Intermarket-Matrix

When gold rises unusually sharply — does oil follow? When bitcoin drops, does the S&P 500 go with it? This page does not test such questions one at a time but all of them at once: every pair among 18 markets, in both directions. And it accounts for the fact that asking this many questions guarantees false hits.

Important: an analysis of historical price data, not a forecast and not investment advice. Since no relationship passes the test, there is nothing here to read as a signal. Conspicuous cells are labelled hints — which is emphatically not the same thing as a finding.
 

Asking 470 questions at once produces roughly 24 apparent hits by chance alone — even if there is no relationship between any of these markets. Anyone who computes such a matrix and then picks out the most striking cells will always find something. To count as a finding here, a cell has to be larger than the best cell of a matrix built from pure noise — and show the same thing in both halves of the history. None does.

The matrix

Row = the market that moved. Column = the market whose development is measured afterwards. The number is the excess in percentage points over what that market does in ten trading days anyway. An empty field means: too few events, not evaluable. Click a cell to see all its numbers.

The most conspicuous cells — and why none of them is a finding

Sorted by the strength of the standardised effect. t puts the excess in relation to its own dispersion — two percentage points in bonds is a different statement from two percentage points in bitcoin. The H1/H2 column shows how many events fall into the first and the second half of the period.

Related: the most conspicuous relationship in this matrix — sharp rises in long-dated government bonds and what happens in equities afterwards — has been examined separately and deliberately, with controls for market regime and sub-periods: Bonds as a leading indicator. The same for crypto: Does bitcoin lead the stock market?

Try it yourself: the Shock Analyser lets you choose trigger, target, threshold and time window freely. It computes every combination on its own — without the correction this page applies. That is precisely why the two pages belong together: explore there, then check here what survives.
Methodology — how this was computed

What counts as an event

A move in the signalling market over ten trading days that ranks among the strongest ten percent of its own history. The threshold is deliberately not a fixed percentage: five percent is a rarely seen jump in gold and the most ordinary move in the world in bitcoin. This makes an event equally rare in every row, so the rows become comparable. At least 41 trading days separate two events, so the measured windows do not overlap.

What is measured

Only what comes after the event: the following ten trading days. A simultaneous move would not be a forecast, merely the observation that two markets moved on the same day. Every number is read against the base rate of the target market — what that market does in ten trading days anyway — and not against zero.

The correction, and why it governs everything here

A matrix asks many questions at once, and the more questions you ask, the more certainly you find answers that are not there. So the question is not “is this cell conspicuous” but: how large does the best cell get when the same matrix is built from pure noise? To find out, the price series are shifted circularly against each other 2000 times — which destroys the temporal relationship while preserving the way quiet and turbulent phases cluster in blocks. Each round uses one shift shared by all cells, so that the dependence between similar pairs is preserved: technology, semiconductors and software largely measure the same thing.

Why the comparison is in standard errors

A first pass compared raw percentage points. Bitcoin and uranium as targets then set a bar of eleven percentage points that no calmer market could ever clear — and the matrix would only have measured which market swings hardest. The comparison therefore uses the effect divided by its own standard error.

The four hurdles for a finding

  • The pair must belong to the primary family, connecting two different categories. That the Dow follows the S&P 500 is not an intermarket hypothesis but the same market under another name. Such pairs are still shown, as replication cells — they cannot produce a finding. The rule separates by category, not by measured correlation; otherwise the choice of family would itself be a search for the result.
  • The standardised effect must clear the max-T bar.
  • The effect must have the same sign in both halves of the period.
  • Each half must contain at least 20 events. Twenty in total can mean three in one half and seventeen in the other — “holds in both halves” would then be a statement about three cases.

What this page cannot do

  • Large moves cluster in crises. If something follows, that may be down to the shared market environment rather than one market leading another. The matrix does not separate the two; the dedicated bond study does.
  • Bonds 20Y+ and high yield are ETF prices here, not yields and not spreads. A rising bond ETF corresponds to falling yields.
  • The price series are dividend-adjusted. On distribution days this carries a small artificial markdown into the series — more noticeable in monthly-paying bond ETFs than in gold.
  • Every pair uses its own shared history. Uranium is available from 2010, ether from 2017, the S&P 500 from 1993. A cell with 6000 observations carries more weight than one with 1200; both numbers therefore appear in the detail view. Mag 7, with history only from April 2023, never reaches 20 events and does not appear in the matrix at all.