Methodology — Conditioning Day Sets (All | Up | Down | Stress)

Transparency is a feature. This page explains what the conditioning day sets are and why we compute them — in plain terms. The precise formulas and numeric conventions are documented in our internal methodology; the summary here is deliberate, not an omission.

Several risk displays ask how a book behaves on a subset of days — on days the market fell, or on genuinely bad days. Those subsets are resolved once, by one shared service, and every conditioned number reads the same sets. If two panels ever disagree about a "stress" figure, it is because they compute different statistics — never because they quietly picked different days.

The four sets

The conditioning series is the benchmark you selected (SPY by default), aligned day-for-day with the book's own return window:

• All — every day in the aligned window.
• Up — days the benchmark rose.
• Down — days the benchmark fell.
• Stress — the benchmark's worst-decile days: the roughly one-in-ten worst benchmark returns of the window, selected by an exact cutoff at the decile boundary. Days exactly at the boundary are all included, so the set can be slightly larger than a tenth — the payload reports the honest count. Days the benchmark was exactly flat belong to neither Up nor Down; we disclose how many there were rather than silently assigning them a side.

A window too short to have a decile (fewer than ten days) gets an empty stress set — not an invented one. Without a benchmark there are no sets at all, and anything that needs them says so instead of guessing.

Why day counts alone aren't enough

Sixty stress days drawn from a single crash are one dated market event — a fact worth showing, but not a basis for estimating anything. So beside the day count we report how many distinct market episodes the stress days span, using the same peak-to-recovery drawdown-episode definition as the drawdown table (no new clustering rule). A stress-conditioned estimate renders only when the set holds at least 50 days spanning at least 2 episodes; below that, the days render as dated fact and the estimate is withheld, with the reason shown.

An honest warning about conditioned correlations

Correlations measured on down or stress days come out mechanically higher than the full-window number even when nothing about the true relationship changed — selecting volatile days does that by construction (Boyer, Gibson & Loretan, 1997, Pitfalls in Tests for Changes in Correlations, Federal Reserve Board IFDP 597). Every display built on these sets therefore carries a mechanical baseline beside it, so only collapse beyond the mechanical effect reads as information.