A constellation of earnings models.

Most earnings forecasts are a single call made once a year. Atlas treats the forecast as a running estimate instead. It starts before the year begins and updates at every quarterly report, so the number always reflects everything known so far.

Each update is its own model. As real quarters replace estimates, the remaining uncertainty shrinks and the error falls. By the third quarter, with only the final quarter left to estimate, the forecast is at its sharpest.

Banks and operating companies do not report earnings the same way, so one recipe cannot fairly cover both. Atlas is built as a constellation of specialist models, each named after a star, with a router that sends every company to the one built for it.

Behind each specialist sit more of the constellation, a set of forecasting models that read a company from different angles. How those stars connect is the part kept under the hood.

A forecast is only useful if you know when to trust it. Alongside every estimate, a separate model grades how reliable that call is likely to be, before the result is known.

Calls are graded HIGH, MED, or LOW. Atlas acts on the dependable ones and sets the rest aside, rather than presenting a shaky number as if it were precise. As the year reports, more companies become reliable, from about 71% before the year to 85% by the third quarter.

Across more than 112,000 company-years from 2015 to 2025, the forecast beats a naive carry-forward baseline at every stage, and the gap widens as the year reports.

Average error falls from $2.02 before the year to $0.55 by the third quarter. On the calls graded reliable it reaches about $0.13, and on the HIGH tier alone about $0.10. Fit to the realized number climbs the same way, from an R-squared of 0.78 across the whole universe to roughly 0.99 on the reliable calls.

The 80% range is calibrated, not cosmetic. Real coverage lands at 77.9% overall, 77.5% for banking and 78.4% for operating, close to the 80% target and slightly tight rather than loose, which is the safer side for a reader.

The charts above are aggregates. These are individual companies, each a real walk-forward call, so the forecast for a year never saw that year. Within every panel, each strip is one fiscal year and the four points are the quarterly stages.

It holds on names you would recognize too. These are large, familiar companies over their last two years, picked by name alone rather than by how well they scored.

Three months before the books close, the stage-3 estimate usually lands on the realized number, and often well away from the prior year. On the names the model is most confident about, the median miss is about $0.03, and every one of them was already flagged HIGH before the year began.

A model that gets sharper as quarters arrive is exactly the kind to suspect of peeking. So the results are built to be checked, with the same discipline as the rest of the data work.

Beyond those output checks, every input feature was rebuilt from data public at each cutoff and compared with what the model was actually fed. Points that land on the diagonal were knowable in advance, so anything off it would be leakage. The stage contract and target isolation match exactly, the peer-rank scatter is symmetric noise from different peer-set definitions rather than a one-sided tilt, and the macro join lands on the prior-quarter reading.