Six AI Agents Forecast Earnings Per Share: Three Matched the Analysts

What happens if you give a language model a company's filings, an investor call transcript and the news, then ask it to predict earnings per share? Researchers at the Georgia Institute of Technology built a benchmark called EPSBench to find out. EPS (earnings per share) is net income divided by the number of shares.

The benchmark uses the diluted GAAP figure, which follows standard US accounting rules, and leaves out the "adjusted" numbers that companies like to report separately. The test covers 149 quarterly reports from 86 large US companies across eight sectors. The announcements ran from December 2, 2025 to June 24, 2026.

Six models from four developers were compared with the analyst consensus as of the same date. The result: three models are no different from the professionals in point-forecast accuracy, and three make noticeably larger errors. All six, however, are overconfident.

Forecasting line by line, not with a single number

The key idea is how the task is structured. Earlier benchmarks such as FutureX and FinDeepForecast simply ask the model to name an EPS figure, and the path to it stays buried in the model's reasoning text. An analyst works differently: revenue first, then costs, then everything below that. EPSBench forces the agent to work the same way.

For each company-quarter pair, the authors build a three-part scheme in advance:

• a base: the values of the report's line items for the same quarter a year earlier;
• drivers: revenue growth rate, change in the cost-of-sales share, growth in operating expenses, tax, share count;
• a graph of accounting identities that assembles these items into the final EPS.

For Microsoft's third quarter of fiscal 2026, the drivers are growth in product revenue and in services revenue, a shift in the cost-of-sales share, growth in operating expenses, other expenses, tax and share count. Several models built the schemes in turn. The check is simple: on year-earlier data, the graph reproduces the actual EPS to the cent in 146 of 149 schemes, and to within four cents in the remaining three.

The agent estimates each driver, states which document it relies on, and sends the values to a shared calculator. The calculator returns an income statement and the final EPS. The agent also gives p10 and p90 bounds, that is, an interval in which it believes the actual result will fall with 80% probability. The calculator handles the arithmetic, so any error can be traced to a specific line item and a specific citation.

A closed sandbox and protection against peeking

The agent operates in a fixed environment. It has access to three types of sources, which it reaches through six tools on an MCP server:

• 10-K and 10-Q filings from SEC EDGAR;
• call transcripts from Seeking Alpha, where management guidance is usually given;
• news from CC-News.

Everything is cut off the day before results are published. The server blocks the report and the call for the target quarter. News items with a late or unknown date are discarded. Shell, files and internet access are disabled in the agent harness. Each configuration was run three times per task, which gives 447 runs per model.

Accuracy is measured by price-scaled error: the absolute difference between forecast and actual, divided by the share price. This avoids the problems of percentage errors when EPS is close to zero. Intervals are scored by hit rate, width, and an interval metric that cannot be improved unless the bounds are stated honestly. There are three baselines: year-earlier EPS, previous-quarter EPS, and the analyst consensus from the IBES database.

Where the models go wrong and what to do about it

The results:

• three models are statistically indistinguishable from the consensus in point-forecast error, and three lose to it;
• the nominal 80% intervals covered the actual EPS in only 37–70% of runs in which an interval was returned at all;
• most misses lie above the upper bound, meaning the forecasts are too low. The analyst consensus, incidentally, has the same bias;
• the gap between the strong and weak trios shows up in quarters where earnings move sharply away from the year-earlier level. The leaders more often shift their estimate far enough and in the right direction;
• almost all of the difference sits in the operating line items, that is, in the estimate of operating profit.

One detail concerns GLM 4.5 Air with its 128-thousand-token context window. In 49 of its 447 runs, the harness compressed the context so that everything would fit. Those runs were scored as they were, and they account for all of the model's missing intervals.

The side findings are useful too. News and filings carry partly different information. Changing the harness and the depth of reasoning moves interval coverage more than it moves the accuracy of the number itself.

The practical takeaway for anyone building financial agents: extract the structure of assumptions from the model, not a single number. That way you can see at which step an error entered the forecast. Stated intervals should not be trusted without calibration. If an agent says "80%," the real figure may be half that. The authors publish the schemes, tools, prompts, forecast records, scores and agent trajectories, so the comparisons can be reproduced.