TOKYO | The Bank of Japan's Institute for Monetary and Economic Studies hosted its 11th annual Finance Workshop on November 28, 2025, in hybrid format, presenting three research papers on how machine learning and artificial intelligence can sharpen financial analysis. The most immediately relevant finding for markets: general-purpose large language models including ChatGPT, Claude, and Gemini generated measurable excess stock returns when applied to corporate disclosures, while fine-tuned specialist models and traditional dictionary-based methods did not.

---

LLMs Produce Alpha Where Fine-Tuned Models Fall Short

Prof. Katsuhiko Okada and colleagues from Kwansei Gakuin University tested multiple approaches to sentiment analysis against Japanese Annual Securities Reports. The study pitted polarity dictionaries (a standard rulebook approach), a fine-tuned version of BERT (a widely-used language model adapted specifically for the task), and three pre-trained commercial LLMs against the same dataset. Only the pre-trained models produced signals that translated into excess equity returns. Fine-tuned BERT and dictionary methods produced no comparable result.

The implication is counterintuitive for developers who assume domain-specific training always wins. For on-chain analytics platforms such as Nansen and Santiment, which already parse governance forums and social sentiment to generate token intelligence, the finding suggests that layering a general-purpose LLM over raw text may outperform a smaller model trained narrowly on crypto-specific language, if the pattern holds across asset classes. The build-versus-fine-tune question has a clearer answer here, at least for sentiment-driven signal generation.

This finding aligns with broader global evidence on LLMs in institutional finance. An IMF working paper published in June 2025, analyzing 74,882 documents across 169 central banks, found that LLM-based analysis of central bank communications reliably predicts policy rate movements. The BOJ workshop result and the IMF study together point to a consistent pattern: general-purpose language models extract actionable signals from institutional financial text across languages and jurisdictions.

IMES Director-General Shingo Watanabe opened the workshop by setting the technology's progress in context: "Machine learning and AI technologies have been utilized in the field of finance from an early stage, and AI technologies including Large Language Models (LLMs) have made remarkable progress." In later remarks during the workshop discussion, Watanabe also flagged real limits, noting that "various biases that have been reported in LLMs, such as a bias toward generating answers favorable to users," are active concerns for financial applications.

---

Explainability Gets a Computational Upgrade

A second paper from Waseda University professor Junnosuke Shino and Kazuhiro Hiraki of the IMF addressed a persistent problem with ML-based financial models: explaining why they produce a given output. The dominant method currently in use is SHAP (SHapley Additive exPlanations), which assigns credit to each input variable but scales exponentially with model complexity. Shino and Hiraki proposed alternative methods grounded in cooperative game theory that reduce that scaling from exponential to linear while preserving comparable accuracy.

For DeFi credit protocols and institutional risk tools that rely on ML scoring, this is a regulatory consideration as much as a technical one. The EU AI Act identifies explainability as an increasingly important requirement for high-risk AI systems, though practitioners should verify which provisions are currently in force given the Act's phased implementation timeline. Japan and India are developing similar frameworks. Linear-scaling interpretability methods lower the cost of compliance without sacrificing model performance.

---

Foreign Ownership in Bond Markets Has a Breaking Point

The third paper focused on Japanese Government Bond liquidity. Researchers Toshiyuki Sakiyama of IMES and Satoko Kojima used machine learning on granular bond-level transaction data to show that foreign financial institution participation improves JGB market liquidity up to a threshold, then reverses. Beyond a certain share of transactions, liquidity deteriorates.

This finding has direct relevance to bond markets across Sub-Saharan Africa. Sovereign debt markets in Nigeria, Kenya, Ghana, and South Africa face recurring stress tied to foreign investor concentration. The non-linear relationship identified in Japan mirrors documented dynamics in those markets, where sudden foreign outflows create outsized liquidity gaps. The BOJ is sharing these methodologies openly via IMES newsletters, giving capacity-building teams at the Central Bank of Nigeria and the South African Reserve Bank a freely accessible research baseline.

---

Regional Dimension: Non-English Markets Are the Next Frontier

The LLM sentiment paper used Japanese-language filings, which matters beyond Japan. Most general-purpose LLMs were trained primarily on English text, yet the BOJ results show that these models extract useful signals from non-English corporate disclosures. That is a direct proof of concept for regulators and fintech developers in South Asia and Africa working with Hindi, Bengali, Urdu, Tamil, or other regional languages in financial data.

Japan's Financial Services Agency published its AI Discussion Paper (Version 1.0) in March 2025, laying out risk frameworks for AI in regulated finance. Central banks in India and across West and Southern Africa that are building data science units now have both a methodological template from the BOJ and a regulatory template from the FSA to draw on simultaneously.

---

Parallel Tracks: CBDC and Blockchain Sandbox Advance Separately

The workshop sits alongside a separate institutional track at the BOJ. Governor Kazuo Ueda confirmed in early 2026 that the bank is expanding its wholesale blockchain sandbox, with the bank conducting "technical experimentation on settlement using central bank money in the form of current account deposits on a system that uses blockchains." A decision on retail CBDC issuance is expected later this year. The BOJ is also participating in Project Agorá, a multi-central bank initiative targeting tokenized cross-border settlement, which carries downstream relevance for trade finance corridors connecting Japan to East African partners.

The convergence of AI-driven market analysis, CBDC infrastructure, and Japan's active regulatory debate over reclassifying crypto assets under its Financial Instruments and Exchange Act positions the country as one of the more methodically active regulatory environments in digital finance heading into the second half of 2026.

---

Sources: IMES Newsletter (BOJ, February 2026); CoinDesk (March 2026); IMF Working Paper on LLM Central Bank Communication (June 2025); Japan FSA AI Discussion Paper (March 2025); Global Legal Insights Japan 2026.