RAG FOMC Update (March 2nd, 2026)
Current status:
The notebook runs end-to-end with Qwen 0.5B, builds a semantic RAG index over the FOMC corpus, and shows that semantic retrieval improves factual QA compared to no-retrieval and keyword baselines.
Remaining issues:
-
Statement-to-decision label alignment is still sensitive to date mismatches and depends on the chosen merge strategy (release date vs decision date).
-
Citation scoring shows 0% because model outputs rarely match the strict citation regex format, even when grounded.
-
Policy classification performance is weak: the model heavily defaults to predicting “Hold” for most statements, indicating class bias and insufficient signal extraction from short excerpts.
-
Differencing remains shallow and does not reliably link textual changes to policy shifts.