At the first Think Forward Summit last year, participants defined three areas where people experience challenges in their financial decision-making. These challenges gave birth to five projects, which participants were busy working on throughout the past year. This is an overview of the Social media analysis-project, their research and the practical solution being developed based on the research.
Anticipating and avoiding harmful borrowing through social media
By Kristin Glass and Rich Colbaugh
Have you ever borrowed money to buy something and later regretted it? Perhaps you even wondered if you had temporarily lost your mind?
If so, you’re not alone, at least judging by recent data on credit card debt, loan defaults and home foreclosures. Fortunately, help may be on the way. We’re leading a research project with our predictive analytics start-up Periander, aimed at helping people anticipate and avoid making poor borrowing decisions.
Traditional approaches to assisting people with their finances, which emphasise things like improved access to information and increased financial literacy, are often ineffective because decision-making is influenced by emotions and relationships in addition to rational contemplation. Recognising this, we’re trying to understand and quantify these ‘nonrational’ factors by tapping into a novel source of information – social media. Using
social media as a window into what’s going on when a person decides whether to take out a loan or enrol in a savings plan has several advantages over more standard forms of data.
Can I afford this?
For one thing, social media posts are informal and volunteered, which facilitates understanding emotional states and personality types, even in settings that may be embarrassing or distressing. Additionally, this data is cheap, easy to collect, and user-specific, which opens up the possibility to offer affordable, personalised decision-support ‘apps’ in the financial realm.
Another novel aspect of the project is its use of machine learning, the technology that helped the robot Watson win at Jeopardy and drives Google’s driverless cars. Consider the task of instructing a computer to read an individual’s Twitter posts, infer his/her emotional state and personality from this content, and then help decide whether an auto loan is a good idea.
Explicitly programming the computer to perform the intricate and deeply contextual reasoning involved is profoundly difficult. Instead, we present the computer with many examples of Twitter posts and subsequent auto loan outcomes, both good and bad, and let the computer learn for itself what patterns of social media activity are predictive of loan defaults.
Process of avoiding negative borrowing
The combination of social media data and machine learning provide the foundation for a process, depicted below, for helping people foresee and avoid negative borrowing outcomes. In this process, a user’s social media activity is analysed to characterise his/her emotional state, personality profile and social influences, using word patterns and other features. As one example, the machines learning algorithms have discovered that depressed users tend to exhibit excessive focus on a small group of network peers and increased use of first-person pronouns and negative-affect words.
The inferred emotion and personality traits are then used to predict if and when a person is likely to engage in harmful borrowing. For instance, the algorithms reveal that low conscientiousness together with high neuroticism, as estimated from a user’s social media activity, increases the probability that a person will act rashly when distressed.
Finally, the analysis identifies the main drivers of the anticipated detrimental borrowing event and uses these insights to help individuals avoid undesirable outcomes. This last step is being designed and implemented by a separate ‘practical solution’ team (BankNXT). While the project is in its initial stages,early results are encouraging. For example, Twitter-based prediction models have been learned for six categories of poor borrowing outcomes: credit card delinquency, auto loan default, student loan delinquency, peer-to-peer loan default, borrowing at excessive interest rates, and debt in collections. In all cases, the models predict loan outcomes with accuracies approaching or exceeding 90% when using only an individual’s social media activities as the basis for prediction. These models also reveal the main drivers of the undesirable outcomes. Perhaps surprising, emotional state and peer effects in the weeks leading up to a borrowing event are actually more important behaviour-drivers than rational considerations, with key emotions including worry, anger, and unhappiness. In conclusion, this research project suggests that, if you feel you must have temporarily lost your mind to have agreed to that auto loan, you may be right: non-rational factors can override logic and reason, resulting in regrettable – but perhaps avoidable – outcomes.
Just picture this: you are looking online to buy a new couch. You are wondering if you can afford it and want to know if you have enough funds, but also if you will not feel uncomfortable afterwards. This is where the practical solution “Can I afford this?” comes in.
It is a built-in tool in your browser or shopping app that helps you determine whether you can afford it by simply saying “yes” or “no”. Of course you will be provided with an explanation of the outcome. The beauty of this solution is that the answer is combined with a social media data analysis which will give us a very good indication of emotional behaviour towards personal finance management. Research has proven that this is a very reliable predictor of financial delinquency.