​Deep NLP Learning NewsTrack Analytics

​How NewsTrack uses newsfeeds in real-time to generate state-of-the-art analysis and forecasting

about NewsTrack Analytics

NewsTrack is a leading-edge NLP deep learning toolkit that processes unstructured and high-frequency textual information to generate risk related signals at real time

This model can be implemented as an overlay/early warning and monitoring tool element for risk-related events

The model is based on a fully flexible development pipeline such that it can be applied to new use cases with only little effort

The NewsTrack toolkit can be used across a variety of news sources, NLP/transformer models and other languages

NewsTrack can process both structured and unstructured data at the same time and can efficiently triangulate across heterogeneous data sources to maximize information gain

NewsTrack Analytics has a striking impact on efficiency, accuracy, and objectivity, where benefits are realized in real-time

Efficiency

Accuracy

Objectivity

10% loss reduction for the loan book of a leading financial institution

200+bps performance uplift for the investment book of a large insurer as well as a pension fund

+50% of credit events predicted at least 30days ahead, including market-related events such as bond-price drops and CDS spikes

Anticipation of “singular” events based on unstructured e.g. WireCard, Greensill, etc.

Various testimonials from leading clients globally confirming the value add

publication

Natural language processing and transformer models for credit risk

News feeds are factored into models to predict credit events by Emanuel Eckrich, Phillip Escott, Rainer Glaser and Christoph Zeiner.

Read the article 

*To read all article you need to log in to Risk.net

Selected applications

Prediction of rating changes / defaults

Prediction of rating changes and defaults based on non-qualitative IRB model components

NewsTrack shows significant predictive power when tested against observed defaults and rating

Events, improving predictive power of non-qualitative IRB model components

News articles

News analytics model

Rating event prediction

70% Recall vs. rating downgrades
Sector / country level early warning system

Early warning system for economic and political risks

Newstrack allows for early warning regarding economic and political risks at sector and geography level

Current and historical risk scores

Newsfeed content, weighted by early warning model

Comprehensive real-time dashboard

COVID-19 Impact

COVID-19 impact on credit customers

Newstrack uses state-of-the-art leaning algorithms for superior predictions of credit risk signals, e.g. COVID-19 induced downgrades

70% Recall
Predicting Market Price Movements

Prediction of short-term price fluctuations based on news articles

Newstrack can predict market prices efficiently and support optimal investment decisions

Newsfeed content, weighted by machine learning model

Fundamental values as additional cross-check

Newstrack Price Predictor for price evolution

Know-your-Customer / Financial Crime

Robust and effective controls for customer due diligence

NewsTrack offers a data-based approach to effectively ensure KYC / customer due diligence compliance via intelligent triggers and thresholds 

Increased accuracy

Streamlined, quicker  operations

Continuous optimisation

Lower cost

Improved document management

Affordability

Affordability estimation in Retail

Newstrack can efficiently triangulate all available information – both structured and unstructured to estimate and monitor client affordability

Historical Behavioural data

Account movements

Payment purpose/ counterpart

Interaction log and loan documentation

Specific information (e.g. external reference data, indices)

Client testimonials and references

We are happy to provide dedicated client references and testimonials – please reach out to us under NewsAnalytics@oliverwyman.com

Send e-mail

How NewsTrack works in application

Retrieve and pre-process news articles by a global news provider

Use state-of-the art transformer models (BERT) to encode texts into a machine readable vectors

Supply the vectors into classic machine learning models

Use model predictions with confirmed performance across a variety of use cases

Modularity and flexibility

NewsTrack analytics was built to allow for maximum flexibility – providing a powerful tool for a broad range of applications

Only minor adjustments are necessary to connect model to other news sources and providers, including languages other than English

Code base is fully portable (Python) and was developed / tested using client IT infrastructure

Can be easily enhanced to integrate other transformer models (e.g. XLNet) and languages (e.g. German)

The model was designed to directly learn from large datasets – expert labelling is no longer required

Can be trained for a variety of classification tasks and corresponding machine learning algorithms

Flexibility contains universal and covering various use cases, agnostic of language, platforms, counterparty types, data sources, etc., and clients and users across sectors, for example Financial Services, Automotive, Energy and Commodity, HLS, supervisory authorities.

01. Client types and Sectors

Financial Services, HLS, Automotive, Supervisory authorities, Energy and Commodities

02. Business Use / Use Cases

Credit (event) prediction, COVID-19 impact and vulnera, Early Warning and Monitoring

03. Economic and system environment 

Languages, Platforms, Input data, Data sources 

team

contact 

If you have any questions or comments regarding NewsAnalytics, please contact us! 

Send email