News and earnings sentiment agree, mostly at the extremes

This is a very preliminary result (snapshot June 15, 2026) It rests on 21 earnings events from a single three-week window, and every number below should be read as a first sighting, not a finding. We are publishing it now to describe a pattern and to set a baseline we can check against as the sample grows.

What this is about

We run two signals over the same companies.

News sentiment summarizes how the day’s news coverage of a company leans, as a single number from minus one to plus one. A Claude classifier reads each headline about the company, marks it positive, negative, or neutral, and we average them, so a day of all-negative coverage scores minus one, an all-positive day scores plus one, and a balanced or mostly-routine day sits near zero.

What we ask the classifier. It is not asked for generic sentiment. It is asked a target-specific question: is this article positive, negative, or neutral for this particular company as an investment, today? Three rules shape the answer. The default is neutral, and a headline only earns a positive or negative label if it carries genuinely new, company-specific information (an earnings beat, a guidance change, an analyst action, a material event); routine commentary and forward-looking hype are pulled back to neutral. The model is told to weigh meaning over tone, so a record high reported alongside heavy insider selling is negative, a rival’s new product is negative for the incumbent, and a regulatory win that invites oversight is negative. And it is source-aware, treating opinion platforms as neutral by default while giving more weight to catalysts from the wire services. It returns a label and a one-sentence reason.

Our core idea is sound but underexplored. Most validation work in NLP-based finance checks a single signal against returns. Instead, we ask whether two independently constructed signals, one reading news coverage, one reading the earnings event itself, agree with each other. That is a meaningful test of internal consistency before tackling the harder question of what either signal predicts. If the two agree, it suggests both are capturing something real about the quarter rather than noise. If they disagree systematically, that divergence might in itself carry information.

Details and Results:

Earnings sentiment summarizes how a company’s quarterly earnings report came across. It comes from a separate model that we call the Sentiment Surprise Model, or SSM, and have written about elsewhere in this series. It produces two things: a score from 0 to 100, and a five-level label, from STRONG SELL to STRONG BUY, which we call the rating. The label is just the score in bands. In these 21 events the labels fall like this:

Rating

Earnings score (this sample)

STRONG SELL

23

SELL

38

NEUTRAL

45 to 63

BUY

65 to 74

STRONG BUY

81 to 98

These are the score ranges that appear in this sample, not the model’s official cut points, and the two bearish labels are a single stock each.

How the score is built. The SSM reads the earnings-call transcript and scores management’s language with a finance-specific dictionary, but its distinctive step is what it compares that tone against. Rather than asking whether management sounded upbeat in absolute terms, it fits a simple baseline to the company’s own history of past calls (an AR(1) model, which just expects each quarter to sound like the recent ones, drifting toward the company’s long-run average) and measures the gap between what management said and what that baseline predicted. The score is therefore a surprise relative to the company’s habit: a normally cautious CFO sounding mildly positive registers as a real positive, while a perennially enthusiastic CEO having another good quarter does not. That tone surprise is combined with the quarter’s hard results, the earnings and revenue surprise and any guidance change, and placed on the 0 to 100 scale. The exact weighting is the SSM’s own method, set out in its dedicated post.

One signal fires every weekday. The other fires once a quarter, on the day a company reports. The question is whether they agree, and whether the daily news leads the earnings score or only follows it.

What we are (and are not) comparing

It is worth being precise about what these two numbers are. Both are assessments of the same earnings event, made in different ways. News sentiment is how the press talked about the company. Earnings sentiment is the SSM’s read of the quarter itself, built from the hard results and the tone of the call. Neither one is a stock return.

So the correlation below measures whether two ways of reading the same quarter agree with each other. It does not measure whether either one predicts what the stock did. The STRONG BUY and STRONG SELL labels read like trading calls, but here they are the model’s verdict on the quarter, not a realized outcome and not a forecast checked against returns. Whether news sentiment or the rating actually tracks forward returns is a separate and more consequential question, and this snapshot does not touch it: there are no returns in the data. That is the study we would want next.

How to read the numbers

Four terms carry the analysis. N is the number of earnings events in a comparison, and here it is small. Pearson r measures how tightly two quantities move together on a straight line, from minus one (perfect inverse) through zero (no relationship) to plus one (perfect positive); around 0.6 is moderate-to-strong. Spearman ρ asks the same question of ranks rather than values, so it is robust to a few outliers and to curved relationships, and when it agrees with Pearson, one or two stray points are not driving the result. A significance threshold is the value r must clear to be unlikely to come from chance at the usual p < 0.05 bar, and it depends on N: about 0.43 at N = 21, about 0.55 at N = 13. Below the threshold means too little data to detect a relationship, not proof there is none.

The two signals sit on different scales (news sentiment from minus one to plus one, the earnings score from 0 to 100), but correlation needs no shared scale; it only asks whether they rise and fall together. That last point, that a small N can hide a real relationship, matters more than any single number below.

What we did

For each earnings event we attached two windows of daily news. The pre-window is the 14 calendar days before the report: the run-up, where expectations form. The post-window is the 5 calendar days after, when the immediate reaction concentrates before attention rotates to the next quarter. Inside each window we took an article-count-weighted average of the daily sentiment ratio, so a single loud article on a thin day cannot swing a well-covered week.

Our news pipeline only began recording on May 16, and most Q1 calls landed in late April, so the clean sample is late-May and early-June earnings, where both windows fall inside our coverage: 21 events with a clean post-window, 13 with a clean pre-window.

The result

Window

N

Pearson r

Spearman ρ

Critical r (p<0.05)

Significant?

Post-earnings (5 days)

21

0.627

0.529

0.43

yes

Pre-earnings (14 days)

13

0.034

0.030

0.55

no

Post-earnings news tracks the earnings sentiment; pre-earnings news does not. Over the 21 post-window events, Pearson r = 0.627 and Spearman ρ = 0.529, both clearing the N = 21 threshold of 0.43. Because Pearson and Spearman agree, the association is not the work of one stray point. The 13 pre-window events give r = 0.034, indistinguishable from zero, but with a critical value of 0.55 at that sample size, this is best read as no signal detected yet, not no signal possible.

Figure 1. Each point is one earnings event: news sentiment on the vertical axis against the earnings score on the horizontal, with the rating band shown by color and the fitted line dashed. The relationship is clear, but much of it is set by the points at the far left and far right.

The post/pre window asymmetry. The finding that post-earnings news tracks the earnings score (r = 0.627) while pre-earnings news does not (r = 0.034) is the most structurally interesting result in this paper. The explanation that the earnings model draws heavily on the call itself, information the public news stream doesn’t have before report day is plausible and, if it survives scale, has a useful implication: the earnings model is not just repackaging what the press already knew. That would be the minimum bar for the model to add anything.

Why the post-window tracks and the pre-window does not

There are three structural readings of the post-window correlation, and this analysis cannot separate them.

Mechanical, and probably the largest. A genuinely good quarter produces high earnings sentiment and positive coverage at the same time, because both are describing the same surprise, the same guidance, the same call. Two instruments pointed at one event will agree without either causing the other.

News follows the earnings sentiment. After a company reports, analysts publish notes and journalists write follow-ups, and the tone of that coverage tracks the result the model already scored.

Complementary descriptions. The two may be parallel readings of one investor narrative, neither strictly upstream of the other. This is the most accurate and the least quotable.

The common thread is timing. The clean correlation lives after the call. By the time the news agrees with the earnings sentiment, the event they agree about has already happened. That is a description of agreement, not a forecast.

The pre-window null has its own readings. The honest leading one is that N = 13 is simply too small: at that size we would miss anything short of a fairly strong lead. Beyond that, there is a substantive possibility worth stating. The earnings model draws heavily on the call itself (surprise, guidance, management tone), information that does not exist in the public news stream until the company reports. If the earnings signal is mostly made on report day, there is little for two weeks of prior coverage to correlate with, and the run-up news that does exist tends to be diffuse positioning chatter the market has already absorbed. We will learn which story is right once the sample fills in.

The agreement lives in the tails

The most important caveat is not in the headline r. Sort the 21 events by rating and average the news sentiment inside each bucket:

Rating

n

Mean news sentiment

STRONG SELL

1

-0.33

SELL

1

-0.16

NEUTRAL

7

+0.07

BUY

7

+0.06

STRONG BUY

5

+0.32

The buckets climb in order, which is what produces the rank correlation. But the middle is flat: NEUTRAL (+0.07) and BUY (+0.06) sit almost exactly on top of each other. The lift is at the ends. Strongly bearish ratings get clearly negative coverage, strongly bullish ratings get clearly positive coverage, and the large quiet middle barely moves.

The correlation leans on those ends. Drop the two bearish events and r falls from 0.63 to 0.42. Drop the four corner anchors and it falls to 0.25, below the significance threshold. And the bearish end is not a population: STRONG SELL and SELL are one stock each, LULU and WMT. The negative arm of the whole relationship is two data points.

Figure 2. News sentiment by rating bucket; red bars are bucket means. The means rise from STRONG SELL to STRONG BUY, but NEUTRAL and BUY are nearly identical and the spread inside each middle bucket is large.

The tails diagnosis shown by the bucket analysis reveals that NEUTRAL (+0.07) and BUY (+0.06) mean news sentiments are essentially identical is an important contribution. The headline r of 0.63 sounds moderately strong; the bucket table shows that it is almost entirely anchored by the extremes, with the large middle of the rating distribution producing no detectable news differentiation. This is well-observed and candidly presented. It also suggests a natural next question: if the two signals agree mainly where both are obvious (LULU clearly bad, ROST clearly good), what is the marginal information content of having two signals at all?

Why the middle is flat

A few explanations, none of which the data can yet decide between.

Substantive. Unambiguous quarters (large beats, large misses) push both instruments the same way. Mixed quarters, which fill most of the middle of the scale, are exactly where the two signals read different things: the rating weighs hard fundamentals like surprise and guidance, while the news weighs tone and narrative, and those diverge most when a quarter is genuinely in between. Low correlation in the middle may reflect the two signals measuring different aspects of an ambiguous event, not measuring it badly.

Arithmetic. Pearson r is sensitive to extreme points, which sit far from the average and exert outsized leverage on the fitted line. A relationship that is ordered across buckets but flat within the middle can still post a respectable r purely because the endpoints anchor the slope. Part of the headline number is leverage, not broad agreement.

A floor on the news side. Our classifier defaults to neutral unless a headline carries a real catalyst, so routine coverage of mid-rated tickers clusters near zero. With little spread to work with in the middle, there is little there to correlate with anything.

The one clean disagreement: DOCU

Inside the middle, the news scatters across both signs. Among the seven NEUTRAL events, PANW’s news sentiment ran +0.35 while AVGO’s ran minus 0.13; among the seven BUY events, MDT ran +0.27 while DOCU ran minus 0.17. DOCU is the cleanest contradiction in the set: a BUY rating (a score of 70), and post-call news sentiment that averaged minus 0.17 across 63 articles in seven days.

DOCU as a prototype for productive disagreement. The analysis flags DocuSign, rated BUY 70 by the earnings model, but with news sentiment averaging −0.17 across 63 post-call articles, it is the most interesting disagreement in the dataset. The logic is right: systematic disagreements between the two signals are exactly where one of them might be detecting something the other misses.

Three explanations are possible and telling them apart means reading the actual full articles, not the sentiment score. It could be real negative news the rating did not weigh: a post-call downgrade, a guidance worry, a read-through from a weak peer. It could be our classifier over-reading negativity into ordinary coverage. Or it could be a genuine post-call re-rating, the market moving against the result the model scored. If the first pattern recurs across tickers and quarters, it is the most useful thing in this analysis, a flag for a rating the coverage will not support. If it is the second, it is a note for the classifier. We do not yet know which.

How preliminary this is

To be explicit, because it would be easy to over-read the 0.63: this is a snapshot, not a study. Twenty-one events from a single three-week window is not a sample in any robust sense. The window leans tech and consumer because of when our coverage opened. The whole bearish side of the relationship rests on two stocks. No causal direction is established, and our best guess is that the agreement is mostly two instruments watching the same earnings event, which is the least exciting of the possibilities. None of the numbers here should be cited as effect sizes. And nothing here speaks to returns: we compared two readings of the quarter against each other, not either reading against the stock’s actual move. The right reading is that post-earnings coverage and the earnings score move together, clearly at the extremes and weakly in the middle, on a small and recent sample, and nothing stronger.

What’s next

The sample grows on its own. By Q3 the pre-window backfills for nearly every event in the panel, and the lead-versus-follow question becomes answerable rather than underpowered. The more promising thread is the middle, where the two signals disagree, because that is the only place could be carrying information the other lacks. The next concrete step is to pull the articles behind the DOCU-style splits and ask whether the coverage is real news the rating missed or noise the classifier mistook for signal. Either answer is useful, and the data to settle it accumulates every weekday. We will re-run this in a quarter and report whether the pattern survived.

Why two models, not one

On clear quarters the news and the earnings model mostly agree, so the news alone would say much the same thing. The earnings model earns its place elsewhere: it marks the exact quarter, it stays clean in the sectors where news sentiment is swamped by macro noise, and it hears management’s tone against the company’s own habit in a way no headline does. Where the two disagree, as on DOCU, is where the second model is most likely to be telling us something the first cannot.