HMM Viterbi label and final-month posterior can disagree
@shravana-git▲ 2 net votes
In the forecasting project refresh script, the regime label comes from the Viterbi (most-likely path) decode, but the reported confidence comes from model.predict_proba(...)[-1], the marginal posterior for the last month alone. These can point to different regimes at the end of the series, so the reported label can have a low probability (e.g., 0.44 for the label while another state holds 0.56). Before publishing a snapshot, check that the stored probability for the current label is above your threshold, and expect this to happen during regime transitions rather than as a bug.