Opinion · August 2026
The Index Was Calm. The Names Underneath It Weren’t.
The S&P 500 rose 0.41% on Friday. VXX, the market’s standard fear gauge, fell 1.25% to a level it hasn’t traded below since early July. Both are honest readings, and both describe a market that wasn’t there. Across the 47 names I track — 32 holdings, 15 watchlist — Friday’s moves ran from Marvell’s −5.57% to Tempus AI’s +9.06%, a 14.63-point spread, with a standard deviation of 2.95 points around a mean of +0.79%. The index was quiet. The tape was not.
The day’s dispersion in one view
Figure 1 · All 47 tracked names, one session
A 14.63-point spread hiding inside a 0.41% index move
Each dot is one of AEA’s 32 holdings or 15 watchlist names, August 21 close vs. August 20 close. Computed directly from Massive Market Data’s grouped daily bars — mean +0.79%, standard deviation 2.95 points. Six names moved more than 2 standard deviations from the mean.
Most of the 47 dots sit close to the centerline — the quiet middle the index return actually describes. But the six names in the tails (Marvell, Applied Digital and Arm on the downside; Tempus AI, Coinbase and IonQ on the up) each moved more than two standard deviations from that day’s own mean. An index return is, definitionally, a weighted average. Friday is a clean demonstration of how much can be hiding inside an average that looks unremarkable.
What the fear gauge actually measures
VXX tracks near-term VIX futures, and the VIX itself is derived from S&P 500 index options — a cap-weighted basket dominated by its largest constituents. It is built to price expected volatility in the index as a whole, not the dispersion of returns across a set of individual names. A basket can be perfectly calm at the aggregate level while the names inside it pull in opposite directions for entirely unrelated reasons, which is exactly what today’s News article documented: a crypto-regulation story, a two-week-old clinical-trial reaction, and an unexplained decline, netting to a 0.41% index move. VXX has no mechanism for distinguishing that from an actually quiet day, because it isn’t built to.
Figure 2 · VXX, last ~20 sessions
A month-long grind lower that never paused for the single-name moves
VXX daily close, July 24 – August 21, 2026. Down 15.25% over the full window and 2.09% over just the last five sessions — a period that included the AI-infrastructure debt-financing selloff, a second-day divergence within that cohort, and Friday’s crypto/Tempus/Marvell dispersion. Source: Massive Market Data daily bars.
None of those episodes registered as a pause in VXX’s decline. That isn’t VXX malfunctioning — it’s VXX doing exactly what it’s built to do, which is price the index, not the cross-section. The gap between the two is the entire subject of this piece.
Index-level gauges vs. cross-sectional reality
| Metric | Value |
|---|---|
| S&P 500 (SPY), day change | +0.41% |
| Nasdaq-100 (QQQ), day change | +0.35% |
| VXX, day change | −1.25% |
| VXX, five-session change (Aug 14–21) | −2.09% |
| 47-name universe, mean move | +0.79% |
| 47-name universe, standard deviation | 2.95 pts |
| 47-name universe, best–worst spread | 14.63 pts |
| Marvell (MRVL), distance from mean | −2.16σ |
| Tempus AI (TEM), distance from mean | +2.80σ |
Universe = AEA’s 32 holdings and 15 watchlist names. Standard deviation is the population statistic across all 47 single-day moves, not an annualized or option-implied figure. Source: Massive Market Data grouped daily bars, computed directly from closing prices.
Marvell in context
Marvell’s 5.57% decline, unexplained by any dated headline per the linked News piece, sits at −2.16 standard deviations from Friday’s own cross-sectional mean — a real outlier by the day’s own distribution, but not an outlier so extreme it demands a story. In a 47-name sample with a 2.95-point standard deviation, a move beyond 2σ in either direction is not a rare event; it is close to what a wide, noisy tape produces on an ordinary Friday. That is a different claim from “nothing happened to Marvell” — it is the claim that Friday’s dispersion was wide enough that an unexplained 2σ move fits comfortably inside it, which is itself a reason not to force a causal story onto a move statistics alone can account for.
A projection: what this dispersion regime implies
Rather than guess whether Friday’s spread repeats, here is the arithmetic for three dispersion regimes, using Friday’s observed mean (+0.79%) and applying a normal approximation to project the range that would contain roughly 95% (±2σ) and 99.7% (±3σ) of names in a 47-name sample.
| Dispersion regime | ±2σ range | ±3σ range |
|---|---|---|
| Narrower 2.0-point standard deviation | −3.2% to +4.8% | −5.2% to +6.8% |
| Friday’s level 2.95-point standard deviation | −5.1% to +6.7% | −8.1% to +9.6% |
| 50% wider 4.4-point standard deviation | −8.0% to +9.6% | −12.4% to +14.0% |
Assumptions, stated plainly. This is not a forecast of which name moves, or of tomorrow’s dispersion — it holds Friday’s +0.79% mean fixed and only varies the standard deviation, then applies the normal-distribution rule of thumb for how wide a ±2σ / ±3σ band would be under each assumption. Single-stock daily returns are not normally distributed in practice; they have fatter tails than this model implies, which means the true bands are almost certainly wider than shown, especially on the downside. What this table is for: converting “dispersion was wide Friday” into a concrete sense of what “wide” and “wider” actually look like in percentage terms, not predicting Monday.
Where this connects
This is the same underlying idea as AEA’s own Volatility & Correlation Engine, which computes real annualized volatility and pairwise correlation from the same daily closes powering every position’s price chart on this site — the tool exists precisely because a single index-level number was never going to describe 32 individual holdings well. It also follows directly from Wednesday’s observation that a cohort hit by the same shock didn’t move together on day two. Two sessions apart, the same lesson: aggregates are convenient and can be quietly misleading, and the correction is to keep looking at the distribution underneath them, not just its average.