[Paper Review] Patterns of Individual Shopping Behavior

Coco Krumme, Manuel Cebrián+1 more|arXiv (Cornell University)|Aug 15, 2010

Human Mobility and Location-Based Analysis4 citations

TL;DR

This study analyzes individual shopping behavior using anonymized transaction data from 10,000 bank accounts over a three-month period, revealing that while daily shopping sequences are unpredictable due to interleaved events, overall behavior is highly predictable at longer time scales. Key findings show that wealthy individuals exhibit higher entropy due to greater variety in store visits and are more likely to bundle shopping trips, contrasting with poorer individuals who show greater routine predictability centered on core stores.

ABSTRACT

Much of economic theory is built on observations of aggregate, rather than individual, behavior. Here, we present novel findings on human shopping patterns at the resolution of a single purchase. Our results suggest that much of our seemingly elective activity is actually driven by simple routines. While the interleaving of shopping events creates randomness at the small scale, on the whole consumer behavior is largely predictable. We also examine income-dependent differences in how people shop, and find that wealthy individuals are more likely to bundle shopping trips. These results validate previous work on mobility from cell phone data, while describing the unpredictability of behavior at higher resolution.

Motivation & Objective

• To investigate the predictability of individual shopping behavior at high temporal resolution using real transaction data.
• To examine how demographic factors such as income influence shopping patterns, including trip bundling and store diversity.
• To compare information-theoretic entropy measures (random, uncorrelated, true) in shopping behavior to assess the role of sequence versus frequency in predictability.
• To validate and extend prior mobility studies based on mobile phone data to financial transaction patterns.
• To explore whether behavioral routines or individual choices dominate at short versus long time scales.

Proposed method

• The study uses a random sample of 10,000 anonymized financial accounts from a major U.S. bank, covering all card transactions, cash withdrawals, and wire transfers from Q1 2010.
• Merchants are classified using Merchant Category Codes (MCC), and individual income is estimated from inflows into accounts.
• Shopping sequences are modeled as networks where nodes are merchants and edge weights represent transition probabilities between stores.
• Information-theoretic entropy is computed in three forms: random (log₂Nᵢ), temporally uncorrelated (sum pᵢ(j)log₂pᵢ(j)), and true (based on sequence complexity via Kolmogorov complexity approximation).
• Monte Carlo simulations randomize and sort shopping sequences to isolate the effect of temporal order on entropy.
• Wealthy and poor individuals are segmented based on annual income inflows: < $16,000 (poor) and > $80,000 (wealthy), with comparisons of store variety, visit variance, and bundling behavior.

Experimental results

Research questions

• RQ1 To what extent is individual shopping behavior predictable at longer time scales despite apparent randomness at the daily level?
• RQ2 How do income levels correlate with shopping behavior, particularly in terms of store variety, visit frequency, and trip bundling?
• RQ3 Does the sequence of store visits significantly affect the true entropy of shopping behavior, or is it primarily driven by the number and frequency of visits?
• RQ4 How do information-theoretic entropy measures compare between financial transaction data and prior mobile phone mobility studies?
• RQ5 What behavioral mechanisms—such as trip bundling or preference for variety—explain differences in predictability between wealthy and low-income individuals?

Key findings

• Wealthy individuals visit a significantly greater variety of stores than poorer individuals, resulting in higher random entropy (p < 0.01), with a 31.4% chance of overlap in merchant type between two highly predictable individuals versus only 11.2% for a predictable and an unpredictable individual.
• Despite higher overall entropy, wealthy individuals show a greater difference between random and uncorrelated entropy, indicating that their behavior is more influenced by temporal distribution of visits than by mere variety.
• The variance in number of visits per store is higher among wealthy individuals, supporting the hypothesis that their higher entropy stems from preference for variety rather than temporal distribution.
• Wealthy consumers are more than twice as likely to bundle shopping trips, as evidenced by higher variance in daily activity bins, suggesting intentional coordination of purchases across multiple stores in a single trip.
• True entropy remains largely unchanged when shopping sequences are randomized, indicating that sequence order contributes little to predictability; however, sorting sequences over weekly intervals significantly reduces entropy, approaching levels seen in mobile phone data.
• Zipf’s law holds for store visit frequency (s = 4 ± 0.031), showing that the probability of visiting a store at rank N follows a power law, independent of total number of stores visited.