Advertisment

Changing the scope of financial services with natural language processing

There are numerous advantages to be gained from the added insight and efficiency that natural language processing can bring to the lender and the borrower

author-image
DQINDIA Online
New Update
Accenture

When a business applies for a loan, the management team generally submits a business plan including a spreadsheet with the business projections. In order for us to analyze the data, it is first necessary to map items in the borrower’s financials to those in our common model so that a like-for-like comparison with other cases and outputs in a common format is possible. This problem is a classic application of natural language processing (NLP) that can be bootstrapped by a skilled human credit analyst providing a glossary of possible mappings in advance.

Advertisment

Efficient document analysis with natural language processing

A lot of time in conventional credit analysis is spent scanning through borrower financials and various contract documents, searching for important information. NLP can help the human analyst to be more efficient in these cases. For example, it can help to identify important statements in financials or clauses such as indemnities and guarantees in contracts. It is a relatively small investment for an analyst to train a classifier to recognize these clauses, and they receive a significant benefit in time savings as the model is able to process documents far faster than a human once trained.

Use of sentiment analysis for credit insights

Advertisment

Sentiment analysis can help to summarise qualitative text data into scores that are more readily included in other analysis. Here we present an example of a restaurant chain, comparing review scores on a popular rating site with the sentiment of text in the reviews. The size of the bubble indicates the number of reviews in the sample. Reviews of restaurants on the diagonal of this plot are normal (a good score corresponds to a positive sentiment and poor scores also correspond to more negative sentiment). It is more interesting to note the restaurants where the score is positive but the sentiment of textless so or vice versa. Our most anomalous example (labeled ‘R1’ in Figure 1) is a restaurant with few reviews, so it may be that the score has been unduly affected by a few negative reviews and will revert to the diagonal as more data points are added. Repeating this analysis for multiple review sites and for a peer set including this chain’s competitors would give a sense of the external perception of the quality of these restaurants and how they compare with peers. This would be important to the understanding of the credit case and allows data-driven validation of the set of comparable businesses used for analysis.

Artifical intelligence

Figure 1: Online reviews of a restaurant chain: ratings versus sentiment of text

Advertisment

Use of recommendation systems for comparable analysis

An important task in the analysis of any company is choosing a peer set of comparable companies against which to benchmark the financial and operating data in the plan. This enables the analyst to assess whether the borrower’s projections are realistic and compare the theoretical outcomes in the plan with the actual realized financial performance of real companies. In order to do this, we must first determine which companies should be included in the peer set. This is a very nuanced question, requiring a great deal of judgment and expertise from the analyst, but a recommendation system (similar to that which is used by Netflix and others to recommend films on the basis of the user’s responses to the previous viewing) can help with this task. We begin by presenting the analyst with a series of potential comparables (similar businesses that can potentially be used as benchmarks for valuation purposes), ranked according to certain factors in the data. In the case of companies, these factors could include the size of the company, the sector and many other variables. Once the analyst has chosen which companies to include in the report, the system can use this information to adjust the weightings on factors that would rank those companies more highly in future, thus learning overtime to improve its recommendations.

Identification of hidden trends using clustering

Advertisment

Clustering is an umbrella term for a set of techniques that enable a data-driven segmentation of a given data set into a set of categories. As such, this allows the external validation of assumptions that the analyst might have a priori. For example, we were performing an analysis of a spirits manufacturer. When looking at sales data for many brands in the industry, it was difficult for the analyst to discern any particular trends. Clustering, however, highlighted two groups of brands that were very different (see Figure 2).

Artifical intelligence

Figure 2: Spirits brand share over time

Advertisment

One was a cluster of brands seeing a significant sales decline, and the other a single brand seeing corresponding sales growth. This was related to a demographic change in drinking fashions: young people were changing their habits from one brand to the other. This is an interesting example of artificial intelligence actually making data more intuitive and understandable — once the analyst saw the clusters, they were immediately able to explain them.

It is clear that while we may never achieve100 percent automation of complex credit analysis through the use of artificial intelligence and machine learning, doing so may not be a worthy goal. Rather, there are numerous advantages to be gained from the added insight and efficiency that these techniques can bring to the lender and the borrower in the process.

By Sean Hunter, CIO, OakNorth

Advertisment