A constellation of earnings models.

A conventional earnings estimate is one call, made once, and left to age. Atlas treats it as a running quantity: an estimate is set before the fiscal year opens and revised at each quarterly report, so it always reflects the results filed to date.

Each revision is a separate model with its own cutoff. As reported quarters replace estimated ones, less of the year is left to predict and the error falls with it. By the third stage only the fourth quarter is unknown, and accuracy is at its highest.

Banks and operating companies report earnings on different bases, and a single model fitted across both is worse at each. Atlas is a set of sector specialists, each named after a star, with a router that assigns every company to one of them.

Each specialist is itself several models, reading a company from different angles and combined into one estimate. How they are combined is not published.

An estimate is only usable if its likely accuracy is known in advance. A second model grades each forecast before the result is filed, using only what was available at the time.

Grades are HIGH, MED or LOW. The low-confidence estimates are flagged and withheld rather than reported at a precision they do not have. The actionable share rises as the year files, from about 69% before it opens to 85% after the third quarter.

The 80% range is held to the same standard. A stated range is only worth quoting if the reported figure lands inside it about 80% of the time, and out of sample it does.

Across 28,000 company-years from 2015 to 2025, the forecast beats a carry-forward baseline at every stage, and the margin widens as the year files.

Mean absolute error falls from $2.09 before the year opens to $0.59 after the third quarter. On the estimates graded reliable it reaches about $0.18, and on HIGH alone about $0.12. Operating companies drive that figure: their reliable stage-3 error is about $0.12, against roughly $0.43 for banks, whose earnings are harder to pin down. Fit follows the same path, reaching an R-squared of 0.78 across the whole universe at the third stage and roughly 0.98 on the reliable subset.

The chart on the right compares the same forecast with the paid sell-side consensus. That test is smaller and stricter: 21,209 company-years across 2,269 firms, limited to companies the analysts cover, and scored against the GAAP-basis consensus so both sides target the same definition of profit.

The model is closer to the reported number than the paid consensus at every stage, and the margin widens as quarters report: 7% closer before the year begins, 34% closer after two quarters, and 65% closer after three. Across all stages its error is 0.72 times the consensus error.

Realised coverage of the 80% range is 80.2% overall, 80.7% for operating and 78.6% for banking. On target, with banking marginally tight.

The figures above are aggregates. Below are individual companies. Every point is a walk-forward call, so no forecast for a year saw any data from that year. Each strip is one fiscal year and the four points are the four stages.

The first panel is companies the model tracked closely over four years, which is the system at its best rather than its average. The second is selected on name recognition alone. Typical accuracy is the aggregate above.

The same holds on widely followed names: large companies over their last two fiscal years, selected by recognition rather than by score.

Three months before the books close, the stage-3 estimate generally lands on the reported figure, and often well away from the prior year. Across these names the median absolute miss is about $0.11, and all but one were graded HIGH before the year opened.

A model whose error falls as the year reports is exactly the kind that should be suspected of having seen the answer. What follows is the evidence against that, together with the conditions under which the model loses.

The model does not win everywhere. Before the year starts, on the largest earners, the paid consensus is still closer: analysts have guidance and management access that the model does not. By the third quarter that advantage is gone in every size band.

Nor is it unbiased. Across the matched sample it forecasts below the reported number by about 5% of prior-year EPS, worst before the year starts and close to zero by the third quarter. The consensus leans the other way and much harder, about 13% high on average and 19% before the year begins, the long-documented sell-side tilt.

That bias is a level effect rather than shrinkage. Regressing the reported change in earnings on the predicted change gives a slope of 0.95, 1.00, 1.02 and 1.00 across the four stages, so the model is calling the full size of the move rather than hedging toward the prior year.

Those are checks on the output. On the input side, every feature was rebuilt from the data public at each cutoff and compared with the value the model was actually fed. A point on the diagonal was computable in advance; mass off the diagonal would be leakage. The stage contract and target isolation match exactly, the peer-rank scatter is symmetric noise from differing peer-set definitions rather than a one-sided tilt, and the macro join lands on the prior-quarter reading.