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.

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. About two thirds of estimates or more are graded MED or HIGH at every stage (71%, 70%, 71% and 63% from stage 0 to stage 3); the bar for those grades tightens as the year files.

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 lands inside 77.6% of the time (2024–2025 test years), slightly narrow.

Across 23,400 company-years of operating companies 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 $1.67 before the year opens to $0.63 after the third quarter, against about $2.80 for the carry-forward baseline. On estimates graded reliable (MED or HIGH) it reaches $0.18 at the third stage, and on HIGH alone $0.12. Fit follows the same path, reaching an R-squared of 0.93 across all operating companies at the third stage and 0.97 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: 16,991 forecasts on 5,604 company-years across 1,389 operating companies, FY2010–2025, graded MED or HIGH, scored against the GAAP-basis consensus nearest the forecast date, and limited to firm-years where the two reported actuals agree within 1%, so both sides are graded on the same number.

The model is closer to the reported number than the paid consensus at every stage, and the margin widens as quarters report: 5% closer before the year begins, 10% closer after two quarters and 24% closer after three. Across all stages its error is 0.925 times the consensus error.

Realised coverage of the 80% range is 77.6% in the 2024–2025 test years, slightly narrow.

The figures above are aggregates. Below are six companies. Every point is a walk-forward call: the model that made it was trained only on earlier years. Each panel shows FY2024 and FY2025, and the four points in each year are the four stages.

UNP, XOM and COST are widely followed names, shown whatever the result; PCG, EXC and AWR are among the companies the model tracked most closely and are a best case, not the average. Costco is graded LOW throughout, and in FY2025 its reported figure lies above the range from the first quarter on. Across the six, the median miss at the third stage is $0.06, against $0.065 for the consensus, and the reported figure falls inside the 80% range in 43 of 48 forecasts.

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. Among the larger earners the consensus stays closer through the first quarter; by the third quarter the model is closer in every size band.

Nor is it unbiased, but it leans the same way as analysts, only less. Across the matched sample its mean signed error is +8.8% of prior-year EPS, against +21.0% for the consensus. By stage the model is +17.1%, +9.7%, +6.5% and +0.4%, close to zero by the third quarter; the consensus is +35.9%, +22.6%, +16.1% and +6.7%, the long-documented sell-side tilt.

Regressing the reported change in earnings on the predicted change gives a slope of 1.01, 1.04, 1.02 and 0.99 across the four stages, so the model calls the full size of the move rather than hedging toward the prior year.