Overview / Usage

Instead of studying all credit derivatives, we focus on Credit Default Swap (CDS), one of the popular credit derivatives that were considered the culprit of the 2007-2008 financial crisis. A credit default swap is a particular type of swap designed to transfer the credit exposure of fixed income products between two or more parties. In a credit default swap, the buyer of the swap makes payments to the swaps seller up until the maturity date of a contract. In return, the seller agrees that, in the event that the debt issuer defaults or experiences another credit event, the seller will pay the buyer the securitys premium as well as all interest payments that would have been paid between that time and the securitys maturity date.

CDS is traded over-the-counter, thus there exists little public information on its trading activities for the outside investors. However, such information is valuable. CDS is designed as a hedging tool that the buyers use to protect themselves from potential default events of the reference entity. Besides, it is also used for speculation and liquidity management especially during a crisis.

Before SEC has requested more frequent and detailed fund holdings reporting at the end of 2016, mutual funds filed the forms in discrepant formats. This made it extremely difficult to effectively extract information from the reports for carrying out further analysis. There exist some previous studies that explored how mutual funds have made use of CDS (Adam and Guttler, 2015, Jiang and Zhu, 2016), but only examined a fraction of institutions over a short period of time. In this project, we aim to extract as much CDS-related information as possible from all the filings available to date to enable more thorough downstream analysis. This information appears not only in the form of charts but also in words, thus Natural Language Processing (NLP) is the key.

Methodology / Approach

We first crawl the US Securities and Exchange Commission website to get the reports of all the Mutual Funds from 2003- 2016. We do this using Python scripts with web crawling frameworks such as Beautiful Soup and Scrapy with Selenium drivers. We also take some open source crawlers such as SEC Edgar crawler which can be used to download all the companys periodic reports filings and forms from the Edgar database.
We then use Perl and regular expressions to extract data from the reports downloaded. Perl is useful and better than Python in this because:

1. It borrows its syntax from C and other UNIX commands like awk, sed etc. due to which it has powerful and built in regex support without importing any third-party modules. While doing string and regex operations like replacement, matching, Python is outperformed by Perl since Perl can achieve the same result as Python in a single
   3
   line. Exception handling and many file I/Os are also done much faster on Perl.
2. Perl can handle OS operations using built-in functions. On the other hand Python has third-party libraries for both the operations i.e. re for regex , sys for os operations which need to be ensured before doing such operations.
   We will be using Natural Language Processing with Recurrent Neural Networks to process the complex sentences in the extracted data. The best way to represent text meaning is through words. This is why we use word embeddings which lets us represent words in the form of vectors. The aim of this is that words in a similar context or similar words are close together in the vector space while other words end up far away from each other. We will be using word2vec and glove techniques to determine these embeddings.
   We represent each word through a unique index mapping. We will have to learn the representation of V distinct words in the form of some D dimensional vector each. Since our vocabulary size is V and we need a D dimensional representation of each word, we need a projection layer consisting of a V*D matrix. Both Glove and word2vec are just 2 different methods to generate the projection layer. Word Embeddings are an unsupervised technique and hence are a very useful starting point for implementing neural networks using Natural Language Processing.
   Once we start embedding words, it is useful to use Recurrent Neural Networks because it makes use of sequential information. RNNs have a memory which is used to capture information about what has been done so far. We will be further conducting time series analysis using LSTM-RNN to forecast the buying/selling activities of Credit Default Swaps by Mutual Funds.

Repository

https://github.com/sudhamstarun/Understanding-Financial-Reports-using-Natural-Language-Processing

Other links

• https://i.cs.hku.hk/fyp/2018/fyp18067/