Financial supervision: SupTech with machine learning in an early warning system

What it is about

A conversation about the research paper “Empowering financial supervision: a SupTech experiment using machine learning in an early warning system” and its conclusions. For financial supervisors, information encoded in text is a valuable complement to the balance sheet data routinely used to assess the soundness of institutions.

The study uses natural language processing (NLP) techniques and network analysis to detect anomalies in the Spanish corporate system and identify idiosyncratic and systemic risks. Sentiment analysis detects anomalies in the behavior of specific companies, while network metrics monitor system risks. In supervisory technology (SupTech), these systems are a proactive tool for financial authorities and can give early warnings of possible financial stress or systemic risks.

Session summary

The speaker presents an exploratory experiment at the Bank of Spain to detect early warnings of risks in the financial system from press news, with three components. The first is entity recognition, which keeps only the news items that refer to the companies of interest (for example, telling a bank from a city with the same name) and is complemented by a language model that gathers all the ways each entity is named. The second is sentiment analysis, using a financial dictionary method and another based on language models pretrained in Spanish, and a traffic-light that triggers when sentiment departs from its average.

The third is graph analysis: the connections between entities that appear together in the news are mapped and centrality measures are computed to see which entities might spread a risk. He shows comparisons with episodes of falling share prices and with a market volatility indicator to validate the tool, and cautions that it is not yet in production and that press bias must be watched. The questions covered the source of the news, how to measure the quality and usefulness of the alerts, stress tests with synthetic news, use in anti-money-laundering and reputational risk, and the possibility of applying it to other sources.

Topics covered

• Press news as a source of early warnings for supervision
• Entity recognition and language models to identify companies
• Sentiment analysis and a traffic-light of alerts
• Graph analysis and the spread of risks between entities
• Validating the tool against share prices and volatility indicators
• Limits: press bias, exploratory phase and lack of deployment
• Questions: data sources, quality of alerts and uses in anti-money-laundering