The McClellan Oscillator Manual

Research note. Every figure computed at render from our daily breadth dataset (from 2017-03-28, reconstructed common-share coverage) — the same file behind the A/D line and the Tape Score's breadth dimension. Methods in How we checked it; editorial standards in How we use AI.

What is the McClellan Oscillator?

Take each session's advancing stocks minus declining stocks — the same raw material as the A/D line — but instead of summing it forever, run a fast and a slow exponential average and subtract: the 19-day EMA minus the 39-day EMA. The result oscillates around zero and answers a sharper question than the line does: is breadth accelerating or decaying right now? Sherman and Marian McClellan built it in 1969; it survives because it compresses the market's participation trend into one number that turns quickly at panics. As of the September 18, 2026 close ours reads −34.

Where it stands today

−34 — the 16th percentile of every session in the covered history, on a day when net advances came in at -1,149. For calibration: the oscillator has spent 51% of its life above zero, its deepest print was −129 (March 12, 2020) and its highest +116 (April 29, 2020). The same construction is scored daily inside the Tape Score's breadth dimension, and a negative oscillator is one of the four conditions of the Hindenburg Omen.

How it is built — and the two decisions that matter

Two construction choices separate one McClellan from another, and both change the numbers. Ratio adjustment: raw net advances grow with the number of listed issues, so we use the scaled form — 1000 × (advancers − decliners) ÷ (advancers + decliners) — which accounts for changing population size. Universe: the classic oscillator counts NYSE-listed issues, bond funds and preferreds included; ours counts eligible US-listed common shares, which removes rate-sensitive noise but shifts the scale. One honesty detail most sources skip: an EMA has to warm up from its seed, so we exclude the first 100 sessions of the record from every statistic and chart on this page — without that cut, the "deepest readings ever" would all be initialization artifacts.

How to read it — the honest edge test

We tested the folklore against our own record: every session in the covered history, bucketed by the oscillator's closing value, with SPY's return over the following month. The all-days baseline is +1.2%.

Two findings, both against the folklore. Deep oversold was the best bucket — panic breadth mean-reverted — but the tilt is half a percentage point, too small to trade. And deep overbought was not bearish: +60 readings mostly mark breadth thrusts, the beginnings of advances, which is the same lesson our Zweig Breadth Thrust study teaches. The oscillator's real job is context — telling you whether the tape's engine is accelerating or sputtering — not generating entries.

The deepest readings on record

The eight deepest prints (one per episode), computed at render. Read the list: it is a catalog of the era's panic days — and one fizzle we keep in deliberately.

The COVID low's neighborhood, the 2022 September flush and Christmas Eve 2018 illustrate why traders compare deep breadth washouts with subsequent buying panics. And then April 2012, which looked identical on this gauge and went nowhere for months. That row is why the edge test above reports a tilt and not a rule.

The Summation Index

The McClellan Summation Index — the running cumulative total of the oscillator — is the usual companion, and we publish it, together with NYSE-listed and Nasdaq-listed variants that stand in for the NYSI and NASI series other charting packages carry.

We previously refused to, and the reason we gave was wrong. The argument was that our universe's daily net advances carry a structural negative drift — delistings concentrate among decliners — so a running sum trends unboundedly and its level stops meaning anything. That is entirely true of the A/D line, and false here, for a reason the old text stated in its own next sentence: the drift washes out of a difference of averages. The Summation Index accumulates the oscillator, where that cancellation has already happened, not net advances. Measured across our record the oscillator's mean is −0.45 against a standard deviation of 124 — 0.36% of one standard deviation — and the running total oscillates between −688 and +1,785 rather than running away.

What survives from the old caution is narrower and still applies: the index accumulates from an arbitrary start date, so its level carries no absolute meaning. Read direction, zero crossings and divergence against price. And the venue variants are not $NYSI and $NASI — those count every listed issue, roughly half of which are ETFs, preferreds, closed-end funds and warrants rather than operating companies, while ours count common stocks only.

Where this gauge will mislead you

• Levels don't transfer between sources. NYSE-composite, Nasdaq, and common-stock constructions all print different numbers on the same day. ±60 is stretched here; it may be routine elsewhere. Compare shapes, not values.
• The edge is a tilt. Half a point of monthly outperformance from oversold extremes, with 2012-style failures inside the sample. Position sizing on this alone has no support in our data.
• Overbought is not a sell. The +60 bucket beat the neutral bucket. Treating breadth thrusts as exhaustion is the single most common misreading of this gauge.
• It is two EMAs of one number. Everything the oscillator knows is already in daily net advances; it adds smoothing and memory, not information. When it disagrees with the raw tape, the tape is newer.
• Divergence stories need discipline. Oscillator-vs-price divergences are read into charts freely after the fact. Our systematic divergence work lives on the A/D line, where the rules are stated.

The last 12 sessions

Underlying daily breadth dataset: ad_line.json.

How we checked it

Advancers and decliners come from our daily breadth dataset; the oscillator is EMA19 minus EMA39 of 1000 × (A−D)/(A+D), with the first 100 sessions dropped as EMA warm-up. Percentiles rank the latest value against every included session. The edge-test buckets assign each session by its closing oscillator value and average SPY's return over the following 21 sessions; windows overlap, so effective sample sizes are smaller than the counts shown. The deepest-readings table keeps one print per episode (minimum 40 sessions apart). Everything recomputes at render.

Frequently asked questions