Forecasting Quarterly Losses from SEC Filings: A Replay Study with a Guarded Agent Design

Quarterly reports are assembled after the quarter closes, so managers learn about losses, cost overruns and lost deals late. This paper specifies an agentic AI system that works from a company's records during the quarter. Specialist models produce the numbers. A large language model agent proposes adjustments through a restricted set of actions, and a named reviewer approves every report before it is issued. The data are 266,134 company-quarter records of US public companies, drawn from SEC filings for 2015 to 2026. Project and staffing data are simulated, because no public dataset covers them. Forecasts were tested by replay on historical filings. The next-quarter loss-direction model passed its acceptance rule. A random forest reached an AUC of 0.926 at horizon 1 and 0.920 at horizon 2, against 0.868 and 0.862 for logistic regression, and 0.826 and 0.813 for last quarter's loss status. The financial-item forecasts did not pass their acceptance rule. The win model on CRM deals did not beat chance. The agent's document reading and revision loop have not been run. The results are a replay on historical filings, not a prospective test. The paper makes no claim that artificial general intelligence exists or is needed.