ECB Perception of Non-Banks: A Machine Learning-Based Approach (European Central Bank)

Summary

• The authors construct an index tracking the European Central Bank's perception of non-bank financial intermediaries using 2,830 speeches delivered between July 1998 and June 2024.p.8
• The entire speech corpus comprises around 315,400 sentences, from which 4,000 sentences were randomly drawn for manual benchmark labeling.p.9
• A zero-shot classification with GPT-4o-mini achieved an accuracy rate of 0.93 for topic identification, flagging 15,017 sentences as relevant to the non-bank sector.p.13
• Out of the 4,000 manually labeled sentences, 1,300 specifically addressed the non-bank financial intermediary sector and served as the fine-tuning dataset.p.13
• The yearly perception index displays an overall downward trend, falling to low levels such as -0.25 during the European sovereign debt crisis.p.17

Key insights

1. 1. The ECB's publicly communicated perception of non-bank financial intermediaries has been predominantly negative and has declined over time, turning almost exclusively negative after the Global Financial Crisis.p.5
2. 2. Negative sentiment in ECB communication is primarily driven by concerns regarding financial stability, liquidity, leverage, bank-non-bank interconnectedness, and calls for tighter regulation.p.5
3. 3. Positive sentiment occurs during periods when the ECB highlights the contributions of non-banks to financial innovation, diversification, market functioning, and real-economy credit supply.p.6
4. 4. A distinct positive rebound in the index occurred around 2014-2015, corresponding to the practical rollout and implementation of post-crisis non-bank regulatory frameworks.p.20
5. 5. From 2022 to 2024, negative perception was heavily influenced by systemic concerns surrounding decentralized finance, crypto-assets, and ongoing non-bank data gaps.p.21
6. 6. Standard off-the-shelf sentiment models performed poorly on central bank speeches because central bank tone depends on nuanced domain-specific context rather than typical financial news vocabulary.p.14
7. 7. Applying up-sampling strategies to minority positive and neutral classes did not improve sentiment classification performance and in several instances led to overfitting.p.25