The Hindenburg Omen Manual

Research note. Every figure computed at render from our daily breadth dataset (from December 27, 2016) — the same file behind the live Hindenburg Omen tracker. Methods in How we checked it; editorial standards in How we use AI.

What is the Hindenburg Omen?

A breadth checklist from technician Jim Miekka, popularized in the 1990s and named — by a colleague of his, for maximum menace — after the 1937 airship disaster. Its idea is simple under the branding: a healthy rally lifts most boats, so a market making many new 52-week highs and many new 52-week lows at the same time, while the index itself is still rising, is internally split in a way that has shown up near some serious tops. It measures disagreement, not doom. As of September 18, 2026's close the conditions are not met (the last signal day was August 17, 2026), with 1.1% of issues at new highs and 5.6% at new lows.

Is it active right now?

As of the September 18, 2026 close: no. New 52-week highs sit at 1.1% of issues and new lows at 5.6%, against the 2.2% trigger both must clear on the same day. The most recent signal day was August 17, 2026 — 23 sessions ago. 2026 has produced 19 signal days so far. The tracker updates this after every close; the cluster table below puts whatever today says in context.

What triggers a signal? The recipe

Four conditions, all on the same day: new 52-week highs and new lows each at least 2.2% of issues traded; the index above its 50-day average (the split has to happen inside an uptrend — that's the whole point); the McClellan Oscillator negative (breadth momentum already deteriorating); and new highs no more than twice new lows. We compute it across our own universe of roughly 4,750 US common stocks rather than the classic NYSE-composite lists, and single days that qualify get grouped into clusters when they land within 30 sessions of one another. The tracker page documents every departure from Miekka's original spec in a comparison table. One consequence worth knowing: counts differ between sites because universes differ. Our breadth-momentum input now uses dated stock classifications and available archived prices; the high/low component retains its existing filter. Both have incomplete historical coverage.

How accurate is it? The honest record

Since 2016, our variant has produced 119 signal days, grouping into 25 clusters. In the 63 sessions after a signal day, SPY averaged +0.1% against a +3.4% baseline. And yet 65% of those windows still finished positive. Both facts matter. The difference from the baseline measures what followed the signal in this sample; individual outcomes varied. Anyone selling everything on a single trigger has been run over repeatedly; anyone ignoring a cluster entirely has met a few famous accidents.

Famous fires, famous fizzles

These 5 selected dates still qualify under the current calculation. SPY's next-quarter return is computed from each listed signal day.

Selected examples cannot establish the signal's predictive value. The complete table below includes all 25 clusters, and the tracker compares confirmed clusters against the unconditional forward-return baseline.

Every cluster in the covered history

The complete record, newest first — each cluster's first signal day, how many qualifying days it contained, and SPY's return over the following quarter. Every covered cluster is included.

Where this gauge will mislead you

• The name oversells it. "Omen" promises crashes; the data delivers below-average forward returns with a 65% positive rate. It is a caution flag with excellent marketing.
• Every implementation disagrees. Thresholds, trend filters and universes vary by source, so "the Omen fired" means different things on different sites. Ours is documented line by line on the tracker precisely because of this.
• It repeats what other gauges already say. When we audited our Swing Score, the Omen's ingredients were already counted by its breadth checks — so we demoted it to an unscored alert rather than let it vote twice. Treat it the same way in your own process: confirmation, never headline.
• History here starts December 27, 2016. Claims about 1987 or 2008 come from other datasets and other recipes; our record covers 25 clusters across 10 calendar years and can't validate the older folklore either way.

How we checked it

Signal days come straight from our daily breadth dataset — the same computation the tracker charts — with clusters formed by a stated rule (a new cluster requires a 30-session gap between qualifying days). Forward returns are SPY averages from each signal day; windows overlap heavily across days inside a cluster, so effective sample sizes are far smaller than the day counts. The named-cluster table computes each next-quarter return at render, so revisions to the underlying data flow through automatically.

Frequently asked questions