The financial crisis since 2008 shows how important it is for bankers and investors to have a credit analysis system as highly developed as possible. One cannot ignore the Lehman case [1], a symbolic illustration of credit risk in which warning signs appeared as early as spring that year, before the banks started cutting their lines of credit on this compensation, when the day of their sudden bankruptcy arrived 15 September. Every institution exposed on Lehman who detected this sign, would have had almost four months to take safety measures: cancellation of credit authorisation, security transfers, contract terminations, CDS [2], conclusions, position netting, etc...
Every institution exposed on Lehman who detected this sign, would have had almost four months to take safety measures: cancellation of credit authorisation, security transfers, contract terminations, CDS conclusions, position netting, etc…
Bruno MATHIS & Jean DELAHOUSSE
Lehman is not the only example. The fall of Enron at the end of 2001, of Parmelat in 2003 and more recently, of MF Global [3] were also preceded by several warning signs.
The first important sign is the discovery of discrepancies in the numbers published or communicated by a company, thereby causing doubt concerning the sincerity of their book-keeping. This leads to the questioning of any analysis that was carried out solely on the base of quantified data..
The second sign is short selling, a sign of distrust. It suggests that all useful information was not made public and is therefore not considered in stock prices. This sign is even more believable since, according to a study [4], the seller of short sales better interprets the information available on a company.
- Jean DELAHOUSSE
Naturally, the style of this exercise is easy to do afterwards [5]. The raid of the social media these last years, through an increase in the volume [6] and in the number of information sources, have changed things in a fundamental way. For example, something as simple as that of tweet from one?s mobile phone during a conference of analysts. Let us add that national regulations are more likely to protect the" whistleblower"; the Dodd-Franck law for example that was passed in 2010 in the United States suggest a financial interest for anyone who denounces criminal acts within a company that would subsequently lead to a judicial penalty [7] ; such measures improve the possibility of finding failure alerts in digital media.
It is necessary to analyse the content before analysing the source, so according to the company Proxem, semantic web specialist, “the analysis of the week of September 14th, 2011 concerning a major French bank shows that the most tweets systematically associate the name of the bank with the fall of the share and with the term ’bankruptcy’. These 600 tweets were not written by human beings, but transmitted automatically by a robotic software which manages about twenty Twitter accounts...”
Bruno MATHIS & Jean DELAHOUSSE
For the Risk Management department in a financial institution, to encourage credit analysis from different and non structured information is a true challenge. It consists of identifying the signs announcing failure among the chaos of the blogoshpere and twittosphere, of identifying the source, analysing the content, verifying the veracity and confirming the stakes, all in a short space of time as it is necessary to take safety measures before the rumour becomes public on the market.
We will begin by defining a list of compensations that would be followed on the web. They have to be notorious, otherwise they won?t have neither the “buzz” nor verifiable information, and they have to represent a financial issue.
A second stage of the process is to list the types of events that one tries to anticipate: one is not only interested in the assumption of bankruptcy, but also in that of a sudden fall in shares, of a publication of exceptional losses... [8]…
For each type of event researched, we will draw up a list of possible warning signs: questioning the sincerity of accounts, resignation of a financial director, announcement of short selling or the conclusion of a CDS... certain signals can be indirect: for example, a rise in the loan position on a security published by a professional informative site such as Data Explorers, reveals a short sales movement, even if these sellers chose not to make their publications public; other signals are specific to an economic sector: the late delivery of the A380 was a warning sign of the sudden fall of EADS shares. For each signal type, we try to identify their possible communication formats: regulated information sites, specialised press, financial blogs and forums, social networks, Twitter. If they consist of written or oral commentary, we will specify how the signals can be expressed in each language.
Next, one needs to define the perimeter of the observed sources, a delicate exercise: financial institutions, the big state companies are the object of direct or indirect comments made by hundreds of thousands of sources; these sources who have a specific target evolve with time, certain fade out or become less pertinent. New bloggers, new media specialists, new twitter feeds appear each day and give out information and comments on the followed counter-parties; when a new event takes place, new sources appear, in particular the “whistle-blowers” : in the case of Lehman, the first signal was found by an independent auditor [9] on her blog.
- Bruno MATHIS
A significant part of the work which must be automated, is to continually look for new sources that appear for a given target, to identify them, qualify them and evaluate their reliability. Different techniques can be used to identify these sources, in particular, we can automate the search in general search motors, to see if the sources start to publish information containing the signals researched. Then, using a counter-party that is the object of the signals, one could search among social networks like LinkedIn or Facebook for sources that are directly or indirectly related, and provide elements that either confirm or invalidate the information. One would then organise on the first assisted level of identification, continual management of the sources list, sources which are generally likely to speak about all counter-parties, or specific sources able to give information on a single counter-party.
Signal research in the content transmitted by sources must be done by software. Several software types can be used:
collection software (“crawl”) which finds all editorial content among blogs belonging to a selection
a text analysis software, also known as “text mining” which finds explicit information in this selection such as “we emit the reserves on accounts published by the company X” , “M. X is named financial director of the company Y, and replacing M. Z”, all this is able to be carried out in each language selected. This type of software is already a necessity in some sectors such as scientific and legal editorial sector and life sciences sector.
the coupling of statistic and linguistic strategies, to detect a rise in the volume of negative messages about a counter-party. The graph below illustrates “emotional analysis” in a trimester in the company Societe
Generale. Until now, this technique is especially used for image analysis...
graph analysis software which searches among social networks and detects changes in the functions or relations between those who are linked to the followed counter-parties.
- Daily number of negative and positive documents on SOCIETE GENERALE SA
Once the signals have been identified, withdrawn and formalised, one must compare them with the other signals for the same counter-party and the same predictable event type and to organise them temporarily, identify their source and rate their pertinence. One must then find the redundancies as signals repeat themselves according to their transmission from source to source.
At this stage, human expertise becomes a dominating factor; the expert must be able to respond quickly to several questions:
Is the information about a default risk of the counter-party already known and being dealt with?
Does the information come from a reliable and independent source or could it be due to manipulation?
Is the information based on enough facts, true or false, and enough sources so that the risk on the counterparty is taken seriously by the market?
Is the information based on real facts, and if so, does it concern the financial strength of the counter-party?
Apart from the mode of operation, one has to define an organisation.
This method indeed requires varied skills, which extend beyond those which can be found within the Risk Management department today:
Digital media specialists, to define and adjust the scope and the techniques of the search on a daily basis
Linguists, to develop various paraphrases of the same concept, in various languages
Analysts, to handle the alerts that are raised
Information experts, to identify the origin of false or misleading information
Communications managers, to broadcast the information, in a sensitive manner, with the persons in house.
In conclusion, this autumn 2011 credit risk awareness is higher than ever, even though traditional tools such as agency ratings and probability models have shown their limits, each major financial institution should consider the potential for exploitation of information conveyed by numeric media and the possibilities that technology offers of processing this material.
