Breadth · Signals
The Hindenburg Omen Examined
The condition it looks for is real and worth watching: a market where many issues make new highs and many make new lows at the same time. The claim built on top of it rests on a sample small enough to count on your fingers.
The conditions
The signal is a conjunction. On a single session, all of the following are required to hold, the exact thresholds vary between accounts, which is itself worth noting:
- Both extremes are crowded. New 52-week highs and new 52-week lows each exceed roughly 2.2 per cent of issues traded. This is the substantive condition and the interesting one.
- The market has been rising. The index sits above its level of about fifty sessions earlier, so the signal describes division inside an advance rather than inside a decline.
- Momentum is negative. The McClellan oscillator is below zero.
- The highs do not dominate. New highs are no more than roughly twice new lows.
And a fifth requirement sits outside the session: occurrences are conventionally expected to cluster, with a second signal inside about thirty sessions, before the condition is taken to mean anything.
What the conjunction costs in sample size
Each condition on its own is uncommon. Requiring all four at once, and then requiring the result to repeat within a month, produces something genuinely rare, and rarity is exactly what makes the published claims impossible to support.
| Condition | Assumed probability |
|---|---|
| New highs AND new lows both above 2.2 % of issues | 5 % |
| Index above its level of 50 days ago | 62 % |
| McClellan oscillator negative | 45 % |
| New highs no more than twice new lows | 35 % |
Those figures are estimates and they are deliberately generous: each condition is treated as more common than the literature’s own descriptions imply, which makes the signal look more frequent than it is. The argument is robust to them being wrong by a factor of two in either direction, because the conclusion is about an order of magnitude rather than a number.
And the count is more flattering than the sample. Clustered signals are not independent observations: the conditions that produce one tend to persist, so several clusters can belong to a single market period, the run-up to one decline contributes a handful of them, and they all resolve together. What a test can learn from is the number of distinct episodes, which is far smaller than the number of clusters and is what keeps the confidence interval wide however the arithmetic above is adjusted. A signal that fires four times in one year and then not for a decade has given you one observation, not four.
The base rate, which is the decisive omission
Suppose a decline followed a majority of clustered occurrences. That sounds impressive until the comparison is made explicit: markets fall reasonably often, so any signal is followed by a decline a substantial share of the time.
The question a test has to answer is whether the probability of a decline after the signal is meaningfully higher than the probability over the same horizon without it. With a few dozen episodes, the confidence interval around that difference is wide enough to include zero, which is not a claim that the signal is worthless, but a statement that the available evidence cannot distinguish it from the base rate. Those are different conclusions and they are routinely conflated.
The condition has also drifted
A second problem sits underneath the statistical one, and it is the same one that runs through every breadth page in this reference: what the underlying data contains has changed.
The thresholds are expressed as percentages of issues traded, so they survive the growth in the number of listings, that much was designed well. What they do not survive is the change in composition. The share of exchange listings that are closed-end funds, preferred shares and exchange-traded products has grown substantially, and those instruments reach new highs and new lows together on interest-rate moves rather than on anything happening to equities.
So the central condition, many highs and many lows at once, can now be produced by a bond-market move in a way it could not when the signal was described. A common-stock-only breadth series removes that, and almost no published version of the indicator uses one.
What to keep
The observation, and none of the machinery. A market in which a large number of issues are making new highs while a large number make new lows is genuinely divided, and that state is invisible in a net figure or a difference chart, which is why watching the two counts separately is worth the trouble.
Treat it as a description of the market’s internal state on a dated session, and check later what followed, keeping the occasions when nothing did. That is a cheap habit with real information in it. What it is not is a signal with a known probability attached, and no amount of naming makes a handful of episodes into evidence.
Frequently asked questions
What is the Hindenburg Omen?
A breadth signal that fires when several conditions hold on the same session, the central one being that new 52-week highs and new 52-week lows are both unusually numerous; each above roughly 2.2 per cent of issues traded. The usual formulation adds that the index must be above its level of fifty sessions earlier, that the McClellan oscillator must be negative, and that new highs must not exceed twice new lows. Occurrences are conventionally required to cluster before the signal is taken seriously.
What is the reasoning behind it?
That a market where a large number of issues are reaching new highs while a large number are reaching new lows is internally divided, and that division is a sign of instability. The observation is genuinely interesting and it is one this reference agrees with: the both-counts-high condition is the state that a difference-based new-high/new-low chart averages away, so it is worth looking for. Whether it forecasts anything is a separate question.
Why is the evidence weak?
Sample size. The conjunction of four uncommon conditions is rare by construction, and the version people quote also requires a cluster of occurrences, so even a century of daily data contains few episodes. The figure on this page works the arithmetic through: with generous probabilities for each condition, a ninety-year history holds on the order of a hundred individual signals and a few dozen clusters. That sounds like enough until you notice that clusters are not independent — several of them belong to the same market period — so the number of distinct episodes to learn from is much smaller than the count suggests.
Has it ever fired without a decline following?
Repeatedly, which is what its own advocates report. The usual summary is that a significant decline followed in a majority of clustered cases, with a substantial minority producing nothing. Both halves of that sentence come from the same small sample, and the honest reading of "a majority of a dozen" is that the number is compatible with a wide range of true probabilities, including one no better than the base rate of declines generally.
Is the base rate accounted for?
Rarely, and it is the decisive omission. Markets fall reasonably often, so any signal will be followed by a decline a good share of the time. What matters is whether the probability after the signal is meaningfully higher than the probability without it, over the same horizon, and a comparison against that base rate is what almost no published account of this indicator provides.
Do the conditions still mean the same thing?
Not exactly, and this is the quiet problem underneath the statistical one. The thresholds are percentages of issues traded, so they survive changes in the number of listed issues, but not changes in composition. The share of listings that are funds, preferred shares and exchange-traded products has grown substantially, and those instruments reach new highs and lows together on interest-rate moves. So the both-counts-high condition can now be produced by a bond-market move, which is not what the signal was originally describing.
What is worth taking from it?
The condition, not the omen. Watching whether new highs and new lows are both elevated is a cheap and genuinely informative habit: it identifies a divided market, which a single net figure hides completely. Treat it as a description of the market’s internal state, note the date, and check afterwards what followed, including the many occasions when the answer is nothing.
How would it be tested properly?
Define the conditions in advance, on a common-stock-only breadth series, across every market with a long enough history rather than one. Compare the drawdown distribution after signals against the unconditional distribution over the same horizons. Report the number of episodes, which will be small, and the confidence interval around any difference, which will be wide. That is a publishable result whichever way it comes out, and it is a different exercise from listing the occasions that worked.