Colloquium, Trainings, and Events

  • Image for STAT SPEAKS: Colloquium on Statistical Sciences

    STAT SPEAKS: Colloquium on Statistical Sciences

    Last updated: Sept. 24, 2020, 9:24 a.m.

    Speaker: Mr. Nigel Rimando, First Circle

    Date: Oct. 15, 2020, 5 p.m.

    Venue: via ZOOM

    Abstract: Incomplete information is the main challenge when it comes to assessing credit risk for financial companies. For Small to Medium enterprises (SMEs) looking to borrow capital for growing their business, the difficulty to prepare traditional financial documents or collateral is a common problem. As an alternative, the inflow of trade documents such as Purchase Orders and Invoices provide a snapshot to the capability of the company to repay as a potential borrower. In Trade Financing, these documents are the proof of cash flow that establishes the ability of an SME to repay the capital it has borrowed.

    However, the verification of these documents is a manual process for Risk Operations Analysts. Calling the issuer of the trade documents is the most direct way to confirm the legitimacy of these documents but this is usually a time-consuming process and some borrowers usually prefer their buyers not to be contacted. Alternative ways to verify documents to smoothen the process of financing is a common question in operational initiatives. One such alternative is by comparing submitted documents against a database of previously verified documents, which allows analysts to visually investigate existing documents against new ones. Taking advantage of previously verified data, analysts can potentially assess the validity of a document without having to contact the buyers. This process, however, creates a new problem of requiring analysts to manually sift through multiple documents, thus adding another step into the verification workflow. This is prone to human error as visually identifying documents issued by same buyers relies on accurate encoding of labels (i.e. buyer names, document numbers, etc).

    Using a Siamese Convolutional Neural Network, an image similarity model is trained to identify documents that are potentially similar in nature, without the requirement of human-encoded labelling. Similarity in this context means similarity in terms of category, such as which buyer issued the document, as opposed to pixel by pixel equality. The model takes in two query images and provides a similarity score. The model achieved a training data accuracy of 99.2% and a non-augmented test data accuracy of 100%.

    Downstream applications for this capability involve building an image search query for analysts. Integrating the search query into the verification workflow of analysts can improve the speed as to which analysts can assess documents, such as for identifying altered documents, surfacing duplicates, and document reconciliation, where previously verified documents support automatic verification for incoming documents.

    Assessing Trade Document Image Similarity Using Siamese Convolutional Neural Networks
    by Nigel Rimando, First Circle