Beyond Consensus
Sequential gradient boosting for full-year EPS, and the returns to analyst disagreement
Mathéo Menges · 14 September 2026
Abstract
We develop a sequential forecasting model for full-year earnings per share (EPS) that updates as quarterly earnings are reported. It combines an annual forecast, a forecast of the next unreported quarter, and realised earnings using only public information available at each forecast date. We call this sequential updating process a revision cascade. We evaluate the model against the I/B/E/S consensus without using analyst forecasts as model inputs. Across 16,991 firm-year-stage observations for non-financial firms, at matched information and actuals, the model’s mean absolute error is 0.925 times that of consensus, improving from 0.953 before any quarter is reported to 0.757 after three quarters. The difference between the model and consensus also predicts subsequent earnings surprises and supports a quarterly long–short strategy with a higher Sharpe ratio than the market, and adds about 5–8% a year when overlaid on the S&P 500. The results show that the model provides an independent benchmark for analyst consensus and that the disagreement between the two forecasts contains information about future earnings.
Cite this paper
Menges, M. (2026). Beyond Consensus: Sequential gradient boosting for full-year EPS, and the returns to analyst disagreement. AtlasEQ Working Paper No. 1.
This paper is adapted from research originally developed by Mathéo Menges for the 2026 CFA Quant Awards, an international quantitative-finance research competition organized by participating CFA Societies. This version corrects the one submitted to the competition (Section 2.6 and Appendix D). It is not a submission to, publication by, or endorsement of CFA Institute or any participating CFA Society. The research used academic data sources that are not incorporated into, connected to, or used as data inputs for the AtlasEQ terminal or its production data infrastructure. It is provided solely for research and informational purposes and does not constitute investment advice. Past research results and historical backtests are not indicative of future performance.