Percentiles and short histories
A percentile on Tipsheet is the share of a series’ past readings at or below today’s: “higher than 80% of readings since 2005”. It says where today sits in the series’ own history, so it is only as good as that history.
The rule. A percentile, a median or a usual range is published only when the history behind it covers at least five years and holds enough readings for its frequency:
| Frequency | Readings needed |
|---|---|
| Daily | 1,000 |
| Weekly | 208 |
| Monthly | 48 |
| Quarterly | 16 |
| Annual | 10 (and ten years) |
The counts are four-fifths of what five years would hold, so a series with gaps in it does not pass on its span alone. Below the rule, the statistics bundle carries null for the percentile, and the page says when the readings start and from when they will be ranked (“Readings since Feb 2023, too few to rank against; ranked from Feb 2028”).
Why five years. With a few months of history a percentile describes those months, not the series: on 29 Oct 2026, 101 of the 125 index pages showed valuation percentiles computed from three month-ends, because NSE’s valuation files for those indices reach back only to August 2026 in our archive. Five years is roughly one business cycle in Indian data, and it is the same rule the site already used for a reading’s usual range, so the two never disagree. Studies may set a longer floor of their own, never a shorter one; the exceptions are listed below.
Where it applies. Every statistics bundle (percentile_full_history, percentile_last_5y), the India macro snapshot, the valuation table on each index page and the /valuation/ snapshot (percentiles, median, low and high), the economy registry, composites, breadth and financial-conditions facts, the 10-year yield’s place in its range, the relative-valuation readings (median and percentile null, short: true), CFTC positioning against its full history, and the US VIX percentile to date (blank for its first five years). The window percentiles keep their own extra conditions: ten years of index valuations need 60 month-ends, and the CAPE percentile 60 months of CAPE.
Windows. A percentile over a named window (“within the last 3 years”, “5-year”) is published only once the history fills that window and holds four-fifths of the readings it would: FRED’s 5-year percentiles, the credit spreads’ 3- and 5-year ones, CFTC’s 3-year (125 of 156 weeks) and the Treasury auctions’ 5-year. Before 30 Oct 2026 some of these ranked as few as 26 weeks or 20 auctions.
What sets its own rule. Some studies set a longer floor (the global questions and the global tracker use eight years; the world-India study 60 month-ends). One reading is a regime window by design: the F&O put-call ratio is ranked only since the November 2024 F&O curbs (pcr_5d_pctile_post_curbs, from 60 sessions), because earlier readings come from a different market. Model and strategy inputs (the barometers’ 756-session burn-in, the IPO heat gauge, the lab’s and models’ valuation rules, the sector valuation study) rank inside their own specs and are not readings a page quotes as “where today sits”. pipeline/tests/test_percentile_floor.py lists each of these with its reason and fails on any new percentile outside publish/bundles.py until it is floored or listed.
Code: pipeline/tipsheet/publish/bundles.py (MIN_HISTORY_YEARS, MIN_OBSERVATIONS, enough_history with a window’s years, floor_start for percentiles to date, percentile) and apps/site/src/lib/reading.ts (MIN_YEARS, tooShort).