Expose DynamicFactorMQ EM convergence status
Added an explicit convergence flag to the EM results returned by statsmodels' mixed-frequency dynamic factor model estimator.
statsmodels/statsmodels · #10214
Statistical reliability fix
Automated pipelines can now distinguish genuine EM convergence from a reverted or iteration-limited DynamicFactorMQ fit without parsing warning text.
Problem
When likelihood decreased under the default revert policy, `fit_em` rolled back and reduced its iteration count. The returned result then looked identical to an early successful convergence, and the only failure signal was an easily missed warning.
Approach
Computes convergence from both the termination state and tolerance criterion, then exposes the boolean as `mle_retvals.converged` while preserving all existing result fields, iteration counts, and warning behavior.
Impact and scope
- Lets unattended forecasting and model-selection systems reject numerically unsuccessful fits programmatically.
- Prevents `iter < maxiter` from being misinterpreted as convergence after an EM likelihood rollback.
- Adds a backward-compatible result field without changing estimator inputs or successful-fit behavior.
Validation
- Added regressions for genuine convergence, maximum-iteration termination, and likelihood-decrease reversion with a deceptively low returned iteration count.
- Passed the focused convergence tests, all 59 DynamicFactorMQ module tests, Ruff, and the complete reported CI matrix.
- The implementation commit credits Goutam Adwant; a statsmodels maintainer merged the GitHub-verified commit and GH-10157 closed automatically.