-holiday rules (#204) for Carnival, Easter and similar holidays. 4. **Different holiday counts in period comparisons:** annotate the counts with no normalization. A future feature, outside this milestone: #206. 5. **Learning and integrity stay separate from statistics.** Multiplier learning, variance, maturity and the spike check keep excluding holidays and events, whatever statistics later shows. They protect the counting model and don't describe traffic. 6. **Event Days get the same treatment:** in totals with a marker, out of the ordinary-day references. They don't get a "same event last year" reference. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

gabogg referenced this issue

2026-10-02 15:42:54 +00:00

feat(occupancy): curate monthly historical flow drafts #170

gabogg

2026-10-02 15:43:01 +00:00