Project Nottingham: A Study of User Behavior on Mobile Investing Applications

Abstract

Investing or trading in the stock market has become a task just a button click away. Financial trading applications made this possible with their effortless and user-friendly design to buy and sell stocks. It has been observed that risky trading behavior on such applications has caused volatility in financial markets. This thesis explores the prediction of risky behavior using users’ actions on an investing application. To achieve this, we conducted an experiment where participants generated behavioral data from transactions with a simulated trading app. An unsupervised learning approach was undertaken to cluster users based on the time series data collected from the simulation. We identified distinct clusters of users based on app usage data which reflected degrees of risky behavior. To determine their risky behavior, an assessment of the clusters based on the degree of intrinsic risk associated with their common actions, as well as responses to survey questions made by the participants before the task was conducted. The algorithm which distinguished the user behavior in the most appropriate way was the TimeSeriesKMeans method. The survey data were used as labels for classification task to explore the reliability of psychometric surveys to predict user behavior and SVM (Support Vector Machine) and Logistic Regression classifier provided the most accurate results among other classification algorithms. Moreover, the factors involved in user enjoyment from simulation were also explored. This work demonstrates a step towards identifying a method for conducting and assessing clustering and classification models for the purpose of risky behavior detection using psychometric measures for evaluation and the reliability of those measures for identifying risky behavior apart from the actual behavior.