Non-bank sentiment in central bank speeches worsens after crisis
ECB Paper

Non-bank sentiment in central bank speeches worsens after crisis

European Central Bank (ECB) speeches express increasingly negative views of non-banks, particularly since the global financial crisis, researchers find. Financial stability concerns and calls for regulation dominate the negative tone.

Training a measure of perception

Ami Dalloul and Anisa Tiza Mimun construct a perception index described as covering 1998–2025, using 2,830 ECB speeches.

They manually label 4,000 sentences, including 1,300 about non-bank financial intermediaries, to train and evaluate classifiers.

GPT-4o-mini identifies 15,017 non-bank references with topic-classification accuracy of 0.93. A fine-tuned Llama3-8B model using three-sentence context achieves a weighted F1 score of 0.80, versus 0.19 before training.

The index subtracts negative from positive sentence counts and divides by total sentence counts.

Alternative aggregation methods and the three best models produce similar dynamics.

The authors offer the index and labeled data upon request; their findings do not represent ECB views.

Regulatory reassurance fades into risk

The index reaches levels near -0.25 during the European sovereign debt crisis.

Positive sentiment returns in 2014–2015 alongside implemented regulatory safeguards, then weakens over 2016–2020 amid contagion and monetary transmission concerns.

Sustainable finance contributes to a positive episode in 2021.

Crypto-assets, decentralized finance, liquidity, leverage and monitoring gaps feature in negative narratives.

Positive references emphasize innovation, market integration, diversified funding and credit supply.

Tone is not risk

The index quantifies communicated perceptions, not non-bank vulnerabilities themselves.

Its labeling rules classify calls for regulation as negative and implemented safeguards as positive.

Model robustness supports consistency, but the measure remains tied to those interpretive choices.

Report an error