Independent strategy research
Does intermarket analysis actually work? We tested 400+ signals
Short version: almost none of it predicts direction. We ran 400+ cross-asset directional configurations through honest multiple-testing correction, across equities, bonds, gold, oil, FX and crypto. One survived. The rest are simultaneous co-movement dressed up as a lead.
The promise, and why it rarely holds
The oldest claim in market lore is that one asset leads another. Bonds lead stocks. The dollar leads commodities. Crypto leads risk appetite over the weekend. It is intuitive, it is everywhere in the education content, and it is mostly an artifact of two things moving at once. Intermarket analysis reads a genuine relationship — assets do co-move — and then quietly assumes co-movement means one side arrives first with tradeable warning. We tested that assumption directly.
Here is the trap. Run enough asset pairs at enough resolutions and some combination will look predictive by luck alone. Sample two dozen leads, and one clears a t-stat. That is not an edge. That is arithmetic. The fix is multiple-testing correction with the denominator set to every config you ran, not just the pretty one you kept.
What we tested (name-free — family and failure mode only)
Directional dependence at resolutions from 1-second to monthly, with a hard rule: the bars had to be synchronous. Fixed-time intraday futures snapshots, not misaligned cash closes. Mismatched closes are a classic manufacturer of fake lead-lag — the New York close reacts to something the Tokyo close already saw, and a chart calls it prediction. Every candidate also had to beat a volatility null: a benchmark where the sign is predictable from drift and volatility alone, with zero cross-asset information. If a "signal" can't beat that, it never knew anything about the other market.
| Family (what was tested) | Scope | Verdict | Why it failed |
|---|---|---|---|
| Cross-asset lead-lag, all pairs | 260+ pairs | NULL | no lead beyond simultaneous co-movement |
| Regime-conditional lead-lag | 130+ configs | NULL | conditioning on volatility doesn't rescue direction |
| Nonlinear machine-learning lead grid | 180+ pairs | NULL | "hidden" leads were a single-month artifact; didn't replicate |
| Calendar premia (announcement, month-end, OpEx, auctions) | 40+ configs | NULL | none survives multiplicity |
| Pre-release drift (jobs, inflation, GDP) | 18+ configs | NULL | efficiently priced at the release |
| Divergence-reversal ("SMT") direction | 7 regimes | NULL | doesn't beat a random-divergence placebo |
| News-reaction momentum | — | NULL | the one strong window was a single 2022–24 inflation regime; reverses out of it |
| Weekend crypto → Monday equity direction | 250+ weekends | NULL | already priced at the Sunday futures reopen |
| Scheduled-Fed-decision positioning drift | — | SURVIVOR | a real, regime-amplified directional edge |
Names, authors and scripts intentionally absent. This is asset-class and failure-mode aggregate only.
The honest headline
Across 400+ directional intermarket configurations, essentially nothing survives honest correction. Exactly what the decay literature predicts. Most families didn't just fail after costs — they had no edge against a placebo before a cent of cost went in. The SMT divergence trade is the cleanest example: it does not beat a randomly-placed divergence. If shuffling your signal works as well as your signal, the signal was never there.
The machine-learning result is the one worth sitting with. Point a nonlinear model at 180+ pairs and it will find leads. It found several. Every one of them was a walk-forward mirage — strong in one window, gone the next month. That is what overfitting looks like when it wears a neural network. The grid didn't discover structure. It memorized noise and handed it back with confidence.
The one that lived
One family survived: a positioning drift around scheduled Fed decisions. Not a cross-asset lead — a calendar-anchored, regime-amplified premium that lines up with the published pre-FOMC drift literature. It has its own page, with the effect, the regimes where it shows up, and the honest caveats. Read the pre-Fed drift finding →
Notice what it is not. It is not "bonds told stocks where to go." The survivor is scheduled and mechanical, closer to a pre-FOMC drift than to any lead-lag story. The lead-lag stories — the whole reason people run intermarket panels — are precisely the families that died.
Then what is intermarket data good for?
The second moment. Cross-asset data carries real information about volatility and correlation — how much things move together and how much they'll move — not about direction. Volatility is forecastable. Direction, on this evidence, mostly isn't. That split is the useful takeaway, and it has its own page. Volatility is forecastable. Direction isn't. →
Why believe a null
Publishing a null is only worth anything if the test could have found something. Ours could. The correction was Benjamini–Yekutieli FDR plus a Deflated Sharpe ratio, with N set to every configuration we ran — the honest denominator, not the survivor count. A gap-embargo killed 1-second microstructure artifacts. GARCH-devolatilisation separated a real edge from plain volatility spillover, so a signal couldn't sneak through just because both assets got loud at once. An anchored walk-forward stayed held in reserve.
- A "null" means not multiple-testing-robust on this 2012–2026 data and window — not a proof of impossibility.
- Counts are evergreen floors, never exact drifting totals.
- Costs are modelled where they bite; we report break-even cost, not an assumed net Sharpe.
How we test
Same pipeline we run on everything. Strategies are ported to Python and run against real costs — spreads and commissions modelled from tick data, not a flat guess. Futures come from 13 years of CME via continuous contracts built without survivorship gaps; FX from tick-level bid/ask; crypto as spot and perps. A fast model does the bulk porting, the strongest model then tries to break every apparent winner, hunting look-ahead and impossible fills. Code is hashed, so a signal re-published under three names gets tested once. It is the same process that rejects roughly 78% of the strategies we test — and here, 400-plus intermarket directional configs down to one.
Research and education, not financial advice. No signals, no return promises. Independent, and not affiliated with TradingView.
Want the named verdicts?
This page gives you the aggregate: which families of intermarket signal die, and how. What it doesn't give you is the roster — every strategy and indicator we audited, named, with its verdict and the exact reason it lived or died. That is The No List.
Get The No List → or join the Discord first →FAQ
Does intermarket analysis actually work?
For predicting direction, almost never in our testing. We ran 400+ cross-asset directional configurations through multiple-testing correction; one survived. Intermarket data does carry real information about volatility and correlation — just not about which way price goes next.
Doesn't one market lead another — bonds lead stocks, the dollar leads commodities?
They co-move, which is real. A tradeable lead is different, and the lead-lag families were exactly the ones that failed: 260+ pairs, 0 survivors. Most "leads" are simultaneous co-movement, or an artifact of comparing misaligned cash closes instead of synchronous bars.
Can machine learning find intermarket edges a human misses?
It finds them and they don't replicate. Across 180+ pairs the nonlinear grid surfaced "hidden" leads that were single-month artifacts, gone the next window. That is overfitting wearing a model, not a discovered edge.
So what did survive?
One family: a positioning drift around scheduled Fed decisions — calendar-anchored and regime-amplified, matching the published pre-FOMC drift literature. It is not a cross-asset lead. It has its own page with the effect and the caveats.
Does a "null" mean intermarket signals are impossible?
No. A null means not multiple-testing-robust on 2012–2026 data and this window — not proof of impossibility. We correct with Benjamini–Yekutieli FDR and a Deflated Sharpe, counting every config we ran as the denominator.