A Day Is Now What a Decade Used to Be

How AI is compressing market efficiency, and what it does to the sentiment signal

Here is a question we have been circling for a while, and that the financial sentiment literature has not asked sharply enough.

Why does sentiment predict returns at all?

The textbook answer is that markets are slow. A positive headline drops at 4:01 PM. By 4:30, sell-side analysts at maybe a dozen banks are scrambling to update their models. By 6 PM, three of them have published preliminary notes. By 9 AM the next day, the buy-side has read those notes, decided, and placed orders. By close on day one, the price reflects most of the new information.

The signal that academics measure, the correlation between a positive label at time T and a positive return at T+1, exists in the gap between when the news arrives and when the analytical chain finishes processing it. Mechanistically, it is a measurement of how long the chain takes.

Make the chain shorter and the signal weakens. Make it short enough and the signal disappears.

This is the world being built right now. Claude is in production at JPMorgan, Goldman Sachs, Citi, AIG, and Visa. The same tool can pull data from Moody’s and S&P, build a discounted cash flow model in Excel, draft an equity research note in Word, format the deck in PowerPoint, and ship it to the PM’s inbox before the analyst has refilled their coffee. Norway’s sovereign wealth fund (NBIM, the largest single equity holder on the planet) reports 213,000 saved hours of analytical labor and a workflow that automatically monitors news flow for nine thousand companies.

AI rolls in like the weather. You sneeze; the world is different.

We described one piece of this earlier, in our essay on attractor markets, where we argued that AI-driven convergence in data, models, and objectives produces self-reinforcing basins of behavior: local stability built from shared representations rather than from economic reality. That essay was framework-level, deliberately speculative.

This one is the empirical companion.

The lag+1 sentiment signal, viewed from the right angle, is a measurement of how shallow the basin currently is. It captures the time gap that exists when analysts are still sufficiently diverse and slow that news takes a day to fully reflect in price. As the basin deepens, the gap closes. The signal we have been measuring across two decades of literature is the depth of the basin, viewed from a particular angle. And we are watching it shrink.

What we are watching

The Claude integration into Excel and PowerPoint and Word does not eliminate human analysts. It compresses them.

Where a junior analyst used to spend three hours building a comparable-companies table, the table now exists fifteen seconds after the analyst types the prompt. Where coverage updates used to ship the morning after earnings, they ship while the call is still live. Where the synthesis of news flow across nine thousand portfolio companies used to require a research department, it requires a Claude session.

None of this individually changes market efficiency. Collectively it does something more interesting. It compresses the time-to-decision across every step of the institutional research pipeline simultaneously. The research-to-trade lag, which was the substrate the sentiment signal lived inside, is being eroded from every direction at once.

If we assume reasonable adoption rates (which we do not have to assume, because the institutions are publishing them), the lag in IT and Financials is probably already an order of magnitude shorter than it was in 2023. By the time these tools are mature in Materials and Energy, the lag in those sectors will collapse too.

What does that do to the cross-sector matrix we just published?

A specific, testable prediction

If the analytical-lag mechanism is what produces sentiment-to-return correlation, and AI-driven tooling compresses that lag, then the criterion-by-sector spreads should diminish fastest in the sectors with the highest tooling adoption.

Concretely. The IT-quantitative cell of our matrix, where the strict prompt found a +2.5pp excess spread on lag+1 returns, should shrink first. The Financials-event cell, second. The Materials-quantitative cell, where IT-style tooling adoption is slowest, should hold longest.

And the Consumer Staples cells, which already look weak, are the canary. They may have collapsed already. That would explain why they look the way they do in our data, not because the news is uninformative but because it has been priced before we measured it.

This is testable in a way few claims about market efficiency are. We have the cross-sector matrix on April 2026 data. By Q1 2027 we will have a quarterly version. If the cells contract in the order predicted, the mechanism is real. If they do not, our story is wrong, and something else is producing the heterogeneity.

Either way, we will know more than the literature currently does about whether the sentiment signal is a permanent feature of market microstructure or a temporary artifact of pre-AI analytical bandwidth.

What replaces the signal

The genuinely unsettling possibility is not that sentiment stops working.

It is that sentiment-as-signal is replaced by something we do not yet have the vocabulary for.

Consider what “news” actually is when every major buy-side desk is running an LLM agent that reads, classifies, and recommends action on every public communication in real time. The headline arrives. It is processed by ten thousand instances of Claude or its competitors at fifty institutions, each instance configured slightly differently but trained on overlapping corpora. The recommendations converge within minutes. The trades execute within seconds of the recommendations. The price moves before any human has read the article.

In this world, the lag+1 sentiment signal does not just shrink. It inverts.

The headline at 9 AM is already in the price by 9:05. By close on the same day, the sentiment-positive event has produced its full price response, and the lag+1 return measured the following day is whatever happens after the AI consensus has already played out. What you see at lag+1 is reversal, profit-taking, second-order effects: the noise that follows the signal rather than the signal itself.

This is not a hypothetical. The early signs are visible in the IT sector data we just published, where the aggregate lag+1 correlation is already inverting across multiple classifiers. Our regime adjustment cleaned up some of that. Not all of it. Part of what we see in the IT cells of the matrix may already be the post-AI return regime.

We cannot prove this from sixteen days of data. The shape is suggestive.

What this means for the field

Two decades of financial sentiment research, including the most recent wave of LLM-based classifier work, has been measuring something that is being structurally erased.

Not because the work was wrong. Because the conditions under which the work was conducted are changing faster than the work can adapt.

The benchmarks the field reports (”FinBERT achieves 97% accuracy on FPB,” “this classification prompt produces a +0.5pp spread,” “this prompt outperforms that prompt by X percentage points”) are measurements made on data generated in a world where human analysts were the bottleneck. We are leaving that world. The benchmarks may not survive the transition.

Three implications follow. They are uncomfortable enough that we will state them plainly.

First. Most published sentiment-trading strategies are probably already obsolete. The strategies were calibrated on pre-LLM data, when the analytical-lag mechanism was at full strength. They are being deployed now, when the mechanism is collapsing. The backtests look fine. The forward returns will not.

Second. The academic literature on sentiment classifiers is becoming an exercise in measuring the past. A 2027 paper that reports “our model achieves 0.X correlation between sentiment and lag+1 returns on the 2018-2022 corpus” is doing solid work, but it is doing solid work on a phenomenon that is structurally different in 2027. The publication cycle is too slow to track the underlying change.

Third. The buy side knows this. The adoption rates suggest the practitioners have figured out what the academics are still calibrating. NBIM does not save 213,000 hours by accident. They save it by reorganizing around tools that compress the analytical lag, which means the firms that move first are extracting whatever signal remains while it remains. By the time everyone has the tool, no one has the edge. We are perhaps three years from that point.

What might survive

Not all sentiment work is in the same boat. A few categories of signal seem more durable than the lag+1 headline correlation, and they are not random. They are, in the language of our earlier essay, instances of dimensional arbitrage, the dimensions that dominant optimization systematically underweights, which is exactly where the surviving signal lives.

Sub-headline information. Order book dynamics, insider trading patterns, options flow, language patterns inside earnings call audio rather than transcripts. These operate at timescales the AI tools have not yet collapsed. The lag from a CFO’s voice cracking on the call to the price reflecting that crack is still measured in minutes. The tools that read transcripts are not yet listening to audio. The basin formed around text; the signal moved into voice.

Genuinely private information. Primary research, expert networks, channel checks. By definition not amenable to LLM aggregation. If it is in your head and not on the public internet, no Claude can read it. This is the part of the analyst job that is most labor-intensive and most defensible. The basin cannot form around information it cannot ingest.

Long-horizon analysis where the signal arrives over months rather than days. Sector rotation, secular thesis development, structural macro positioning. These do not depend on the lag+1 mechanism. They depend on judgment about which trends matter and how much to weight them. That is exactly what humans plus LLMs are jointly better at than either alone. The same-day pricing regime structurally cannot produce these signals because the relevant feedback arrives outside the optimization window.

Each of these is the same kind of trade. Positioning in dimensions the dominant system underweights. The basin deepens around the dimensions that can be measured cheaply and reacted to quickly. It cannot deepen around the dimensions that are private, slow, or sub-symbolic.

If the lag+1 sentiment signal is the canary, what comes after will be different work. More private, more long-horizon, more dependent on judgment that frontier models do not yet replicate. Probably less democratically accessible, since these advantages are exactly what scale lets large firms produce.

What we are doing about it

We are running the cross-sector criterion analysis on a quarterly window. If the predictions above hold, we will see it. If they do not, we will say so.

We are watching the IT and Financials cells specifically, because they are the leading indicators. If the lag+1 correlation in IT is meaningfully weaker on Q3 2026 data than on Q3 2025 data, the compression is real and probably accelerating.

We are also asking a question we did not know to ask six months ago. What is the sentiment-to-return relationship at lag zero (same-day, intraday) in sectors with high AI tooling adoption? If the signal is moving from lag+1 to lag-zero, the field needs to relearn how to measure it.

There is something disorienting about publishing a methodology in the same window that the methodology is becoming obsolete. We are doing it anyway. The cross-sector heterogeneity holds for the data we have. By the time it does not, the absence itself will be the finding: visible evidence of basin formation, measurable in the disappearance of a signal whose existence depended on the diversity that is being optimized away.

The work continues, as the ground continues to move.