Short-Term Data Divert Analysts’ Attention from the Long Term

Meta reports its second-quarter results after U.S. markets close on July 29, 2026. The company has guided revenue to $58 billion to $61 billion for the quarter and expects capital expenditures of $125 billion to $145 billion in 2026. Across the largest technology companies, S&P Global estimates that hyperscaler capital spending could rise from $384 billion in 2025 to $682 billion this year and $878 billion in 2027. Quarterly revenue can be measured within weeks. The return on infrastructure built today will emerge over years.

That split between what can be observed now and what must be forecast further ahead sits at the heart of research by Thierry Foucault, Olivier Dessaint and Laurent Frésard. Their study asks whether the expansion of alternative data improves financial forecasting. It finds that data focused on current activity can make analysts’ near-term forecasts more informative while reducing the information contained in their forecasts several years ahead.

The paper, Does Alternative Data Improve Financial Forecasting? The Horizon Effect, develops a model of how forecasters divide their effort across different horizons, then tests its predictions with U.S. equity analysts’ earnings forecasts and the expansion of StockTwits. This is a major social media platform designed for sharing ideas between investors, traders, and entrepreneurs. The evidence covers more than three decades and separates a broad historical trend from a more focused test of analysts’ exposure to social-media data.

How can more data reduce information?

Alternative data include social-media posts, web traffic, credit-card and point-of-sale records, geolocation, satellite images and employee reviews. Many of these sources reveal what consumers, investors or employees are doing now. Existing research therefore finds that they are most useful for predicting outcomes over the next year, with much less evidence that they improve forecasts several years ahead.

The researchers model an analyst who must forecast both short-term and long-term earnings. The analyst can collect information about assets and activity already in place, which helps at both horizons, and information about growth opportunities, which matters mainly for the longer term. Collecting and processing either type of information requires effort, and handling several forecasting tasks carries a cost.

When short-term-oriented data become cheaper and easier to obtain, analysts have an incentive to use more of them. That improves the precision of their short-term information. It can also shift effort away from information that is useful only for long-term forecasts. The model therefore predicts a change in the allocation of attention: better forecasts close to the present, accompanied by less informative forecasts further out.

This prediction depends on the horizon covered by the new data. Information about long-term outcomes should make long-term forecasts more informative and can draw effort away from shorter horizons. The paper’s argument concerns the direction of the information, rather than alternative data as a single category.

What happened to forecast quality?

The authors first examine earnings forecasts from the Institutional Brokers’ Estimate System between 1983 and 2017. They calculate forecast informativeness for every U.S. analyst, every day and every available horizon from one day to five years, producing more than 65 million analyst-day-horizon observations.

Their measure asks how closely an analyst’s forecasts across the firms they cover line up with those firms’ eventual earnings. A higher figure means that the forecasts explain more of the differences in realized earnings and leave less uncertainty after they are observed. A value of 100% would mean that the forecasts remove all such uncertainty at that horizon.

Forecasts naturally become less informative as the horizon lengthens. Across the full sample, informativeness falls by about 12 percentage points for every additional year. The more revealing result is how that relationship changed. Since 2000, average informativeness at the one-year horizon rose from roughly 60% to 70%. At five years, it moved in the opposite direction, from about 40% to 30%. The gap between the two horizons consequently doubled from around 20 to 40 percentage points.

The divergence accelerated after 2005 and appeared across most industries. It was also more pronounced in industries covered by analysts who used more alternative data. These historical patterns fit the model, although the authors stress that the long-run trend alone cannot establish that alternative data caused the change.

What did StockTwits add?

For a tighter test, the researchers study StockTwits, an investor network introduced in 2009. Users post short messages linked to company ticker symbols, mark views as bullish or bearish and place firms on watchlists. The platform expanded gradually and unevenly across companies, exposing analysts to different amounts of newly generated social-media information depending on the firms they followed.

The paper first establishes that this information is oriented toward the short term. StockTwits ratings predict subsequent growth in sales, earnings before interest, taxes, depreciation and amortization, operating earnings and net income at short horizons, but the predictive relationship disappears further ahead. The authors also present evidence that analysts consume the information: analysts issue forecasts more frequently after activity rises on the platform, including on days without news from traditional sources, and their upgrades and downgrades move with bullish and bearish messages.

The scale of the test is substantial. The StockTwits analysis contains 31,623,819 daily analyst-horizon observations from 2005 to 2017. The average analyst covers 10.35 firms. Once the platform is established, each covered firm is followed by an average of 321 StockTwits users and generates the equivalent of 4.6 hypothetical messages a day under the researchers’ exposure measure.

The authors use two measures designed to capture information generated by the platform rather than outside news. One tracks growth in the number of users placing covered firms on their watchlists. The other estimates messages using each firm’s fixed share of overall platform activity. Both measures are set to zero before StockTwits begins, and the analysis controls for persistent differences among analysts and for events affecting all analysts on the same date.

A one-standard-deviation increase in exposure to StockTwits data raises short-term forecast informativeness by 0.54 to 0.66 percentage points for horizons of up to one year. For forecasts beyond two years, the same increase reduces informativeness by 1.51 to 1.92 points. The researchers find no significant effect between one and two years, placing the point at which the effect changes direction inside that interval.

The pattern remains when the researchers hold each analyst’s 2009 portfolio of covered firms fixed and use the later growth of StockTwits around those firms to predict exposure. The term structure of forecast informativeness effectively rotates around the one-year horizon: the near end improves, the distant end weakens, and the one-year point itself changes little.

Who is most exposed to the horizon effect?

The effect is stronger for analysts who cover more firms. This result matches the model’s prediction that shifting effort between horizons matters more when multitasking is costlier. It is also stronger when the earnings of covered firms are less persistent. When current earnings provide less information about future earnings, a short-term signal carries less value into the distant forecast, making the loss of specifically long-term information more consequential.

The authors provide further evidence that social-media data had entered analysts’ working environment. Prior research cited in the paper ranks social media as the second most frequently used form of alternative data among securities analysts, behind app usage and level with point-of-sale information. In the study’s own matching exercise, 35% of 7,655 analysts’ names correspond to StockTwits account holders. 

The size of the StockTwits effect is modest for any single forecast, but meaningful relative to the historical movement in the relationship between horizons. The estimated change in the slope represents about 19% of its long-run variation at the aggregate level, 16% across industries and 7% across individual analysts.

Does every new data source create the same trade-off?

The effect depends on what the new source can predict. The paper’s evidence concerns data that are informative about outcomes within roughly one year. A source that reveals research pipelines, durable investment, product development or another genuinely long-term outcome should improve long-term forecasts under the model, potentially while reducing effort devoted to short-term information.

The findings also concern the information contained in forecasts across firms, rather than the absolute error of every individual estimate. The historical sample ends in 2017 and does not test generative AI, current data vendors or the present investment cycle. StockTwits provides a setting in which a short-term-oriented source arrived gradually and could be measured, allowing the researchers to examine the allocation predicted by their model.

The central distinction is therefore the forecasting horizon a dataset can inform. A larger volume of current signals can improve the next earnings estimate while leaving questions about growth opportunities, strategic investment and earnings several years ahead with less analytical attention.

Where does today’s earnings season leave the long term?

When Meta releases its figures on July 29, analysts will receive exact quarterly revenue, costs, margins and updated guidance. Those disclosures will quickly refine forecasts for the months ahead. The economic return on the company’s $125 billion to $145 billion investment program will remain a multi-year forecast.

The study does not analyze Meta or artificial-intelligence spending. The current reporting cycle displays the same separation between horizons that the research measures. With hyperscaler capital expenditure projected at $682 billion in 2026 and $878 billion in 2027, the decisive information question is whether the expanding data stream reaches the years in which those investments must produce earnings, or mainly sharpens the view of the next quarter.

If analysts’ long-term forecasts have become less informative, as the study finds, investors face a greater risk of misvaluing investments whose returns will emerge only years later. The error can run in either direction. Overvaluation can contribute to a bubble and overinvestment. Undervaluation can produce the less frequently discussed problem of underinvestment. Understanding how securities analysts form forecasts at different horizons therefore matters not only to traders in financial markets, but to the economy as a whole: it can affect whether investment in long-term projects rises above or falls below what would be optimal.