Modeling User Redemption Behavior in Complex Incentive Digital Environment: An Empirical Study Using Large-Scale Transactional Data

Technical Summary: Modeling User Redemption Behavior in Complex Digital Incentive Environments

Problem Statement
The digital economy increasingly relies on complex incentive structures, specifically Partnership Loyalty Programs (PLPs), where multiple firms jointly issue and redeem reward points. While existing literature examines PLPs from firm-level performance or cross-purchasing transitions, there is a significant gap in understanding individual-level redemption behavior in real-world settings. Traditional social science methods struggle to analyze massive consumer logs within these multi-currency environments, particularly regarding the valuation of digital points against traditional money and the impact of atypical point acquisition methods (e.g., large, exogenous government grants). Furthermore, prior research on behavioral biases and non-standard point acquisition predominantly relies on laboratory experiments rather than actual transaction data.

Methodology
This study leverages a unique, large-scale dataset from Zaim, a popular third-party personal finance application in Japan. The dataset includes over 5.5 million unique purchase transactions and 5.9 million point transactions from approximately 42,000 customers, capturing both monetary spending and point-based transactions across multiple major PLPs. The authors employ three distinct empirical approaches to address their research questions:

1. Topic Modeling and Clustering (RQ1): To analyze redemption patterns, the authors preprocess text descriptions of point transactions and apply Latent Dirichlet Allocation (LDA) to identify latent topics. These topics are labeled using a Large Language Model (GPT-4o) and manually validated. User profiles are constructed as feature vectors based on these spending patterns and clustered using the k-medoids algorithm to identify distinct user groups.
2. Causal Inference via Natural Experiment (RQ2): To assess the impact of exogenous point grants, the study utilizes a quasi-experimental design based on Japan's "second My Number Point campaign" (late 2022). A Difference-in-Differences (DID) approach with doubly robust estimation is employed to compare a treatment group (users receiving a one-time JPY 7,500 grant) against a control group. This method controls for selection bias by estimating propensity scores based on pre-intervention demographics and behavioral variables.
3. Embedding-Based Exploration-Exploitation Analysis (RQ3): To quantify shopping styles, the authors apply a Skip-gram with Negative Sampling (SGNS) model (similar to word2vec) to learn vector representations of stores based on user visit sequences. Exploration and exploitation tendencies are calculated using the cosine similarity between a user's current day's store vector and their rolling mean of the previous six days. This metric is then correlated with point redemption diversity across PLP categories.

Key Results

• Demographic and Category Correlations: Point redemption patterns are not random but align predictably with demographics. Younger users (20s) allocate a significantly larger share of points to "Finance/Transportation" and "Investment" categories compared to older cohorts, who tend to redeem points for commercial facilities. This suggests a shift in redemption priorities from financial services toward discretionary goods as consumers age.
• Asymmetry in Point Flows: A fundamental asymmetry exists between point earning and spending. Points are earned incrementally but redeemed in larger, consolidated sums. Notably, a distinct group of users redeems points for financial services (e.g., mutual funds) rather than where they were earned (e.g., restaurants), indicating that consumers do not treat points as equivalent to cash, consistent with mental accounting theories.
• Impact of Exogenous Grants: The JPY 7,500 government grant significantly stimulated point redemption. In the first month, the treatment group redeemed approximately 16% more points than the control group. However, the grant had no statistically significant effect on monetary (cash) spending, suggesting that point grants do not directly substitute for cash and may not be effective for inducing broader monetary consumption. Recipients appeared to "hoard" points initially, with redemption peaking in the third week, suggesting deliberation rather than impulsive spending.
• Shopping Style and Redemption Diversity: There is a clear link between shopping styles and redemption patterns. "Exploitative" shoppers (those who frequent a narrow set of familiar stores) concentrate their point redemptions, while "exploratory" shoppers (those trying diverse merchants) redeem points more broadly across categories. Age is a stronger predictor of exploration tendency than redemption category, with younger cohorts exhibiting significantly higher exploration scores.

Significance and Contributions
The paper claims to advance the field by providing empirical evidence from a massive real-world economic ecosystem, moving beyond the limitations of laboratory experiments and single-firm datasets.

• Theoretical Contribution: The study extends the understanding of Loyalty Programs (LPs) to collaborative PLP contexts, offering empirical evidence on consumer rationality (or irrationality) in point usage, specifically regarding mental accounting and responses to exogenous stimuli.
• Managerial and Policy Implications: The findings offer practical insights for optimizing PLP design and targeting promotions based on demographic segments and shopping styles. For policymakers, the results suggest that while point-based public policies can effectively stimulate specific point-based consumption, they may not be effective tools for driving broader monetary spending or substituting for cash transfers.
• Methodological Contribution: The paper demonstrates the viability of using embedding techniques and natural experiments on third-party financial data to model complex, multi-currency consumer behaviors that were previously difficult to study with traditional methods.