Modelling credit risk

The overall aim of this project is to increase the efficiency with which finance is allocated by lenders to companies so increasing the productivity of all stakeholders involved: banks, companies and data companies that may collect such data & use our models.There is an extensive literature on predicting financial distress for large companies.The differences in predictive performances between different algorithms is relatively well known & relatively small. However, the use of completely untried covariates is unknown. This project aims to assess the contribution that one type of completely new information that emanates from the recent growth in vast amounts of social media data can make to predict financial distress of large companies. We test the hypothesis that various aspects of Twitter messages enhance the predictive accuracy of models that may be used to predict financial distress. If sentiments in publicly announced information reflect issues within a company we would expect that, sometime before a company experiences financial distress, the sentiments expressed in its Tweets or other aspects of its tweets like their frequency, length, timing and so on would differ from those issued by companies not so at risk. The model we propose differs from most in the literature.We do not aim to explain the underlying relationship between aspects of a firm's behaviour and the probability of bankruptcy. Rather, we aim to model the perceived probability of bankruptcy given the information the analyst has available at the time a prediction is made. That is, when new relevant information becomes available, the perceived risk may change even if the actual risk had changed before.The project involves 3 work packages. WP1:Create a sample of companies meeting three criteria: 1 financial distress indicators can be convincingly created, 2 Twitter data available; 3.sentiment indicators from Moody's available. WP2: For each company collect annual profit & loss & balance sheet data, daily market price data, corporate governance data, and Twitter data. WP3:Use different classifier algorithms to estimate predictive models for a training sample & test their accuracy on an out of sample-out of time test sample. In the case of logistic regression we will also estimate sample selection models that correct for bias that may occur because only certain companies will produce tweets & so others will be unobserved. The data is available at different frequencies. The measure of distress-bankruptcies-available daily.Twitter data available daily and macroeconomic data available monthly. Financial & corporate governance data available only annually. We will collect Twitter data for a large sample of financially healthy and financially stressed companies weekly for 3 months. Our aim is to estimate a model that relates the probability of bankruptcy to information on certain variables that influence the assessment of this probability as soon as this information is available. This means we can include variables in the model that are of different frequencies. The model will predict risk monthly. We have chosen to incorporate Twitter data rather than other types of social media data into corporate models to avoid the ethical and data protection issues that may arise in the use of social media data relating to individuals as would be necessary to predict the probability of consumer loan defaults. Use of corporate tweets is legally and ethically acceptable because we will not be using data that relates to persons. We will collect annual financial variables from Compustat, market value variables from CRSP, Twitter data from Twitter, macroeconomic variables from the ONS and bankruptcy data from the financial press. The outputs of the project will be a proof of concept & initial model that can be used by data collection & dissemination companies like Experian, lenders including banks & companies themselves where they have in-house teams that assess trade credit risks.