Regulations pose hardly manageable challenges for banks as more and more voluminous requirements are added to an already extremly complex environment.
Often, the regulations require significant changes in the banks’ methods, processes and/or systems and can have an additional impact on the capital requirements and hence the business model of the banks.
Projects dealing with regulatory issues cost the banks in total several billions of Euros and bind immense expert resources. Furthermore, the timelines for fulfilling the requirements are very tight and require fast decisions.
On the other hand, banks already have enormous knowledge resources for dealing with the challenges: External and internal texts, like the regulatory texts themselves or project documentations and audit findings, contain abundant information about a wide range of issues.
Hence, an efficient analysis of the information can provide substantial enhancements of the banks’ capabilities to deal with regulatory issues. RiskDataScience has already developed respective concepts and tools and is further enhancing them together with partner companies.
The article at hand deals with the exemplary machine learning / natural language processing-based analysis of a new regulatory text with our specially developed tools. The emphasis of the analysis presented lies on speed, since the self-imposed target is to obtain all the results “in 15 minutes” once everything is properly set up.
The analyzed regulation is the “RTS on Procedures for Excluding Third Country NFCS from CVA Risk Charge“. Mainly, it’s about excluding transactions with non-financial counterparties established in a third country from the own funds requirement for credit valuation adjustment risk.
Our 15-minutes analysis starts with the superb Python “summarize” tool which provides short summarizations of texts. In this case the summarization is
EBA final draft Regulatory Technical Standards on the procedures for excluding transactions with non-financial counterparties established in a third country from the own funds requirement for credit valuation adjustment risk under Article 382(5) of Regulation (EU) No 575/2013 (Capital Requirements Regulation Œ CRR) RTS ON PROCEDURES FOR EXCLUDING THIRD COUNTRY NFCS FROM CVA RISK CHARGE 9 EUROPEAN COMMISSION Brussels, XXX [–](2015) XXX draft COMMISSION DELEGATED REGULATION (EU) No –/..
Hmm, sounds familiar. The algorithm apparently has extracted the basic information of the text.
The next step is a cosine similarity comparison with other regulatory texts via the semantic analysis method LSI (Latent Semantic Indexing). The regulatory texts at hand were Basel 2, Basel 3, CRD IV, CRR, and EMIR.
As one should expect, the similarities to CRR and CRD IV should be very high, there should be some similarities to EMIR, few to Basel 3 (due to the CVA context) and nearly none to Basel 2.
These are the results:
- CRD IV: 0.964
- CRR: 0.891
- EMIR: 0.667
- Basel 3: 0.327
- Basel 2: 0.082
Again, as expected.
Now, the speed advantage of automatic analyses comes into play, as the last analysis deals with a sentence-by-sentence comparison. For this, our program splits the RTS text and finds out each CRR sentence with the highest similarity. The – at the first glance plausible – results can be downloaded here as tab-separated csv:
This concludes our analysis – since the timeline is just below 15 minutes.
The method can be enhanced to capture more complex cases and issues, like the identification of the legal basis of inquiries or the examination of “regulatory gap networks”.
We are developing respective methods and tools and support our customers in obtaining an overall perspective of the regulatory data in use.
Dr. Dimitrios Geromichalos
Founder / CEO
RiskDataScience UG (haftungsbeschränkt)
Theresienhöhe 28, 80339 München
Telefon: +4989244407277, Fax: +4989244407001