Innova Castle
All findings
ResultEvidence: Real external data

On the S&P 500, structural tension runs higher during declines, and separates them better out of sample than in sample

Over 24,772 daily sessions since 1927, the Market Stress Index reads 22.5 points higher during decline phases, on the segment it was never calibrated on.

Market Stress Index (MSI)

Time Series Analysis Platform

Validation

01

The question

Most risk models are re-tuned after each downturn, which makes it impossible to tell whether they measure stress or merely fit the past. We wanted to know how a single frozen reading behaves across a market history it was never calibrated on.

02

Method

  1. The Market Stress Index reads a market as a number from 0 to 100 with four states. It measures the structural state of the system; it does not predict prices, dates or probabilities.
  2. Data: Yahoo Finance (^GSPC), a price index at daily close. 24,772 sessions from 1927 to 2026, the S&P 500 only.
  3. Decline phases were located separately, from price alone, without the index. The index is never told that crises exist; it only reads the price series.
  4. The reading was set using the first half of the century. The second half, which the method never saw, was used to check it. Everything reported here is measured on that second half.
  5. Every step is white box: no neural network and no training set, so any reading traces back to the price that produced it.
The price, on its own

No index here. This is the input the reading is computed from, and the only input there is.

S&P 500 · 24,772 daily sessions
19291949196919892009
S&P 500 closeDecline phase
Vertical axis: logarithmic, so the same percentage fall takes the same height in any decade

03

Result

S&P 500 sessions, 1927-2026
24,772S&P 500 sessions, 1927-2026
points higher during decline phases, out of sample
+22.5points higher during decline phases, out of sample
major declines of the century with high tension
9 / 12major declines of the century with high tension
of those, with tension above the median for months beforehand
8 / 12of those, with tension above the median for months beforehand

On the segment it was never calibrated on, the index reads 22.5 points higher during decline phases than outside them. It separates the two better out of sample than in sample, the opposite of what an overfitted model does.

Of the twelve major declines of the century, nine coincided with a high-tension reading. In eight of the twelve, that tension had sat above the median for months beforehand, measured with the same yardstick. This describes the state that preceded them; it is not a claim that the index announced anything.

High tension is not the same as a coming decline. Episodes of high tension that corrected without a decline lasted 95 sessions at the median; those that ended in a decline lasted 273. A market can sit in a tense state for a long time.

A century measured

The monthly reading of the index, 1929-2026. Shaded: the century's twelve declines of 20% or more, located from price alone.

S&P 500 · 24,772 daily sessions
19291949196919892009

The six declines with a name of their own, out of the twelve marked above. The fall is the price. The reading is what the index showed while that fall was running, not a warning issued before it.

  • 1929 · Crash
    Price fall
    -86.2%
    Highest reading
    86 / 100
  • 1973 · Oil shock
    Price fall
    -48.2%
    Highest reading
    94 / 100
  • 1987 · Black Monday
    Price fall
    -33.5%
    Highest reading
    72 / 100
  • 2000 · Dot-com
    Price fall
    -49.1%
    Highest reading
    85 / 100
  • 2008 · Financial crisis
    Price fall
    -56.8%
    Highest reading
    94 / 100
  • 2020 · Covid
    Price fall
    -33.9%
    Highest reading
    56 / 100
NormalElevatedHighSevere
Dashed line: historical median, 58 / 100
Price and reading, together

To read one against the other: when the reading moved, and what the price was doing then.

19291949196919892009
S&P 500 closeReading, 0-100
Left axis: price, logarithmic · right axis: reading, 0-100 · the relative height of the two lines means nothing
State, month by month

Each cell is the median of its month, not its worst day. Scroll right to walk the century.

Arrow keys to walk the grid, or hover a month
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
1930
1940
1950
1960
1970
1980
1990
2000
2010
2020

Marks: a dot from high, a bar at severe, so the states read without colour

NormalElevatedHighSevere
Columns: years · rows: months

04

What this does not show

Limits of this finding

  • This is measured over history that already existed. It is not evidence of behaviour on a market whose outcome is still unknown.
  • The index measures tension, it does not anticipate declines: it predicts no price, direction, date or magnitude, and high tension is not a coming decline.
  • This is calibrated on the S&P 500 and holds for the S&P 500 only. It does not carry over to other indices, assets or markets; that would be separate work, evidenced on its own.

05

Related findings

See what the research becomes.

Explore the technology emerging from our work, or start a conversation about where it could be useful.

Explore the technologyStart a conversation