Interview with Olivier DUCHENNE & Sophie EOM
Conducted by Jooyong JUN
C&S: Please make a summary presentation about Solidware.
Olivier DUCHENNE & Sophie EOM: We (Solidware) build Machine Learning-based predictive models for finance companies. They need a predictive model for their business as they are related to various kinds of risks, such as underwriting, product offering, or customer retention if we name just a few.
Many financial companies already have the enormous size of valuable data. Their risk management system, however, has been mostly based on a simple model, which hinders their ability to fully utilize the data as it is plagued with many human assumptions and biases. As a result, financial firms fail to maximize their revenues while consumers pay more for loans and insurance premiums.
With our machine-learning based data analytics solution, DAVinCI LABS, we analyze clients' data and find the best combination of different machine learning algorithms such as deep learning to generate the most accurate risk prediction as far as possible, without wasting any information in the data. In all, we help to find and minimize risks, and eventually generate significant additional values.
A lot has been talked, but few success stories were observed in machine learning (ML) and FinTech at this point. Which area of finance do you think is best-suited for machine learning? (e.g. credit information discovery, robo-advisor, failure or default prediction), and which is not?
(Sophie) At this point, credit scoring is best-suited for ML. Robo-advisor does not necessarily work well because more often than not the amount of data is insufficient. It is also far harder to predict the market.
How far can ML replace the implicit knowledge that banks have and use in relational banking? What do you think the role of human traders/investors/analysts will be like in the future after the widespread of ML in finance, if there is any?
(Sophie) Humans will be able to spend more time building better "strategies" based on the insights extracted through machine learning, rather than trying to spend time extracting insights from themselves.
If every financial firm and/or investor uses FinTech services based on the similar ML algorithm, would not it homogenize the financial system and increase systematic risks?
(Sophie) I want to emphasize "No free lunch theorem". There's no single algorithm that works the best for all cases. Moreover, different datasets will give different results.
From your experience, what do you think would be the value of social networks' data or publicly available data, which are claimed to be used by many FinTech lending firms for the credit evaluation of an individual at this point? What do you think it would be like in the future? Is there any difference between individuals and firms?
(Sophie) Social Network Service data is biased, not complete, and usually difficult to match with target variables such as default and fraud probability that financial companies are interested in. In my opinion, it is not really worth spending time and resources to use SNS data for our business at this point.
If most of the financial market participants use ML, decision makings may become more homogeneous. Would not it worsen the probability of systematic risks such as a (bank) run?
(Olivier) I think if everyone has the same data, the correlation of decisions will increase with adopting ML. However, individuals, companies, and organizations have huge amount of private data now and will collect more in the future. As we said before, machine learning applied to different datasets will not lead homogeneous behaviors among financial market participants.
Many economists are still reluctant to adopt ML and big data because ML finds correlations in big data while not identifying causality which is important for policymaking. In your opinion, how can we use ML for policymaking decisions?
(Olivier) Well, I am not very familiar with the application of ML to social science and policymaking. In my opinion, for many cases, the data may not be "big" enough to justify the use of ML. Those applications require supervised machine learning, which implies that you need to already know which option is "more correct" than others. Still, ML may be useful if we know the underlying mechanism, the policy making decision is very specific (e.g. fine tuning of sales tax rate or interest rates), and there exist sufficiently big data for the process.
What are the barriers which may handicap the ML usage in banks and financial companies (potential impacts on the employment, regulation, etc.)?
(Sophie) I think one of the barriers is that banks and financial companies have to "explain" the result of their predictions to their customers and regulatory authorities. For example, if a bank rejects to give out a loan to a certain individual, that individual will ask the bank why, and if the bank does not provide a clear answer, the customer may file a complaint to the government, which will be a big trouble to the bank. However, ML is like a black box and it's difficult to explain logics behind prediction results.
What is the difference between desired objectives in computer science and in finance when it comes to applying machine learning?
(Sophie) In computer science, it is about finding the algorithm that beats the state-of-the-art, world-best one at the moment. On the other hand, in finance, it is about finding the explainable algorithm beating the incumbent models while working fast enough.
You had a choice to start your business either in France or in Korea. What factors affected your decision making?
(Olivier) First, compared with France and other EU countries, starting a company in Korea requires less complicated processes. Taxes and other regulatory burdens are also lighter in Korea. Second, compared with the US and the EU, compensation costs for engineers in Korea are lower. Third, the level of competition in the Korean market is also lower.
What are the points that you might emphasize in managing a company with people from different countries (France, Korea, Russia, and Sweden in Alphabetical order) and cultures like Solidware?
(Sophie) The common language must be English (Very important) all the time. All official documents and all talks are done in English. We emphasize task-based management. No hierarchy. (Olivier) Some of our engineers are not very comfortable with speaking in English, but they can communicate well with writing.
Do you want to stay specialized on the financial sector? For example, is there any room in your vision for a business model in which you would license your ML technology to banks and insurances? Who is your more relevant competitor?
(Sophie) We're trying to be more vertically integrated in the financial sector. That is, we are trying to adapt our solution to specific forms of datasets that financial companies often use. Potential competitors may be financial companies which are our clients as of now, if they try to build their own machine learning system.
How do you see the future of your company?
(Sophie) Bright! (Olivier) The market for ML application in finance grows fast in Korea and our revenue also does. Second, big names in ML such as IBM and Palentir are in the market, but their performance here is not on par with their reputation. We can cover more tailored and specified needs of our customer companies. As Sophie mentioned, potential competitors may be financial or credit information companies which want to internally have both data and technologies. Personally, I think it requires quite a long time for them to have both.