User defined interpretability of machine learning algorithms applied to financial time series data

User studies have be underrepresented in Explainable AI (XAI) literature. While many papers seek to understand interpretability through Folk Psychology or Theory of Mind explanations, the XAI literature also needs to continue to explore rigorous measures of interpretability from the user standpoint. To accomplish this, we employed a straightforward, accurate learning model for automated allocation of capital from financial time series data, and investigated user reactions to, and ability to interpret, varied explanation strategies. After an initial survey gauging a user’s experience with machine learning models in general, the user engaged in an interactive demonstration with the model. Then, we measured the user’s notions of their own understanding and the model’s interpretability and trustworthiness. Interestingly, text explanations were the only type of explanation that increased user trust in a statistically significant way. At the same time, historical examples pulled from the dataset actually negatively impacted user understanding, and were the only explanation type to do so.