[Paper Review] Know Your Clients’ Behaviours: A Cluster Analysis of Financial Transactions

John R. J. Thompson|arXiv (Cornell University)|Jan 25, 2021

Financial Markets and Investment Strategies68 references19 citations

TL;DR

This study applies a modified RFM (Recency, Frequency, Monetary) model and k-prototypes clustering to 50,000+ client financial transaction records, revealing that Know Your Client (KYC) data explains little of investor trading behavior—instead, trade frequency and volume are the strongest predictors of investor behavior clusters. The findings challenge the assumption that KYC data reliably informs suitability, suggesting regulators and advisors should adopt more dynamic, behavior-based metrics.

ABSTRACT

In Canada, financial advisors and dealers are required by provincial securities commissions and self-regulatory organizations—charged with direct regulation over investment dealers and mutual fund dealers—to respectively collect and maintain know your client (KYC) information, such as their age or risk tolerance, for investor accounts. With this information, investors, under their advisor’s guidance, make decisions on their investments that are presumed to be beneficial to their investment goals. Our unique dataset is provided by a financial investment dealer with over 50,000 accounts for over 23,000 clients covering the period from January 1st to August 12th 2019. We use a modified behavioral finance recency, frequency, monetary model for engineering features that quantify investor behaviours, and unsupervised machine learning clustering algorithms to find groups of investors that behave similarly. We show that the KYC information—such as gender, residence region, and marital status—does not explain client behaviours, whereas eight variables for trade and transaction frequency and volume are most informative. Hence, our results should encourage financial regulators and advisors to use more advanced metrics to better understand and predict investor behaviours.

Motivation & Objective

• To investigate whether KYC data (demographics, risk tolerance) reliably predicts investor trading behavior.
• To develop behaviorally informed financial features using a modified RFM model from behavioral finance.
• To identify distinct investor behavior clusters using machine learning and validate their consistency with KYC profiles.
• To assess the suitability of investor behaviors relative to KYC-generated risk tolerance and regulatory expectations.
• To inform regulatory and advisory practices by identifying gaps between KYC assumptions and actual investor behavior.

Proposed method

• Engineered behavioral features using a modified RFM model, incorporating recency, frequency, and monetary value of trades.
• Applied k-prototypes clustering to group investors based on mixed numerical and categorical features, minimizing a similarity cost function.
• Used silhouette coefficient and Davies-Bouldin (DB) score to evaluate clustering quality.
• Visualized clusters using t-distributed stochastic neighbor embedding (t-SNE) for interpretability.
• Compared observed risk tolerance (from trading behavior) with anticipated risk tolerance (from KYC) across clusters.
• Planned future analysis of asset mix, security risk ratings (SRR), and portfolio returns to assess suitability and outcomes.

Experimental results

Research questions

• RQ1 To what extent does KYC data predict actual investor trading behavior in a real-world financial advisory setting?
• RQ2 Do clusters of investors with similar KYC profiles exhibit consistent and predictable trading behaviors?
• RQ3 How does the observed risk tolerance derived from trading behavior compare to the risk tolerance inferred from KYC data?
• RQ4 Are investors at the extremes of risk tolerance (high or low) adequately captured by current KYC frameworks?
• RQ5 Can machine learning-based behavioral clustering improve suitability assessments and investment outcomes?

Key findings

• KYC data, including age, income, and self-assessed risk tolerance, showed little to no correlation with actual trading behavior across investor clusters.
• The average observed risk tolerance across all clusters was 3.12 out of 5, with a standard deviation of 0.73, indicating consistent moderate risk behavior regardless of KYC profile.
• Cluster 1 (Active Traders) exhibited a risk tolerance of 3.19/5 but was anticipated to be 5/5, indicating a significant mismatch between KYC expectations and actual behavior.
• Cluster 4 (Older Investors) had the lowest observed risk tolerance (2.95/5) but was anticipated to be 1/5, suggesting KYC underestimates risk aversion in older clients.
• The distribution of risk tolerance across clusters was remarkably similar, indicating that KYC-based risk scores do not reliably predict or reflect actual trading behavior.
• The study concludes that KYC criteria may create rigid, narrow categories that fail to accommodate extreme risk preferences, especially among highly risk-averse or aggressive investors.