Computational Assistance in Risk Perception: A UCL BIT seminar
In his book “Black Swan” Nassim Nicholas Taleb talks about highly improbable, and therefore unexpected, events that have an enormous impact on the world, beyond what is expected in history, science, finance and technology. It is a theory taken from the archives when Covid-19 came upon us earlier this year, as Taleb cites pandemics as an example to illustrate Black Swans. He is convinced that events like Covid-19 are unpredictable:
1. because scientific methods cannot detect such highly unlikely events
2. because of the psychological biases that blind people, both individually and collectively, to uncertainty and the massive role of rare events in history.
Our latest speaker, Dr. Jochen Leidner, research director at Refinitiv, the former finance and risk department of Thomson Reuters, spoke about a risk-mining instrument he invented based on the Taleb’s Black Swan theory.
According to Dr. Leidner “risks lurk everywhere in the world, in our private lives and in our professional career. Things can go wrong geopolitically as we experience this year, perhaps more strongly than in the last couple of decades.” Especially at the moment “we can sense that things are not running smoothly, companies are going bankrupt and devastating events are happening all around the world, in some more than others.”
The question that arises is whether those events could have been anticipated and even more importantly prevented?
The reasons us humans are quite bad at calculating risks is because “we are not very good at dealing with probabilities. Especially not when they are unusual, apart from the norm and very small. Although a large amount of very small probabilities can actually add up to something that is quite a realistic probability mass. This means, bad things do happen for a reason after all.”
Dr. Leidner is therefore convinced that every company’s “ risk register”(a set of all the risks that an entity like a company or even a person or a project is exposed to) should include what will be done when each of the risks that might happen actually happens. “Risk is pervasive, and company risks, in particular, are valuable pieces of information”. Dr. Leidner and his team therefore set themselves the goal to automatically create or extract a risk register. They attempted to use computational help in making up such risk registers, to make them comprehensive, so they would include things that are not so common and therefore very unlikely to happen.
What they came up with “is a process where an inventory of what the risks are actually called is established.” It’s important to remember that “new risks come out all the time, and sometimes these risks have a new name, for example COVID-19 or Brexit. With regard to language, you have to consider that these words did not even exist because the things that they were referring to were unheard of, as they are a new phenomenon in the world. The people who discovered them made up new names for them. And that means when you computationally want to analyse risks, you can’t actually start with a close list, you need to create an open and organic process that is capable of learning the name of risky things of riskiness itself how it’s expressed in human language.”
“All of these risks and words are learned from the web using a way of remote controlling Google-type search engines, by giving them certain linguistic patterns that extract pairs of words where one of them is a parent and the other one is a sort of a son or daughter node. Financial risk is a super type of bankruptcy and financial risk itself is a child of general risk, which might be the root node at the very top.” This is how they managed to create this automatic taxonomy of around 10,000 nodes of hierarchy consisting of words and phrases that describe some form of risk.
“The first insight was to make this quite tangible and to link it to companies, and therefore turn it into something that is essentially actionable intelligence. At [Refinitiv] we have a tool that finds pieces of information or company names. Now we have looked up potential risk terms and company names in the same sentence. And that means we can feed sentences that can contain both a candidate risk to bankruptcy and company names.”
For example, this risk relationship classifier analyses the syntax or grammar of each sentence to find out whether, for example, bankruptcy and a company have been mentioned together in a sentence by a journalist. This machine learning algorithm will then output a number that indicates whether a risk mention was found.
This information is then stored in a database and can be used for various purposes such as sending alerts to people who care about this specific company. It can also be used to search for specific risks, such as oil spills. What you get is a computer program that can run over a year’s worth of news and then as a result, it gives you an enormous amount of detail in terms of the list of risks that have been found.
At the end of the day, Leidner says the “software doesn’t do anything magical; it still requires human journalists to talk about the risk. It uses the journalists as a crowdsourcing sensor, or as a social sensor, you might say. But the nice thing is you will not miss it. The thing about human brains is they get tired. They are very slow. We are not consistently performing at a certain level. And the reason for that is just because we are not built to do mechanical tasks. Instead, we are built to operate flexibly in very quickly changing environments. That is what us humans are extremely good at and machines thus far are not. They’re getting better, thanks to machine learning; but they’re still fairly rigid.”
Therefore, they wanted to create something more like risk glasses or a risk filter to help human analysts to no longer overlook risks. Leidner hopes that his invention will be used primarily by risk analysts to detect all kinds of unlikely events that could have potentially serious negative consequences. The future of this work seems promising, and may soon help us calculate and account for even the most unlikely risks, far more successfully.
Thank you Dr. Leidner for a fascinating talk. Stay tuned for more exciting events from the UCL Behavioural Innovations Team!
Twitter: @UCLBIT
FB: UCL Behavioural Innovations Team
LinkedIn: UCL Behavioural Innovations Team
Instagram: @ucl.bit
