Tools
RESEARCH

More Buy Signals Didn't Mean Better Stocks: Our Confluence Ranking, Tested

October 5, 2026

We sorted stocks by how many of our own timing techniques were flashing buy at the same time, then asked whether the top of that board actually did better. Five years of daily data said: not measurably.

GetCoattail's Timing board sorts stocks by confluence: how many of our timing techniques are flashing a buy signal at the same time. The intuition is the one behind most signal-stacking tools — one indicator can be noise; several agreeing should mean something. Before letting that sort order stand in for quality, we tested it as we test everything here: rules fixed in advance, our own production code, a benchmark we could lose to.

The short answer: in five years of daily data, the stocks our board ranked highest did not demonstrate an edge. Across 17 non-overlapping 60-trading-day periods, the top 50 beat SPY by +0.61% on average — and the average stock beat it by +0.78%. The gap between the board's picks and the field, −0.17 points per period, comes with a 95% confidence interval from −1.98 to +1.63 points. The interval includes zero, and most values inside it are small: the board survives as a screening device, not as a ranking of expected returns.

The fifteen techniques, in plain language

The Timing board lists 23 techniques. Fifteen of them are mechanical enough to be judged automatically, and those fifteen are the study's denominator. They fall into four families. This section states what each technique measures, in plain language, so that a reader who sees "7 signals on" knows what those seven are.

Trend (4 techniques). Trend tools ask whether the price is moving in a sustained direction, and whether that direction is strong.

Breakout and entry (7 techniques). These tools look for a price breaking out of a range or a pattern, or for a pullback to a level where a buyer might enter.

Oscillator (3 techniques). Oscillators measure whether a stock is overbought or oversold, or whether momentum is strengthening or weakening.

Volume (1 technique). Volume confirms whether a price move has participation.

The remaining eight techniques on the board — interpretive or intraday methods such as candle patterns and Wyckoff — produce no machine signal and are not in the study's denominator. They are shown on the board for a human reader, but they cannot be judged automatically and therefore cannot be counted in a confluence score.

How the Timing board judges a signal

A signal is "on" if its rule fired within the last three trading days, using only data available on that day. That definition has three parts, and each one matters for interpreting the board.

The rule is the site's own code. The study does not reimplement the techniques; it extracts the same rule functions the live Timing board runs and applies them to daily bars. A signal on the board and a signal in the study are the same computation. That matters because a reimplementation could differ in a threshold or a lookback window and produce different results. Here there is no such gap: the study tests the board as it actually operates.

The three-day window. A technique counts as "on" if it fired within the last three trading days, not only if it fired today. The window exists because a signal is not instantaneous — a breakout on Monday is still relevant on Wednesday, even if no new signal fired on Tuesday or Wednesday. The three-day window is the board's own convention, and the study adopts it. A consequence is that a stock's count can stay high for a few days after the signals stop firing, as the older signals age out one by one.

Only data available that day. The judgement uses only data through the day being scored — no future information. That is the point-in-time rule, and it is what makes the study honest. A signal judged with tomorrow's data would be look-ahead bias: it would know something the trader on that day could not have known. The study's entry is the next trading day's close, for the same reason — signals are computed from closing data, so same-day entry would spend information before it exists.

The ranking. Stocks are ranked by families represented (of four), then by technique count (of fifteen), then by freshest signal. Families come first: a stock with signals from three families ranks above a stock with the same count from a single family, on the reasoning that agreement across different kinds of evidence is more meaningful than agreement within one kind. Count comes second; freshness comes third. That ordering is the board's own, and the study tests it as it stands.

Why confluence is seductive — and why it failed here

Confluence is a seductive idea because each technique on the board is defensible. A moving-average alignment is a reasonable way to identify a trend. A breakout is a reasonable way to identify momentum. A momentum oscillator is a reasonable way to identify strengthening. Volume confirmation is a reasonable way to check participation. Counting agreements feels like counting evidence: one indicator can be noise, but several agreeing should mean something.

The catch, which the article states, is that the witnesses are not independent. Four of the fifteen techniques are trend tools. Seven more are breakout and entry patterns that tend to fire after a stock has already moved. A stock with seven signals lit is often not seven independent endorsements; it is frequently a stock that has just run hard, described seven ways.

That non-independence has a specific consequence. If seven techniques all fire because the stock rose sharply last week, then the "seven signals" are not seven pieces of evidence — they are one piece of evidence (the stock rose) counted seven times. The count overstates the strength of the case. A stock with three signals from three different families (trend, oscillator, volume) may genuinely carry more independent information than a stock with seven signals all from the trend and breakout families. The board's ranking — families first, then count — is designed to capture that distinction, but the study's result suggests that even the family-aware ranking did not produce an edge.

The result does not say the techniques are useless. It says that, in this dataset, ranking stocks by how many of them fired at once did not identify stocks that subsequently outperformed. The board survives as a screening device — a fast way to see which stocks have several textbook setups active at once — not as a ranking of expected returns. That distinction is the article's central claim, and it is a claim about the ranking, not about the techniques themselves.

Reading a confidence interval

The article's headline result is a spread of −0.17 points per period, with a 95% confidence interval from −1.98 to +1.63 points. That interval is the most important number in the article, and it is worth reading carefully.

A 95% confidence interval is a range of values consistent with the data. It does not say the true spread is −0.17; it says that, given the 17 periods observed, the true spread could plausibly be anywhere from −1.98 to +1.63, and that values outside that range are unlikely. The interval includes zero. That means the data cannot distinguish the true spread from zero — from no effect at all. It also means the data cannot distinguish the true spread from a small positive effect (up to +1.63) or a small negative effect (down to −1.98). The interval is wide because 17 periods is a small sample.

The width matters as much as the inclusion of zero. A narrow interval around zero would mean the study had precisely measured no effect. A wide interval around zero means the study cannot tell whether there is no effect or a small effect in either direction. Here the interval is wide — four points from end to end — because 17 periods, each an average over 50 stocks, still leaves substantial uncertainty. The article is explicit that the interval's top end is +1.63 points per period, and that a big, reliable reward for agreement should have shown some sign across roughly 1,500 stocks. What the data rules out is a large premium; what it does not rule out is a small one.

The seven robustness variants tell the same story in different words. Every one of their intervals includes zero: 20-day hold (−1.16 to +0.38), 120-day hold (−2.00 to +7.09), family-only ranking (−1.29 to +2.61), top 20 (−2.63 to +3.42), top 100 (−2.03 to +0.78), same-day entry (−2.07 to +1.62), weekly overlapping starts with Newey-West errors (−1.40 to +0.31). The point estimates disagree about the sign — some positive, some negative — which is what underpowered variations do. The intervals agree on the only statement any of them can support: no variation found a positive spread distinguishable from chance.

A reader who prefers one variant — the 120-day hold, which shows +2.54 points, or the top-20 ranking, which shows +0.39 — should note that this dataset cannot confirm those either. The 120-day interval runs from −2.00 to +7.09, which includes zero and is very wide (it is based on only eight periods). The top-20 interval runs from −2.63 to +3.42, which also includes zero. Preferring a variant does not make its interval narrower. The honest reading is the one the article gives: no variation found an edge, and the dataset cannot confirm a small one in either direction.

How to use the board as a screen

The board's plumbing does not change. It remains a fast way to see which stocks have several textbook setups active at once — genuinely useful as a first pass. What changes is the claim attached to the sort order. Treat a stock's position on the board as a reading list, not a ranking of prospects. This section makes that advice concrete.

Open the stock page and look at which families are firing. A stock with signals from trend alone is a different proposition from a stock with signals from trend plus oscillator plus volume. The families-first ranking already encodes that preference, but a reader should check it directly: are the signals spread across families, or are they all trend and breakout tools describing the same recent rise? If they are all from one family, the count overstates the independence of the evidence.

Check whether the signals are fresh or three days old. A signal counts as "on" if it fired within the last three trading days. A stock whose signals all fired today is in a different position from a stock whose signals fired three days ago and are about to age out. Freshness is the third element of the ranking, and it is the one most likely to have changed since the board was last refreshed. A stale count is a weaker count.

Then look at what the board cannot see. The board sees price and volume. It does not see the filings, the valuation, or the reason the price moved. A stock that has seven signals lit because it just reported strong earnings is a different proposition from a stock that has seven signals lit because it was mentioned in a press release, or because the whole sector rose. The board cannot distinguish those cases; only the filings and the news can. The count tells you where to start reading. In five years of the site's own data, it did not tell you what would win.

Why this question

Confluence is a seductive idea because each technique on the board is defensible: moving-average alignment, breakouts, momentum oscillators, volume confirmation. Counting agreements feels like counting evidence.

The catch is that the witnesses are not independent. Four of the 15 auto-judged techniques are trend tools, and seven more are breakout and entry patterns that tend to fire after a stock has already moved. A stock with seven signals lit is often not seven independent endorsements; it is frequently a stock that has just run hard, described seven ways. That could mark genuine momentum, a crowd arriving late, or nothing at all. We wrote down all three hypotheses before computing anything, because a study that can only publish one answer is not a test.

The test

The rules were fixed in a written plan before returns were computed. Its primary metric — the only one this article draws conclusions from — is, for each non-overlapping 60-trading-day period, the equal-weighted average return (minus SPY's) of the 50 highest-ranked stocks, less the same average for the whole eligible universe. Everything else below is description or a robustness check.

Signals are judged by the site's own code, not a reimplementation: the study extracts the same rule functions the live Timing board runs and applies them to daily bars. Of the 23 timing techniques GetCoattail lists, 15 are mechanical enough to judge automatically — the study's denominator; the other eight, interpretive or intraday methods such as candle patterns and Wyckoff, produce no machine signal. A technique counts as "on" if its rule fired within the last three trading days, using only data available that day. Stocks are ranked by families represented, then count, then freshest signal — the board's own ordering (methodology box).

The universe is GetCoattail's registry: 1,631 stocks with daily bars from September 27, 2021 to October 2, 2026. Sixteen listings with fewer than 200 bars were excluded, leaving 1,615 eligible stocks, each entering only after 200 of its own bars exist (the warmup rule). Entry is the next trading day's close — signals are computed from closing data, so same-day entry would spend information before it exists — and exit is 60 trading days later. Returns use dividend-adjusted closes for the stock and SPY alike. Scoring every trading day as a start gives 1,647,112 stock-days over 1,061 start dates, of which 1,548,971 completed a 60-day measurement; those completions fall on 1,000 start dates — the ones on which a full 60-day window could still be measured to the end (the final 61 start dates have no completions, being too close to the end of the data). The rest are counted, with reasons, under Objections and limits. The pipeline was validated against the live board first.

The headline result

MeasureResult (17 periods)
Top 50, average excess vs SPY per period+0.61%
Whole universe, average excess vs SPY per period+0.78%
Spread (top 50 − universe)−0.17 pts
95% confidence interval of the spread−1.98 to +1.63 pts
Periods with a negative spread11 of 17

Primary metric, pre-registered: non-overlapping 60-trading-day periods, starts from July 13, 2022 to May 11, 2026; equal-weighted; entry at the next trading day's close.

Spread between the confluence top 50 and the whole universe in each of the 17 non-overlapping 60-day periods, with the mean and 95% confidence interval
The spread in each of the 17 periods, with the mean and its 95% confidence interval. The losing periods are shown, not hidden.

Read the result at the strength the evidence allows. The top 50 did not beat the average stock — and this sample cannot distinguish the true spread from zero, or from a small effect in either direction. The period-by-period record is mostly a coin flip weighted slightly negative: spreads ran from +6.69 pts (period starting January 3, 2023) to −7.66 pts (period starting May 11, 2026), with 11 of 17 below zero. What the data does rule out is a large premium: the interval's top end is +1.63 pts per period, and a big, reliable reward for agreement should have shown some sign across roughly 1,500 stocks (the period universes run 1,496 to 1,598).

What the buckets show

Average and median 60-day excess return versus SPY by number of buy signals active at once
Average and median 60-day excess return vs SPY, by number of buy signals active at once. n = 1,548,971 stock-starts over 1,000 start dates (the dates on which a full 60-day window could be measured to completion), July 2022 – October 2026. Bucket figures are descriptive; daily starts overlap.

The full daily sample — every start date, not just the 17 period starts — shows the same shape:

Signals onnMean returnMedian returnMean excess vs SPYMedian excess vs SPYBeat SPY
0332,682+6.26%+3.92%+1.31%−1.06%47.1%
1–2722,162+5.39%+3.11%+0.78%−1.41%45.9%
3–4394,898+3.83%+2.17%−0.19%−1.72%44.1%
5–682,770+3.84%+2.36%−0.11%−1.48%44.2%
7+16,459+3.66%+2.13%−0.22%−1.85%43.5%

Two patterns stand out, both labelled description, not findings. First, the average excess return steps down as the count rises: +1.31% with no signals on, +0.78% with one or two, then slightly negative from three upward, reaching −0.22% at seven or more. The share beating SPY falls alongside, 47.1% to 43.5%; high-overlap states are rare (7+ ≈ 1% of observations). Second, the averages flatter every group: the median stock trailed SPY in all five buckets, including the zero-signal bucket (−1.06%). A minority of large winners pulls each mean above its median (chart 3), so "the average stock with no signals beat SPY by 1.31%" and "the typical stock trailed it" are both true.

Distribution of individual 60-day excess returns by signal count
Distribution of individual 60-day excess returns by signal count. Boxes are quartiles; the long right tail is why means and medians tell different stories.

A caution: daily starts overlap, so these 1.5 million observations share market days and are nowhere near independent; we do not test the bucket slope. The inferential claim is the period metric above — the buckets show the raw shape it summarizes, and a slope pointing the "wrong" way belongs next to a null headline, not in a drawer.

Does the result survive?

We re-ran the primary metric seven ways, all specified in advance. A 20-day hold gives −0.39 pts (52 periods; 95% CI −1.16 to +0.38); a 120-day hold gives +2.54 pts, but on only eight periods (CI −2.00 to +7.09). Ranking by family count alone gives +0.66 pts (CI −1.29 to +2.61). Top 20 only: +0.39 pts (CI −2.63 to +3.42). Top 100: −0.62 pts (CI −2.03 to +0.78). Entering at the signal day's close — the shortcut our displayed backtests use — gives −0.23 pts (CI −2.07 to +1.62), so the null is not an entry artifact. Weekly overlapping starts, Newey-West errors: −0.54 pts (CI −1.40 to +0.31).

Every one of the seven intervals includes zero. The checks agree on the only statement any of them can support — no variation found a positive spread distinguishable from chance — while disagreeing about the sign of the point estimate, which is what underpowered variations do. A reader who prefers the 120-day or top-20 variant: this dataset cannot confirm those either.

Objections and limits

The universe is today's universe. The 1,631 stocks are names listed in GetCoattail's registry in October 2026, applied backwards. A stock delisted during the five years is missing entirely — not one sample stock's price series ends before the final date, so the dataset contains zero delisted names. Delisted stocks tend to have poor subsequent returns, so their absence inflates both groups' average returns. The spread between the groups is less exposed than the levels, but not immune, and this dataset cannot measure the distortion. This is the study's largest known bias, and it points in the flattering direction.

Seventeen periods is a small sample, and one market. The periods span the 2022 bear market through 2026 — several regimes in one sample — and the wide interval is the honest consequence. A regime split chosen after seeing the results would be post-hoc, so we report none.

The count is not 15 independent votes. Trend and breakout techniques dominate the 15, so a high count partly restates "this stock has already risen." The bucket pattern is consistent with the overheating story — signals clustering after a run — but consistent is all it is; the mechanism is untested.

Two measurement choices to know. Signals are judged on the unadjusted (split-adjusted only) prices the site's charts display, while returns use dividend-adjusted closes; around an ex-dividend date the two conventions can judge the same day differently. Forward returns count each stock's own trading days, so after a long halt, "60 trading days" can span far more calendar time than SPY's window.

Missing data, counted. Of 1,647,112 eligible stock-days, 96,526 (5.9%) sit too close to the end of the data to complete a 60-day window, and 1,615 (0.1%) had no next-day price; 324,077 earlier stock-days fell to the 200-bar warmup, and the 16 stocks that never reached 200 bars appear nowhere. Costs, slippage and taxes are excluded throughout.

How to read the board now

The board's plumbing does not change: it remains a fast way to see which stocks have several textbook setups active at once — genuinely useful as a first pass. What changes is the claim attached to the sort order. Treat a stock's position on the board as a reading list, not a ranking of prospects. Open the stock page, look at which families are firing — trend alone, or trend plus oscillator plus volume — and whether the signals are fresh or three days old. Then look at what the board cannot see: the filings, the valuation, the reason the price moved. The count tells you where to start reading. In five years of our own data, it did not tell you what would win.

Methodology and data

Past patterns are not a promise of future results. For education only. Not investment advice.

Revision history

Related

A Famous Fraud Detector Flags NVIDIA. The Detector Is Wrong. · Methodology: how every number is measured · Compare overlap

One email a week: the week’s new filings and research. Subscribe free

← The Pelosi Legend Is an Options Book  ·  A Famous Fraud Detector Flags NVIDIA. The Detector Is Wrong. →