Analysts, traders and computer scientists have been looking at more advanced ways to make or save money for decades. The industry itself is prime for machine learning, as it has immense historical records, large data sets, and a quantifiable method of measuring results. What we will focus on are the following: fraud detection, sentiment analysis, predictive analytics, underwriting, customer service, financial product recommendations, and image recognition in claims.
The most obvious usage of machine learning involves fraud detection. This is nothing new, and has been used for a few years by companies such as PayPal. Every data set is different and requires its own training and fine tuning. Suspicious behavior of customer should be reported to a consultant in a augmented intelligence (AI) manner, to complement human cognitive skills of recognizing fraud patterns. Good UX improves drastically performance of each consultant and their accuracy. Good looking UI accelerates adoption of the tool. We have experience in crafting fraud detection tools, engineering leaner manual review processes and building dashboard to analyze performance of each consultant.
Communication with financial institution is shifting towards messaging and phone/video calls. Thousands of interaction with call centers are impossible to monitor and evaluate with traditional methods. This is where NLP, in combination with sentiment analysis can be immensely useful. Sentiment analysis can improve manual reviewing of phone calls, optimize agent delegation by their individual skills and prioritize interaction with panicked or stressed customers. The same methods could be used while on a chat or call with a debtor to assess their honesty. Furthermore, we could automatically assess the state of a company’s public perception with an aggregated twitter sentiment analysis or a historical news sentiment analysis.
What better way to increase our earnings, than by predicting the future. Predictions are never guaranteed, they only have a confidence score. Behavioral patterns, be they macro or micro, are the most significant features we should look for when doing a historical analysis and trying to use predictive analytics. Overfitted trading models can be a concern as it is quite difficult to account for market volatility. Once we get outside of trading, we find room for all sorts of uses. In subprime lending, we can predict which of our users credit behavior will improve over time. A vehicle loan, that may seem too risky right now, may become acceptable when we take into account the benefits provided when an individual has access to a personal vehicle. Venture capitalists could look at characteristics and behaviors of founders in relation to the financial situation of startups. By combining new datasets, trends, technological development, sentiment analysis and predictive analytics, we could even adjust actuarial models.
Conventional methods of underwriting are not poor, but like any model, could always be better. Securities, banking and insurance fields can all benefit from advances in technology. Data collection and client histories providing us with a significant corpus of data to train our models with. Long histories are significantly more useful, as they can limit potential overfitting and idiosyncrasies arising from market conditions during certain years. Machine learning models decrease the risk associated with current borrows while increasing the number of approvals. Behavioral trends, for example can indicate added risks associated with the driving habits of certain vehicle owners. Such predictive models can assist in serving areas that have been traditionally underserved. These customer segments not have a long financial history, but insights from similar clients can be transferred to more accurately infer the risk involved. Our reproducible models alleviate regulatory issues, as it becomes easier to audit the decision process involved.
Customer Service is one of the most resource consuming processes within finance and insurance industry. From handling paper forms to direct communication with customer. Recent advances in NLP (natural language processing) have increased level of automation in document processing and speed of resolving customer issues. As chat is growing and becoming one of the most prevalent forms of communication between consumers and business, augmented Intelligence solutions allows agents to reply with recommended predefined answers than can be highly personalized. Text sentiment can be detected, allowing an agent to prepare for conflict resolutions and receive additional recommendations on how to lower tension levels. On the other hand machine learning can help in pre assessment of insurance claims or credit request.
Each contact with customer, especially business one, with product recommendation not addressing actual needs has a negative impact on relationship. Personalized product recommendation based on customer data can be applied to a few different aspects of finance. Consumer applications start with analyzing a given customer’s behaviors, habits, and other associated data points. With these analyses in hand, we can disseminate the types of products our users will find most useful. We will be able to tell which users will be the most inclined to take advantage of a new credit card offer or consider an investment opportunity instead of cold calling customers with random offer. The business side of such recommendations work on a much larger scale. Here, we must take into consideration entire swaths of data. Marketing and product teams can then take advantage of this data to create new promotions or packages based on unfilled market opportunities.
Mobile applications have allowed users to take pictures of their assets when filing an insurance claim, thereby increasing their own convenience while lowering costs associated with estimators. But an agent was still required to make an assessment of these images. He would then flag claims that require a human to asses the case in person. Recent advances in deep learning have increased image recognition accuracy to levels that surpass humans. With that in mind, we can automatically reviews claims submitted and flag cases that need manual review, but can do so almost instantaneously. Customers would have their claims paid out in minutes, dramatically increasing satisfactions. Image recognition can be used as well in fraud detection by verification of signatures between documents or data consistency in scanned contracts – there is always a risk that a customer will swap contacts before signing.