cb-policy-tone (v0)

Classifies a sentence from a Bank of Ghana Monetary Policy Committee statement as hawkish, neutral or dovish. Part of the project github.com/Mbibachris/cb-policy-tone, with a demo at huggingface.co/spaces/Kobichris/policy-tone.

Training

• Base model: ProsusAI/finbert, fine-tuned for 3 classes (dovish, neutral, hawkish).
• Data: sentences from Bank of Ghana MPC statements, 2003-2026, split by date (train up to 2018, validation 2019-2021, test 2022-2026). Sentences announcing the rate decision are excluded.
• Labels: weak labels from a rule-based lexicon written by the author from a published codebook. The model therefore learns to imitate the rules.
• Class-weighted cross-entropy; 4 epochs; max length 128.

Results so far

• Agreement with the rule labels (macro-F1): 0.74 on validation, 0.78 on the 2022-2026 test years. This is agreement with the rules, not accuracy.
• Accuracy against hand-labelled sentences: pending (a blind 400-sentence gold set is being labelled).
• Statement-level tone is associated with the next policy-rate decision in-sample, but does not improve out-of-sample forecasts beyond past decisions and inflation.

Limitations

• Trained on one central bank's English statements; other banks or genres are untested.
• Inherits the lexicon's blind spots, e.g. rising inflation reads as hawkish even when inflation is low.
• Research prototype; not financial advice.

Use

from transformers import pipeline
clf = pipeline("text-classification", model="Kobichris/cb-policy-tone", top_k=None)
clf("Inflation pressures have intensified and risks are tilted to the upside.")

In the browser (transformers.js): pipeline("text-classification", "Kobichris/cb-policy-tone", { dtype: "fp16" }) uses onnx/model_fp16.onnx (16-bit, about 220 MB). 8-bit versions were tested and rejected: they gave the same label as the full model on only about 84% of 500 test sentences.

Author: Christopher Mbiba.