The windows on this dashboard differ by a factor of fifty, and the differences come from three different places — only one of which is under your control, and you cannot tell which from the number.
Where the unequal windows come from
It would be reasonable to assume a spread of window lengths reflects decisions someone made. Mostly it does not — most of these are imposed from outside, by different mechanisms worth knowing because they recur in every data product. But not all of them, and the exception is the one that matters, because it is invisible from the surface and it is yours.
Licensing. The high-yield credit spread series is trimmed to roughly three trailing years — the index behind it is licensed, and the public distribution carries only a window. This was verified live rather than assumed: the series starts where it starts even in the public download. No amount of engineering extends it, and no reasonable person would guess it from looking at the chart.
Your own default, wearing an external limit's costume. The policy-rate window is about four decades, and the obvious story — that is how far back the series goes — is false. The public source publishes that rate daily from 1954, seventy-plus years. The window is forty-one because this engine's FRED helper takes a start parameter defaulting to 1985-01-01 and no caller overrides it. The same default silently truncates two currency series that are published from the 1970s and 1981.
That one is worth sitting with, because it is the inverse of the licence case and indistinguishable from it at the surface. Both present as "this series carries N years." One is imposed by a rights holder and no engineering can extend it. The other is a line in your own ingest layer that nobody has revisited since it was written. You cannot tell them apart from the number. Which means the discipline is not just "disclose the window" — it is know, for each series, whether the window is the world's answer or yours.
Scholarly reconstruction. The CAPE inputs reach back to the 1880s because an academic assembled the long series and published it — a genuinely unusual artifact, and the reason one gauge on this dashboard can make a statement about a century and a half at all.
Why not normalize them?
The instinctive engineering fix is to trim everything to the shortest common window so that all percentiles are computed over the same span. It is a clean idea and it is wrong here, for a reason worth internalizing.
Trimming a century of history down to three years to match its shortest neighbor does not make the two comparable. It destroys the most valuable thing in the entire dataset — a genuinely long series is rare, expensive, and irreplaceable — in exchange for a cosmetic consistency. The reader gains nothing: they now have two percentiles that agree in span and both say very little.
The alternative is to keep each series' full history and carry the window as data, so the interface can show it and any consumer can reason about it. Disclosure preserves information; normalization destroys it. When those are the two options, disclosure wins every time.