finbert-ng-financial

ProsusAI/finbert fine-tuned on Nigerian business-news headlines, labelled for what each headline means for Nigerian business conditions: negative, neutral or positive.

It scores the daily news feed of tobacco-price-intelligence, an independent portfolio project on free-tier infrastructure. That project is not affiliated with any tobacco company and uses synthetic sales data. Its dashboard shows the mean negative probability per day as a "news crisis score".

Labels: 0 negative, 1 neutral, 2 positive

Same three classes and id2label as the base model, so it drops in wherever ProsusAI/finbert is used.

Training data

344 headlines from Nigerian business and news RSS feeds, August and September 2026. Headline text only; no article bodies.

The labels were made by models, not people. Two Claude models labelled the same 383 headlines independently, each blind to the other and to the base model's score, following a written guide. Only the 344 on which they agreed were kept (89.8% agreement, Cohen's kappa 0.787): 52 negative, 245 neutral, 47 positive. So the fine-tune learns the guide as those two models read it. It is not expert human annotation. The labels, guide and agreement statistics are in the source repo under data/labels/ and docs/labelling-guide.md.

Training

One run on a Kaggle T4, 2026-09-26. 80/20 stratified split (random_state=42): 275 training and 69 evaluation headlines. The notebook is notebooks/finbert_transfer_learning.ipynb in the source repo.

Evaluation

Accuracy on the 69 held-out headlines, against the agreed labels:

Model	Accuracy
ProsusAI/finbert (base)	46.4% (32 of 69)
Always predict "neutral"	71.0% (49 of 69)
finbert-ng-financial	76.8% (53 of 69)

Read the middle row before the last one. Most headlines are neutral, so a model that never commits scores 71%. The fine-tune beats that by four headlines. Its wide margin over the base model comes largely from neutral headlines: the base model got 37 wrong, and at least 17 of those must be headlines the guide calls neutral, since only 20 are not. With 69 evaluation examples the 95% confidence interval is wide, roughly ±10 points.

Intended use and limits

• Coarse daily aggregates of a news feed, as in the source project. Not for judging individual headlines, and not for trading or investment decisions.
• English headlines about Nigerian business. Outside that domain, expect the base model's behaviour or worse.
• The labels encode one written guide's view of what is good or bad for business. They are not neutral facts.

Licence

The fine-tuned weights inherit the terms of the base model, ProsusAI/finbert.