Which popular stock signals don't predict returns?
By Mehmet Saruhan, Founder·Opulence Alpha Research · Published · Data through
Macro sensitivities, statistical factor loadings, wavelet cycles, crowding scores, 13D/13G filings, anomaly intensity and chart patterns combined with context. Of 69 such features measured on US stocks as far back as 1995, 0 cleared the bar for a measured tendency and 14 showed a faint tilt. Sorted into fifths on 4 of them, each fifth beat the median stock over the next month 49 to 50 times in 100.
Keywords: stock signals that don't work, do macro factors predict stock returns, interest rate sensitivity stocks, 13D filing stock returns, 13G blockholders, crowding score stocks·JEL classification: G11, G12, G14, C12, C58·Concepts: Multiple testing, Point-in-time data, Backtest overfitting, Deflated Sharpe ratio
- Universe
- The same for every study: 1,767 US common stocks in the 11 GICS sectors, 422 of them since delisted; S&P 500 members since 1996 plus large and mid-sized companies outside the index
- Period
- Every Wednesday from 4 Jan 1995 to 19 Aug 2026: 1,634 Wednesdays across 7,959 trading sessions
- Sample
- n = 1,763 stocks of the 1,767-stock universe; 1,889,758 stock-weeks, a median of 1,148 stocks per Wednesday
- Outcome
- Return over the next 1, 5, 21 and 63 trading sessions against the same day's median stock
- Inference
- t on non-overlapping dates; proven only when |t| ≥ 3 and the same sign in at least 60% of years
1Introduction
Every year brings new stock signals: sensitivities to interest rates and commodities, statistical factors, cycle decompositions, crowding scores, activist filings, anomaly counts. Harvey, Liu and Zhu showed that with so many signals tested, the usual bar for significance lets far too many through by luck, and McLean and Pontiff found that return predictors weaken after they are published. This study measures 69 such features on one universe, from 1995 or from the year each one's data begins, with a stricter bar, and reports how many cleared it.
2Data and method
2.1Sample design
Population and frame. The population is US common stocks listed on the NYSE and Nasdaq; funds, ETFs, trusts, preferred shares, warrants and units are excluded. The sampling frame is a fixed universe of 1,767 companies, drawn once when the platform was built and not re-sampled since, in two strata. Stratum 1 is a census of the S&P 500: every company in the index at any time since 1996 whose price history could be recovered, 1,027 companies of which 413 have since delisted; it holds 76% of the index's members in 1996 and at least 96% in every year from 2010. Stratum 2 is 740 large and mid-sized companies outside the index, selected in proportion to the market's sector weights from the stocks that passed a minimum share price of $15 and a minimum average daily trading value of $25 million; 9 of them have since delisted.
Sample. The unit of observation is a stock-week: one stock on one Wednesday. A stock enters a Wednesday's cross-section when it has a valid close that day, a value of the signal and a measured outcome; bars flagged as bad data and returns that cross a change of issuer are left out. This study's sample is n = 1,763 stocks of the 1,767: 1,889,758 stock-weeks on 1,634 Wednesdays from 4 Jan 1995 to 19 Aug 2026, a median of 1,148 stocks per Wednesday (range 820 to 1,371). 331,954 stock-weeks (17.6%) come from the 422 companies that have since delisted. Every stock and every Wednesday carries equal weight.
Representativeness. Table 1 gives the sample by GICS sector beside the S&P Composite 1500: 7.9% of companies would have to change sector for the two to match exactly. By latest market value, 55% of the listed companies with a recorded market value are large (at least $10bn), 38% mid ($2–10bn) and 6% small. Because Stratum 2 was chosen from companies listed at construction, its history carries survivorship bias. Section 4 repeats the lead result among the stocks that were S&P 500 members on each date, the part of the sample largely free of that bias.
Companies in this study's sample, of them those since delisted, the size of the listed companies with a recorded market value, the sample's share of stock-weeks and the sector's share of the S&P Composite 1500.
| Sector | Sample | Delisted | Large | Mid | Small | Stock-weeks | S&P 1500 |
|---|---|---|---|---|---|---|---|
| Information Technology | 282 | 67 | 122 | 76 | 18 | 13.9% | 12.7% |
| Financials | 256 | 60 | 114 | 74 | 8 | 14.6% | 17.2% |
| Industrials | 255 | 53 | 126 | 67 | 11 | 15.9% | 17.5% |
| Health Care | 237 | 47 | 92 | 81 | 16 | 12.4% | 10.9% |
| Consumer Discretionary | 209 | 39 | 69 | 87 | 12 | 12.3% | 12.9% |
| Energy | 105 | 30 | 39 | 31 | 5 | 5.9% | 4.7% |
| Consumer Staples | 101 | 38 | 40 | 17 | 5 | 5.9% | 4.9% |
| Materials | 94 | 34 | 35 | 24 | 2 | 5.6% | 5.1% |
| Real Estate | 84 | 10 | 40 | 32 | 2 | 5.7% | 6.9% |
| Communication Services | 80 | 31 | 28 | 13 | 5 | 3.9% | 3.3% |
| Utilities | 60 | 13 | 35 | 12 | 0 | 4% | 4% |
| All sectors | 1,763 | 422 | 740 | 514 | 84 | 100% | 100% |
Sample: n = 1,763 of the 1,767 companies. Size by latest market value for the 1,338 listed companies with one recorded: large ≥ $10bn, mid $2–10bn, small < $2bn. S&P 1500 shares count the constituents of the S&P 500, MidCap 400 and SmallCap 600 (1,506 companies, lists read 25 Sep 2026). Sectors are each company's current GICS sector.
2.2Signals
| Signal | Definition | Since | Wednesdays |
|---|---|---|---|
| Sensitivity to the 10-year yield | How a stock's returns have moved with changes in the 10-year Treasury yield | 4 Jan 1995 | 1,634 |
| Concentration in one statistical factor | How much of a stock's movement loads on a single one of the market's strongest statistical factors | 4 Jan 1995 | 1,634 |
| Blockholder concentration (13D/13G) | How concentrated the stakes of large holders filing 13D or 13G are | 14 Feb 2001 | 1,317 |
| Crowding composite | A blend of abnormal volume, turnover and co-movement with the market | 4 Jan 1995 | 1,634 |
2.3Measurement
Outcome. The return from the close on the Wednesday to the close 1, 5, 21 and 63 trading sessions later, on closes adjusted for splits and dividends, compared with the same day's median stock; half of all stocks beat the median by construction, so chance is 50 in 100 on every date. No delisting return is added.
Each Wednesday the stocks are ranked on the signal. The study reports how often each fifth of that ranking beat the median stock, and the rank correlation (IC) between the signal and the return that followed.
The t-statistic uses non-overlapping dates only. A result is called proven when |t| is at least 3, it held in at least 60% of years and it is large enough to matter; with 276 tests across the studies, a looser bar would pass dozens by luck.
Full data and methods3Results
What we measured, family by family
Each family was measured the way every study here is: every Wednesday from 1995, or from the year its data begins (the table gives each family's start), stocks were ranked on the signal and followed for a session, a week, a month and three months against the same day's median stock. Of 69 features, 0 carried a measured tendency, 14 a faint tilt that is not reliable on its own, and 55 no measurable edge.
| Family | What it describes | Measured | Proven | Faint | No edge |
|---|---|---|---|---|---|
| Macro sensitivitiessince 1995 · some features from 2004 | How a stock has moved with oil, copper, gold, the 10-year Treasury yield, credit spreads and inflation expectations | Measured24 | Proven0 | Faint2 | No edge22 |
| Statistical factor loadingssince 1995 | Exposure to the strongest statistical factors in the market's returns, and how concentrated it is in one of them | Measured6 | Proven0 | Faint1 | No edge5 |
| Wavelet cyclessince 1995 | How a stock's price movement splits between short and long cycles | Measured13 | Proven0 | Faint1 | No edge12 |
| Crowding scoresince 1995 | A blend of abnormal volume, turnover and co-movement with the market, with its one-year percentile, its trend and its gap to the sector | Measured4 | Proven0 | Faint2 | No edge2 |
| Ownership filings (13D/13G)since 1995 · some features from 2001 | Large holders and activists: new filings, stakes, concentration and accumulation | Measured8 | Proven0 | Faint7 | No edge1 |
| Anomaly intensitysince 1995 | How many unusual price or volume moves a stock had recently, and how extreme they were | Measured4 | Proven0 | Faint1 | No edge3 |
| Chart patterns with contextsince 1995 · some features from 2021 | A pattern combined with short interest, analyst actions or crowding | Measured10 | Proven0 | Faint0 | No edge10 |
Sorted into fifths, the bars stay flat
Every Wednesday the universe is split into five equal groups on the signal, from its lowest fifth to its highest, and each group is followed for a month. A signal that forecast returns would tilt the bars, one end above 50 in 100 and the other below. Here every fifth beat the same day's median stock 49 to 50 times in 100. Even the relations that do measure are small: a tilt across hundreds of stocks, a nudge for any one.
How often each fifth beat the median stock over the next month · since 1995
How often each fifth beat the median stock over the next month · since 1995
How often each fifth beat the median stock over the next month · since 2001
How often each fifth beat the median stock over the next month · since 1995
Why a faint tilt is not enough
Test enough signals and some look good by luck alone: at a t-statistic of 2, about one test in twenty passes when nothing is there. This page covers 276 tests, each feature at four horizons. 18 cleared that loose bar, against about 14 that luck alone would pass. So a relation counts here only at a t-statistic of 3 or more, with the same sign in most years and a size that matters. 3 tests reached 3, and 3 of those were too small to matter.
Context still describes the stock
None of this makes these readings useless. A stock's sensitivity to oil or to the 10-year yield says what it is exposed to; a 13D filing says an activist holds a large stake; a crowding score says how heavily a stock is being traded and how closely it moves with the market. Those describe a position's risks and who else is in it. They did not tell which way the stock went next against the median stock, so Opulence Alpha shows them as context, never as a signal.
Chart patterns belong on this list
On their own, 0 of 15 chart pattern types showed a measured edge over the next month. Combining a pattern with short interest (data from 2009), analyst actions (from 2012) or crowding did not change that: 0 of 10 combinations carried a measured tendency.
Do chart patterns predict stock returns? The full studyWhat did clear the bar
A few relations did: small for any one stock, steady across the market. The strongest is short-term reversal. The crowding composite is only faint as a blend, yet one of its parts, unusually heavy volume measured on its own, is among the small, steady ones.
Short-term reversal: the strongest relation we measuredVolume and liquidity: what heavy trading did next4Robustness
Table 3 re-measures the crowding composite over the next month: 0.007 (t 0.7) across all stocks and 0.011 (t 1.0) among S&P 500 members on the date. Its sign held in 9 of 11 sectors and 3 of 3 decades, but nowhere strongly enough to separate from chance: a steady sign on a correlation this small is not an edge.
Rank IC and its t-statistic on non-overlapping dates.
| Sample | IC | t | Wednesdays |
|---|---|---|---|
| Crowding composite, next month | |||
| All stocks, whole period | 0.0069 | 0.7 | 1,630 |
| By decade: 1995–2004 | 0.0057 | 0.3 | 516 |
| By decade: 2005–2014 | 0.0056 | -0.1 | 518 |
| By decade: 2015– | 0.0091 | 0.9 | 596 |
| S&P 500 members on the date | 0.0111 | 1.0 | 1,578 |
| Sectors with the overall sign | 9 / 11 |
5Limitations
- The universe and period are the same as in every study, but families start at different dates: 13D/13G blockholder data from 2001 and some macro series only from 2004, so their verdicts rest on fewer years.
- Finding no edge holds for this universe, these definitions and horizons up to three months; a signal with no measurable edge here can still describe a stock's risks, and longer horizons are not tested.
- Before costs. Averages exclude trading costs, taxes and market impact.
- Same-close timing. Returns start at the close the signal is computed from; a real trade would start later.
- Survivors among smaller companies. The non-index names were chosen from companies listed when the universe was built; no delisting returns are added.
6Conclusion
0 of 69 features in these families cleared the bar for a measured tendency and 14 showed only a faint tilt; sorted into fifths, every fifth of the 4 charted signals beat the median stock over the next month 49 to 50 times in 100, a coin flip. These readings still describe a stock's exposures and who holds it; they did not tell which way it went next against other stocks. Across 276 tests, 18 cleared a t-statistic of 2 against about 14 that luck alone would pass, which is why a relation counts here only at a t-statistic of at least 3 that holds across years.
References
- Harvey, C. R., Liu, Y. and Zhu, H. (2016). …and the Cross-Section of Expected Returns. Review of Financial Studies 29(1).
- McLean, R. D. and Pontiff, J. (2016). Does Academic Research Destroy Stock Return Predictability? Journal of Finance 71(1).
- Benjamini, Y. and Hochberg, Y. (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society, Series B 57(1).
- Grinold, R. C. and Kahn, R. N. (2000). Active Portfolio Management, 2nd ed. McGraw-Hill.
- Harvey, C. R., Liu, Y. and Zhu, H. (2016). … and the Cross-Section of Expected Returns. Review of Financial Studies 29(1).
- Shumway, T. (1997). The Delisting Bias in CRSP Data. Journal of Finance 52(1).
Appendix A. Questions readers ask
- Do macro sensitivities like interest-rate beta predict stock returns?
- Not reliably. Of 24 measures of how US stocks moved with oil, copper, gold, the 10-year Treasury yield, credit spreads and inflation expectations, measured from 1995 (some only from 2004), 0 carried a measured tendency and 2 showed a faint tilt, not reliable on its own. Ranked by 10-year yield sensitivity, the lowest fifth beat the median stock over the next month 50 times in 100 and the highest fifth 50.
- Do 13D and 13G filings predict stock returns?
- Only as a faint tilt, not a reliable edge. Of 8 features built from 13D and 13G ownership filings (activist filings, blockholder counts, stakes, concentration and accumulation), measured from 1995 (some only from 2001), 0 carried a measured tendency and 7 showed a faint tilt, not reliable on its own. Ranked by blockholder concentration since 2001, the lowest fifth beat the median stock over the next month 50 times in 100 and the highest fifth 50.
- Does a crowding score predict stock returns?
- Only faintly. Measured on US stocks since 1995, the crowding composite, its one-year percentile, its trend and its gap to the sector gave 0 measured tendencies out of 4; 2 showed a faint tilt, not reliable on its own. Ranked by the composite, the lowest fifth beat the median stock over the next month 49 times in 100 and the highest fifth 50.
- Do wavelet cycles, statistical factors or anomaly scores predict returns?
- Not reliably. Across 13 wavelet-cycle features, 6 statistical-factor features and 4 anomaly-intensity features measured on US stocks since 1995, 0 carried a measured tendency; 3 showed a faint tilt, not reliable on its own.
- Why do so many stock market signals fail when tested?
- Two reasons show up in this data: luck and size. Across the 276 tests on this page, 18 cleared a t-statistic of 2, against about 14 that luck alone would pass; 3 reached 3, and 3 of those were too small to matter. Even a real relation is a tilt across hundreds of stocks and a nudge for any one of them.
Data availability and citation
Every figure in this paper is quoted from one frozen snapshot, published as JSON with the sample description, the robustness results and the test counts. The same snapshot feeds the research console, so the two cannot disagree. The paper and the snapshot are licensed under CC BY 4.0: share and adapt them with credit to Opulence Alpha and a link to this page.
Mehmet Saruhan (2026). Which popular stock signals don't predict returns? Opulence Alpha Research, Opulence Alpha Studies, 25 Sep 2026. https://opulencealpha.ai/studies/signals-without-an-edge
Research, not advice. A measured tendency across hundreds of stocks is a nudge for any one of them, never a forecast of its price.