RatingsXpress Reflecting on Automation

In the past year, we have noticed a surge in client demand for digitized solutions. As we continue to revamp our product line, we reflected on the role further automation plays in credit risk modeling and management – and how can we design our products to make it easier for machines, and machine learning algorithm, to perform the tasks of humans at a similar level of accuracy as humans, with a higher speed and scalability that automation brings.

Here we present a few questions that frequently get asked in our client and product design conversations, and some of our observations.

Question #1: To automate or not to automate - Is that the question?

When we first launched the digitized equivalents of our desktop offerings, such as RatingsXpress® (RX) Research on Xpressfeed, which provides bulk access to credit research from S&P Global Ratings for textual data analysis, we were frequently asked if these new offerings could completely replace the web-based platforms (i.e. will bots completely replace humans?).

For processes which the output still needs to be processed by humans, web-based and digitized approaches complement each other.

RatingsDirect®, the official desktop offering for S&P Global Ratings credit ratings and research, presents information in a way that human users can deep dive to uncover insights behind the data with visualization and other analytical tools.  It gives you the “why” behind data points.

On the other hand, the machine-readable RatingsXpress Research on Xpressfeed enables users to:

• Analyze (skim) large volumes of articles
• Create alpha “signals” for investments, and custom metrics or benchmarks
• Identify emerging themes affecting credit, and track spillover. In our recent blog, we tracked how the mentions of “COVID-19” language in earnings call transcripts and S&P Global Ratings reports rippled across geographies and supply chains.
• Provide an early warning indicator via custom filters on key words

In other processes where subjective judgement is not required (e.g., manual transfer of data), machines could replace tasks done by humans. For example, we developed CreditPro® on XpressAPI, which allows users to automatically embed report output into their systems via JSON API. This means users do not have to manually transfer high volumes of report output into their risk systems.  Screenshots comparing the user input from web user interface (UI) vs. XpressAPI are shown in figure 1.

Figure 1: Platform input (left) vs. request body for CreditPro on XpressAPI (right)

^{Source: CreditPro® on XpressAPI programmatic access, S&P Global Market Intelligence. As of July 23, 2021. For illustrative purposes only.}

Question #2: Addressing resistance to change: How do we get users comfortable to migrate from web-based platforms to fully automated processes?

Very early in our product development cycle, we discovered the need for an intermediate solution for clients. Some users simply want a more flexible way to run batch reports because each client slices and dices the reports differently and creating a custom build for each client would not be viable on the UI.  Hence, developed an intermediate CreditPro on API Drive solution, which allows the end user to directly interact with the backend calculation procedures. They modify and append a simple *.txt input file, which is uploaded via Secure File Transfer Protocol (SFTP) into our servers and receive all the reports in their API drive. Screenshots comparing the user input from web UI vs. XpressAPI to the text filed loaded into SFTP via API Drive is shown in figure 2. 

Figure 2: Platform input (left) vs. input text file for CreditPro on API Drive (right)

^{Source: CreditPro® on API Drive, S&P Global Market Intelligence. As of July 23, 2021. For illustrative purposes only.}

In parallel, we released the product on Swagger UI,^[1] which has a look and feel of a web-based platform and provides JSON-API requests and responses. (See figure 3)

Figure 3: Swagger UI for CreditPro on XpressAPI (Programmatic Access)

^{Source: CreditPro® on XpressAPI programmatic access, S&P Global Market Intelligence. As of July 23, 2021. For illustrative purposes only.}

Question #3: What features do end users want to see in their digitized offerings?

In addition to ease of use mentioned above, timeliness and completeness are important. We also uncovered other features in our client outreach:

• Consistent and standardized metrics
  • Since automated algorithms do not use subjective judgment, our metrics must be developed with consistency in mind. For example, the Scores and Factors database, which provides the underlying business, financial, industry and economic risk factors and assessments for S&P Global Ratings credit ratings, was developed to allow end users to run peer comparisons compare, for example, leverage across different industries and geographies.
• Clear organization of information for easy retrieval
  • Whilst humans know how to “skim and skip” over sections (i.e. assign different levels of importance to different sections in an article), machines tend to assign equal weight to all the paragraphs read unless there’s a way to categorize textural data according to similar content and level of importance (e.g., Outlook, Creditwatch, etc.)
  • Furthermore, it also allows machines to read these (“outlook”) sections across entities in the industry, to form an aggregated “outlook” of that industry
• Linking and integration with other content sets
  • Credit risk assessments and modeling often require multiple content sets. One of the benefits of hosting these within the same data delivery platform is the convenience of accessing all data in one place.
  • To enhance this experience further, we have built in additional common identifiers (e.g. ciqCompanyid) and cross reference capabilities to link data (e.g. Loss Given Default) to other inputs into models (e.g. financial statement data available in Market Intelligence).  See Figure 4 on the schema and sample data for US LossStats on Xpressfeed

Figure 4: CreditPro on Xpressfeed (US LossStats)

^{Source: CreditPro® on XpressAPI programmatic access, S&P Global Market Intelligence. As of July 23, 2021. For illustrative purposes only.}

Question #4: Is more data always better? The surprising answer.

During our client outreach, we were pleasantly surprised to find clients say that low coverage does not necessarily translate to low value.

The incremental value of a dataset can depend on:

• The additional explanatory power of this data over what a client already has on hand (i.e. “unique datasets”)
  • This is measured by the correlations of this new content set vs. the client’s existing data.  This also means you can have two clients with the same objective attaching different value to the same dataset.
  • There are also cases where a client finds value from different proxies. For example, economic risk and country risk indicators offered with a credit risk point of view (e.g., our Country Risk (CICRA) and Banking Industry and Insurance Industry Risk (BICRA) indicators) could complement similar indicators provided by economists and those with geo-location data.
• The use cases:
  • There is appetite for having additional early warning signals for downgrades and negative filters, even with lower coverage.
  • Smaller datasets (e.g., our ESG evaluations) are still useful to flag companies with high environmental risks despite lower coverage than our more comprehensive datasets.
  • For risk modeling, a simpler model with fewer factors may be easier to maintain unless there’s clear proven additive value from onboarding a new dataset.
  • It also means that when a client has a new project in mind, a dataset which is previously evaluated to be not valuable can be of value.