<p>Nhaufinance: A Domain-Adapted NLP Benchmark for African Central Bank Discourse and Sovereign Risk Analysis</p>

African central bank communications remain analytically inaccessible to standard financial NLP tools: ProsusAI/FinBERT scores zero F1 on CURRENCY entity recognition for African monetary terminology (ZiG, RTGS dollar, cedi), confirming a vocabulary-level domain gap that motivates the present work.

We introduce NhauFinance, the first domain-adapted language model and annotated benchmark dataset targeting African central bank discourse. NhauFinance-v2 is pre-trained on 40 million tokens of policy text from five African monetary authorities: the Central Bank of Nigeria (CBN), South African Reserve Bank (SARB), Central Bank of Kenya (CBK), Reserve Bank of Zimbabwe (RBZ), and Bank of Ghana (BOG), evaluated on monetary policy sentiment classification, financial distress detection, policy event classification, and named entity recognition. Annotations were produced via an LLM-assisted expert simulation protocol spanning emerging-market practitioner, quantitative research, academic, and central bank validator perspectives across three labelling rounds yielding 1,876 annotated examples. Independent human validation of a stratified subset is in progress and will be reported in the journal submission prior to peer review.

The primary finding is a data efficiency advantage of 6× on financial distress detection: NhauFinance-v2 achieves 0.62 macro F1 with 50 labelled examples, a level FinBERT requires approximately 300 examples to match. On policy event classification, NhauFinance-v2 achieves 0.5491 macro F1 versus FinBERT's 0.4713 (Δ = +0.0778); on distress detection, 0.7209 versus 0.6790 (Δ = +0.0420). These two tasks represent the core policy signal extraction use case for which the model was designed, with applications to emerging-market sovereign risk analysis and contagion early warning systems.

Analysis of annotation patterns reveals that 96.2% of African CB policy event announcements are embedded in institutional survival framing, a structural feature of African sovereign communication absent from developed-market NLP frameworks, with direct implications for how EM sovereign risk signals should be extracted from African CB text.

We document two limitations with full statistical transparency. Sentiment classification exhibits initialisation sensitivity attributable to cold-start classifier adaptation on small datasets, characterised via a 7-seed robustness analysis; NhauFinance-v2 shows 51% lower cross-seed variance on NER, confirming that domain pre-training stabilises token-level recognition. CURRENCY_ACTION and INTERVENTION classes contain insufficient training examples (17 and 10 respectively) for reliable current estimates; expanded annotation is targeted for subsequent benchmark rounds.

The NhauFinance corpus, annotation guidelines, and model weights are released as the first open benchmark for African central bank NLP, providing a reproducible evaluation framework for emerging-market monetary policy text analysis.