Do Stock Scores Work? We Tested Ours on Four Months of Returns
We checked four months of recorded scores for 1,916 US stocks against what each did next. The composite had no edge; fundamentals within sector held up.
Every night since July 15, 2026, StocksScores has saved the score it gave each stock and the price at that moment. That makes a plain test possible: take the scores exactly as they were published, and see what the stocks did next. No refitting, no hindsight about which scorers to keep.
We ran it on the 1,916 US-listed stocks worth $2 billion or more, over the four months to November 14, 2026. The result was not flattering, and it is the reason the score is built differently today.
The short version
- The composite score had no edge. Its rank correlation with the next four weeks’ return over SPY averaged −0.04. Over those four weeks the top-scored fifth of stocks trailed the bottom fifth by 2.1 percentage points on average.
- Fundamentals, compared within their own sector, were the one consistent signal: +0.08 at four weeks, positive on 85% of days.
- Technicals pointed the wrong way: −0.10 at four weeks, and −0.12 among stocks worth $10 billion or more. The most technically bullish fifth trailed the least bullish by 4.1 points.
- News sentiment added nothing (−0.03), which is why it was already shown but never counted.
- On November 12, 2026 we rebuilt the composite around this: 80% fundamentals, ranked within sector, and 20% technicals. Rescored on the same four months it reads +0.04 at four weeks — better, still small, and measured on the very data that suggested it. The honest test is the next four months.
How the test works
On each trading day we line up every stock twice: once by its score that night, once by how it did over the following weeks compared with SPY. Then we measure how closely the two orders agree, using a rank correlation (Spearman’s). In quant finance this is called the information coefficient. A value of 1 would mean the score ordered every stock perfectly; 0 means no relationship; a negative value means higher-scored stocks did worse. Nothing real sits near 1. A signal that averages around +0.05, day after day, is usually considered worth having.
- Entry: the score and price recorded the night a stock was scored, after the close.
- Exit: the same stock’s recorded price 7, 28 or 56 days later.
- Return: the stock’s return minus SPY’s over the same two recordings; SPY’s price is stored in the same row, so both are timed identically.
- Sample: 81, 66 and 47 entry days for the three horizons, about 1,718 stocks a day.
- Scores as published: until November 12 the composite was an even blend of the technical and fundamental pillars, and that is the composite tested here.
We average the correlation across days instead of pooling every observation, so one volatile week cannot dominate. Because a four-week return overlaps the next day’s four-week return, the daily readings are not independent; the significance figures below account for that.
The scorecard
| Signal | 1 week | 4 weeks | 8 weeks | Days positive (4 wk) |
|---|---|---|---|---|
| Composite score (as published: 50% technical, 50% fundamental) | −0.025 | −0.040 | −0.038 | 50% |
| Technical pillar | −0.057 | −0.100 | −0.061 | 35% |
| News sentiment (shown, never counted) | −0.020 | −0.030 | −0.015 | 33% |
| Fundamental pillar, raw | +0.012 | +0.043 | +0.020 | 62% |
| Fundamental pillar, ranked within its sector | +0.029 | +0.078 | +0.070 | 85% |
| Today’s composite: 80% sector-neutral fundamental, 20% technical (in-sample) | +0.012 | +0.042 | +0.045 | 71% |
The top-scored stocks did worst
Sorting stocks into fifths by score makes the problem concrete. Under the old composite, returns fell steadily from the lowest-scored fifth (+0.40% over four weeks) to the highest (−1.74%). The technical pillar alone did the same, more steeply: +1.57% for the lowest fifth against −2.56% for the highest. Fundamentals ranked within sector leaned the right way, mostly by keeping investors out of the weakest names: the bottom fifth returned −1.50%, the third and fourth fifths +0.22% and +0.28%, and the top fifth was about flat (−0.05%).
Why the technical half failed
Our technical total turned out to be mostly one thing. Across the nightly cross-section it correlates 0.98 with its own trend-following component — where price sits against its moving averages, crossovers, slope — because about half of every timeframe’s weight sits on those indicators. So a high technical score largely meant this stock has been rising.
Over a few weeks, recent winners tend to give some of it back. Short-term reversal is one of the oldest documented patterns in stock returns (Jegadeesh, 1990), and it is what this window shows. Price momentum does work in the research, but over longer formation periods — typically three to twelve months — which four months of history cannot test. The finding is narrow: a trend-heavy technical score, judged a few weeks out, ran backwards.
Why fundamentals needed the sector adjustment
Raw fundamental scores averaged only +0.04. A bank and a software company are not on the same scale for margins, leverage or valuation, so raw scores partly sorted stocks by industry: on a typical day, sector alone explained about 15% of the spread in raw fundamental scores. Over a few weeks, a sector’s swing can easily outweigh what separates one company from its neighbours. Ranking each company only against its own sector lifted the four-week correlation to +0.08 and made it positive on 85% of days; at eight weeks it held at +0.07.
What we changed
The changes went live on November 12, 2026, and every score since uses them:
- The composite is now 80% fundamental, 20% technical.
- The fundamental total is sector-neutral: a company’s standing inside its sector, mapped onto the market-wide scale, so a 70 means the same thing in banking and in software.
- Analyst estimate revisions are read over 30 and 90 days instead of 7, and count for more; the price-target scorer measures upside to the mean target.
- A guard keeps earnings-based valuation scores from rewarding cyclical companies at peak profits.
- A timeframe with less than 60% of its indicators available no longer produces a technical total.
We did not tune the new weights to maximise the numbers above. The +0.04 for today’s composite is still in-sample: we built it after seeing these results, on these results.
What this test cannot tell you
- Four months is one market. SPY, weighted toward the largest companies, beat the typical stock over the period: at eight weeks every fifth of every signal was negative against it. A different stretch could read differently.
- Few independent periods. At four weeks, four months hold roughly 3 non-overlapping windows. Adjusted for the overlap, the within-sector fundamental signal has a t-statistic of about 2.2; the technical one about −1.0. Suggestive, not settled.
- Survivors only. The sample is today’s list of stocks above $2 billion, so companies that were delisted, acquired or shrank below the line are missing.
- The engine moved during the test. Scorer fixes shipped in August, inside the window.
- Not a trading result. No costs, no position sizing, no portfolio. A rank correlation says whether an ordering helps, not what it would have earned.
Check it yourself
Everything here can be rebuilt from public data, with no key. The history endpoint now returns the price and SPY’s price recorded with each nightly score:
curl --compressed 'https://api.stocksscores.com/api/analysis/history?symbol=AAPL&days=365'
The daily correlations behind every number on this page are in a CSV: score-ic-2026-11.csv (date, horizon, signal, correlation). The API is described in our free stock API guide, and what goes into each score in how stock scores work. Each stock’s current breakdown is on its page, for example AAPL.
We will run the same test on the new composite once it has four months of its own history, and publish it whichever way it comes out.
Scores are information, not investment advice. Past scores do not predict future returns.