Across thirty years of S&P 500 data, the share of constituents above their 40-week moving average did not order the average return over the following twelve months. It ordered the spread of those returns consistently, and the risk actually experienced during them only in the earlier half of the sample.
Research · Data through 2026-08-31
Market breadth counts how much of an index is participating. The form used here is the share of constituents above their 40-week moving average — approximately 200 trading days. Forty weekly observations, not an average of 200 daily ones; the distinction matters for how the measure behaves and is carried through the study.
It is cited most often as a warning: a rising index on narrowing breadth is described as fragile, and a low reading as a reason to reduce exposure.
We tested that on 360 month-end observations from 1996-09-30 to 2026-08-31, of which 347 have a complete twelve-month forward window. Membership is point-in-time, from Bloomberg quarterly snapshots. Coverage is 99.6% of constituents at the median month and never below 99.4%.
The series reaches its lows where it must: 3.2% in 2009-02, 6.1% in 2020-03, 10.4% in 2002-07 and 12.0% in 2022-09.
360 month-ends, 1996-09-30 to 2026-08-31. The reading at 2026-08-31 is 67.5%, the 54th percentile of its own history, +11.4 points over three months.
Sorting every month into quintiles by its starting breadth:
| Starting breadth | Mean reading | Forward 3m mean | Forward 12m mean | Forward 12m median | Positive 12m |
|---|---|---|---|---|---|
| Q1 lowest | 29.4% | 3.47% | 11.68% | 16.0% | 71.4% |
| Q2 | 54.1% | 2.28% | 7.55% | 12.81% | 71.0% |
| Q3 | 65.5% | 1.69% | 12.59% | 14.62% | 81.4% |
| Q4 | 75.4% | 3.23% | 11.37% | 13.2% | 88.2% |
| Q5 highest | 86.8% | 2.86% | 12.02% | 13.37% | 87.1% |
| All months | 2.7% | 11.05% | 13.74% | 79.8% |
Starting breadth did not order subsequent average returns. The lowest quintile returned 11.68% over the following year and the highest 12.02%, but the second quintile is the weakest of the five at 7.55% and the third the strongest at 12.59%. There is no monotonic relationship at either horizon, and the same is true of the median.
Mean S&P 500 total return over the following 3 and 12 months, by the quintile of the breadth reading at the start of the month. The 12-month means span 5.04 percentage points across the five buckets and do not order with breadth. Lowest quintile 11.68% over 12 months, highest 12.02%, weakest the second at 7.55%.
Sorting instead by the three-month change in breadth — the directional reading — gives the same answer. Months where breadth had fallen most returned 12.31% over the following year; months where it had risen most returned 13.27%; the middle quintile returned 9.03%. Neither the level nor its direction ordered average returns.
The probability of a positive twelve months did rise with breadth, from 71.4% in the lowest quintile to 87.1% in the highest — though not strictly, since the two lowest quintiles are equal and the highest is marginally below the fourth. That is a statement about the frequency of positive outcomes, not about their average size, and the two point in different directions here.
They can point in different directions without contradiction, and the quartiles show how. In the lowest quintile the middle half of outcomes ran from -5.99% to 29.52%; in the highest, from 6.31% to 18.06%. Low breadth was followed by a positive year less often, but when it was positive the upper quartile was eleven points higher. The two effects offset in the mean and do not offset in the spread — which is why the spread is where the signal turns out to be.
The same observations, measuring the spread of forward twelve-month returns across the starting months in each quintile:
| Starting breadth | Forward 12m standard deviation | Q25 | Q75 | Minimum | Maximum |
|---|---|---|---|---|---|
| Q1 lowest | 24.61 | -5.99% | 29.52% | -43.42% | 56.23% |
| Q2 | 18.72 | -7.82% | 20.39% | -36.06% | 41.03% |
| Q3 | 14.76 | 4.22% | 22.23% | -18.18% | 48.21% |
| Q4 | 11.35 | 6.29% | 17.79% | -24.74% | 34.86% |
| Q5 highest | 9.32 | 6.31% | 18.06% | -11.24% | 30.08% |
Every monthly observation from 1996 to 2026, grouped by the breadth quintile it started in. After the lowest-breadth months the outcome fell anywhere between -43.4% and 56.2%; after the highest-breadth months between -11.2% and 30.1%. Overlapping windows — see Robustness.
This ordering is monotonic across all five quintiles. The spread of outcomes following the lowest-breadth months is
2.6 times that following the highest.
A block bootstrap with 12-month blocks and 10,000 resamples, which accounts for the dependence created by overlapping forward windows, puts the Q1-minus-Q5 difference at
13.58 points, 95% interval [5.51, 20.96], with the ratio at 2.479× [1.545, 3.798]. The interval excludes zero.
The spread narrows at every step from the lowest quintile to the highest. A moving-block bootstrap (10,000 resamples, 12-month blocks) puts the Q1-minus-Q5 difference at 13.58 points, 95% interval 5.51 to 20.96.
This is a statement about the cross-sectional spread of endpoint outcomes — how differently twelve-month periods beginning at similar breadth readings turned out. It is not, by itself, a statement about the volatility or drawdown experienced along the way. Those are measured next, and they do not behave identically.
Measured during each forward twelve months, from daily returns:
| Starting breadth | Realized volatility (annualised) | Maximum drawdown | Worst 3m within the window |
|---|---|---|---|
| Q1 lowest | 22.43% | -19.25% | -12.95% |
| Q2 | 17.96% | -17.04% | -11.96% |
| Q3 | 16.22% | -14.02% | -10.32% |
| Q4 | 16.22% | -13.52% | -10.66% |
| Q5 highest | 15.41% | -12.24% | -9.15% |
Both measures come from daily bars inside the forward window rather than from its endpoint. Drawdown deepens at every step as breadth falls, and its bootstrap interval clears zero (0.10 to 11.66). Volatility is higher after low breadth but does not order across the middle quintiles, and its interval contains zero (-0.62 to 8.14).
Maximum drawdown is monotonic across the five quintiles, from
-19.25% to -12.24%. The bootstrap places the difference at 6.59 points, 95% interval [0.1, 11.66] — excluding zero, but only just.
Realized volatility is not monotonic. It is higher after low breadth (22.43% against 15.41%), but the third and fourth quintiles are equal, and the bootstrap interval [-0.62, 8.14]
includes zero. The probability that Q1 exceeds Q5 is 0.94. Directionally consistent; not established.
So the endpoint-dispersion result is stronger than the realized-risk result, and the two should not be reported as one finding.
Expanding-window quintiles. Re-running with boundaries estimated only from breadth observed up to each month — no full-sample knowledge, minimum 120 months of history, n=228 — preserves the dispersion ordering: 26.72, 16.81, 12.36, 10.35, 10.02. Returns remain unordered (11.99% to 11.11%, peaking in the third quintile at 14.68%).
Non-overlapping sample. Using one observation per year — December month-ends only, n=29, five to six per quintile — the dispersion gap survives in direction (26.9 against 9.44) but is no longer monotonic: the third quintile (21.88) sits above the second (18.37). At this sample size that is expected and little should be read into the ordering, but it is reported rather than omitted.
Sample halves, and this is the material caveat. Splitting at 2011-03-31, the dispersion ordering holds in the earlier half (27.43, 19.56, 19.16, 13.12, 8.53) and weakens in the later one (14.16, 12.11, 12.07, 9.15, 10.5).
The same calculation as above, run separately on the first and second 180 months. The ordering is much steeper in the first half (27.4 down to 8.5) than in the second (14.2 down to 10.5), and in the second half the highest quintile is no longer the narrowest.
The drawdown relationship does not survive the split at all: in the earlier half maximum drawdown runs
-23.23%, -22.58%, -14.81%, -13.35%, -12.9%, and in the later half
-12.68%, -12.92%, -13.86%, -13.97%, -11.77% — no relationship between starting breadth and subsequent drawdown in the last fifteen years. The full-sample drawdown result is carried by the first half of the record.
Data source. Every observation with a complete twelve-month forward window ends before the 2025-12-26 transition from Bloomberg to EODHD pricing. The vendor change is therefore outside the twelve-month sample entirely and cannot have influenced these results; it also means the transition remains untested for this horizon, and will only become testable once a year of post-transition data exists.
Breadth stood at 67.5% at 2026-08-31 — the 54th percentile of 360 months, against a full-history mean of 62.2%. It rose 11.4 points over three months and is 1.8 points below where it stood a year ago.
The 54th percentile places it in the third of the five quintiles — the bucket whose mean reading over the full record is 65.5%. On the 347 twelve-month windows measured here, months starting in that bucket returned 12.59% on average with a standard deviation of 14.76, realized volatility of 16.22% and an average deepest drawdown of
-14.02%. Every one of those figures is close to the all-month average of 11.05%, 16.69, 17.65% and
-15.22%, which is what a middle-quintile reading means: it is the part of the range that carries the least information of any part.
Three months ago the reading was 11.4 points lower and a year ago 1.8 points higher. Section 2 tested the three-month change directly and found it did not order average returns either, so the direction of travel does not add to what the level says.
Stated as narrowly as the evidence allows.
Supported. Over 1996-09-30 to 2026-08-31, the spread of twelve-month outcomes following low-breadth months was consistently wider than following high-breadth months. The ordering is monotonic across all five quintiles, survives quintile boundaries estimated without look-ahead, and its bootstrap interval excludes zero. It is the one result in this study that survives every robustness test applied to it.
Supported, but weaker. The deepest drawdown inside the following twelve months also ordered with starting breadth, and its interval clears zero — but only just, and the relationship is absent from the second half of the sample. Realized volatility points the same way and its interval does not clear zero.
Not supported. That breadth ordered the average return, at either horizon, by level or by three-month change. Nothing here says a low-breadth reading is followed by a worse average year; the lowest quintile's mean twelve-month return was 11.68% against the highest quintile's 12.02%.
Not tested. Whether the dispersion relationship is stable enough out-of-sample to act on, what causes it, whether it holds outside the S&P 500 or at other window lengths, and whether any position rule built on it would survive costs. This is a description of thirty years of one index on one measure. It is not a strategy, and nothing here was traded.
Out of reach of the data. The twelve-month sample ends before the 2025-12-26 vendor transition, so the transition cannot have influenced these results and equally cannot be tested against them until a further year of data exists.
This research was prompted by a discrepancy. The live breadth measure on our own surfaces fell from 59.2% to 26.2% across five sessions in which the index rose 0.87%.
The cause was not the market. The underlying price store is designed to hold one bar per week, and the moving averages are defined as a count of bars: forty bars for the long average, ten for the short. Daily bars had begun accumulating in that store, so the same definitions were being computed over roughly two weeks rather than ten. The measure had silently shortened, and a short average tracks price closely enough that the share above it swings hard.
The study above does not use that series. The reconstruction reads adjusted closes from files on disk and contains no reference to the live measures store; the analysis asserts this on every run and refuses to produce a payload if it stops being true.
It is also the point of the piece in miniature. A descriptive statistic that starts moving sharply looks like information. The test is not how much it moves, but whether what follows it differs.
Universe. S&P 500 constituents at each month end from Bloomberg quarterly membership snapshots. Membership is carried forward between snapshots — a month-end that falls between two quarterly observations inherits the membership in force at the preceding one. It is point-in-time to the quarter, not observed monthly. 499 to 517 names per month.
Breadth. Share of constituents whose latest close exceeds the mean of their trailing forty weekly closes, approximately 200 trading days. Percentages are of constituents with a usable observation.
Prices. Chained adjusted closes: Bloomberg total-return history through 2025-12-26, EODHD thereafter.
Returns. Total return of the S&P 500 via SPY from each month end.
Risk. Realized volatility is the annualised standard deviation of daily returns within the forward window (40-week breadth window is unrelated to it). Maximum drawdown is the deepest peak-to-trough decline within the window. Worst 3m is the minimum 63-trading-day return within the window.
Quintiles. Formed on the full sample for the primary tables, and separately on an expanding window for the robustness check.
Bootstrap. Circular moving-block bootstrap, 12-month blocks, 10,000 resamples, quintile boundaries re-estimated within each resample.
Limitations. Twelve-month forward windows overlap, so monthly observations are not independent; the bootstrap exists for that reason and the non-overlapping sample is reported alongside. The series is measured on a total-return basis and is not continuous with the price-return basis our live surfaces display. Quintile boundaries in the primary tables use full-sample information and are descriptive, not a rule that could have been followed in real time. The realized-volatility and drawdown results are not stable across sample halves.
AI assistance. This publication is produced with interactive assistance from large language models, including analysis and drafting. TrendToWealth defines the methodology and editorial requirements and reviews numerical claims against the underlying data and sources before publication.
2026-09 — Initial draft. Sample through 2026-08-31. Methodology version 2.0.
Next expected refresh: semiannual. Conclusions are checked against the previous generated payload on every run; a material change flags EDITORIAL REVIEW REQUIRED rather than silently rewriting the thesis.
Methodology version 2.0. Sample 1996-09-30 to 2026-08-31 (360 month-end observations). Refreshed semiannual; conclusions are compared against the previous run on every refresh.