📄 social_science

Demographic Differences in the Predictability of Consumer Mobility

Using anonymized credit-card transaction data, this study quantifies how consumer mobility predictability varies across demographic groups and temporal scales, revealing that visitation sequences are highly predictable (96–97%) at the individual-store level but significantly less so at the business-type level, with slight seasonal differences observed between younger men and women.

Original authors: Zolzaya Dashdorj

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zolzaya Dashdorj

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Human movement is rarely random. We tend to return to the same coffee shop, the same gym, or the same grocery store, creating a rhythm that repeats day after day. Scientists who study how people move through cities have long known that these patterns are not just habits; they are predictable. By looking at the digital footprints we leave behind—such as the records of where we swipe our credit cards—researchers can measure how much of our future behavior is already written in our past. The core idea is simple: if we know where a person has been, how well can we guess where they will go next? This question sits at the intersection of urban planning and data science, asking whether the chaos of daily life hides a hidden order. The answer matters because understanding these limits helps us design better services, from traffic systems to local businesses, without needing to predict the future perfectly, but rather understanding the boundaries of what is possible to foresee.

A researcher set out to test these limits using a massive collection of credit-card records from a Spanish bank in 2011. They did not try to build a computer program to guess where people would go next. Instead, they used a method based on information theory to calculate the theoretical ceiling of predictability. Imagine trying to guess the next word in a sentence; if the sentence is highly repetitive, the guess is easy. If it is random, the guess is hard. The researcher applied this logic to store visits, measuring the "uncertainty" in a customer's history. They looked at two different ways to view these visits: first, as trips to specific, individual stores, and second, as trips to broader categories of shops, like "restaurants" or "clothing stores." By comparing these two views across different ages, genders, and times of the year, they mapped out how much regularity exists in our spending lives and who follows the most rigid patterns.

The study began by filtering the data to focus on active customers who had made between ten and one hundred visits in a single month. The researcher then analyzed the sequence of these visits. They found that when looking at specific stores, the patterns were incredibly strong. The math showed that if one knew a person's past visits to specific locations, the theoretical limit for guessing their next stop was between 96% and 97%. This means that for most people, their choice of a specific store is highly determined by their history. However, when the researcher grouped those same visits into broader categories, the predictability dropped. Knowing that a person visits a "restaurant" today made it about 10 percentage points harder to guess which specific type of restaurant they would visit next compared to knowing they visited a specific restaurant. This gap reveals that our strongest habits are tied to specific places, not just general types of activities.

The researcher also examined how these patterns changed over time and across different groups of people. They looked at data from different seasons and found that predictability was slightly higher in the winter and summer months compared to spring and autumn. This suggests that people's routines become more repetitive during the extremes of the year, perhaps due to weather or holiday schedules. When they broke the data down by age and gender, a clear trend emerged: predictability generally decreased as people got older. This means younger adults tended to have more consistent routines than older people. There was also a small difference between men and women. In the shorter time frames, such as a single season, younger men showed slightly higher predictability than women of the same age. However, when the researcher looked at the entire year of data, this gender gap disappeared, and the patterns for men and women became nearly identical.

These findings offer a precise look at the structure of human behavior, but the author is careful to define what their numbers mean. The 96% to 97% figure is not a score that a computer model achieved in a test; it is a theoretical limit. It represents the maximum amount of predictability that exists within the data itself, regardless of how smart the prediction tool might be. The study does not claim that a specific algorithm could reach this number, nor does it suggest that every single visit is predictable. Instead, it establishes that the information contained in our transaction history is rich with order. The data comes from 2011, so it reflects the payment habits of that era, and it only captures the movements recorded by one bank, not every place a person might go. Yet, within those boundaries, the results are clear: our daily lives are governed by strong, measurable rhythms, and those rhythms vary depending on whether we are looking at a specific shop or a general category, and who we are.

Technical Summary: Demographic Differences in the Predictability of Consumer Mobility

Problem Statement
While human mobility predictability has been extensively studied using mobile phone trajectories and location-based social media, there is a need to understand how theoretical predictability varies within consumer behavior derived from financial transaction data. Specifically, it remains unclear how the "regularity" of store-visitation sequences differs across demographic groups (age and gender) and temporal scales (monthly, seasonal, annual). Furthermore, a critical distinction exists between predictability at the level of specific individual establishments versus broader business categories. This study addresses these gaps by quantifying the theoretical upper bound of predictability for consumer mobility using anonymized credit-card transaction records, rather than developing a new predictive model.

Methodology
The study utilizes a dataset of anonymized bank-card transactions from BBVA (Spain) recorded in 2011, covering over 300,000 point-of-sale terminals. The analysis focuses on 22,703 active customers who recorded between 10 and 100 store visits and visited 10 to 50 distinct stores during the observation period.

The core methodology relies on information-theoretic measures to quantify uncertainty and derive a theoretical limit on predictability:

Entropy Measures: Four entropy metrics are calculated for each customer's visitation sequence:
- Random Entropy ( $H_{rand}$ ): Based solely on the number of distinct locations visited ( $N_i$ ).
- Temporally Uncorrelated (Shannon) Entropy ( $H_{unc}$ ): Accounts for heterogeneous visitation frequencies but ignores temporal order.
- Conditional Entropy ( $H_{cond}$ ): A first-order Markov measure conditioning the next visit on the immediately preceding visit.
- Sequence-Dependent Entropy ( $H_{est}$ ): A Lempel-Ziv-type nonparametric estimator capturing repeated subsequences.
Predictability Bounds: Using Fano's inequality, the study translates these entropy values into a maximum theoretical predictability ( $\Pi^{max}$ ). This represents the upper bound on the probability that an ideal predictor could identify the next location, given the sequence's entropy and the number of distinct states.
Comparative Framework: The analysis compares two representations of mobility:
- Individual-store sequences: Specific merchant visits.
- Business-type sequences: Visits aggregated into eight broader categories (e.g., Bars/Restaurants, Apparel, Health/Wellness).
Demographic and Temporal Analysis: Results are stratified by gender, age groups (in five-year intervals), and temporal observation periods (monthly, seasonal, and full-year).

Key Results

Store vs. Business Type: Individual-store visitation sequences exhibit substantially higher regularity than business-type sequences. The theoretical predictability bound for individual stores is approximately 96–97%, whereas business-type predictability is roughly 10 percentage points lower (peaking around 86–94% depending on the metric). This indicates that consumers are more predictable when visiting specific establishments than when engaging in general activity categories.
Entropy and Uncertainty: The distribution of entropy measures confirms that incorporating visitation frequencies and temporal ordering significantly reduces uncertainty. While random entropy peaks near 3.68 bits (approx. 12.8 states), conditional and sequence-dependent entropies drop to approximately 0.47 and 0.32 bits, respectively, indicating strong short-term sequential regularity.
Demographic Variations:
- Age: Predictability generally decreases with age for both individual stores and business types.
- Gender: Younger men (under ~45) exhibit slightly higher seasonal predictability than women of comparable age. However, this gender gap diminishes significantly when analyzing full-year observation periods, where values for men and women converge.
- Seasonality: Store-visitation sequences show slightly higher predictability in winter and summer (~~97.74%) compared to autumn and spring (~~96.13%). Business-type predictability remains relatively stable across seasons (ranging narrowly from 88.34% to 88.57%).
Temporal Scale: Longer observation windows (full-year) reveal greater regularity compared to single-month analyses.

Significance and Claims
The paper explicitly positions its contribution as a demographic comparison of entropy-based theoretical predictability rather than a benchmark for state-of-the-art forecasting architectures (e.g., RNNs, Transformers, or Graph Neural Networks).

Theoretical Limits vs. Empirical Accuracy: The author emphasizes that the reported values (e.g., 96–97%) are theoretical upper bounds derived from information theory, not out-of-sample classification accuracy. These figures represent the maximum possible performance an ideal predictor could achieve given the inherent structure of the data, independent of any specific machine learning model.
Model Independence: By using information-theoretic measures, the study provides a model-independent view of the regularity available in mobility sequences. This allows for a direct comparison of behavioral structure across different demographic groups without the confounding variables of model selection or training design.
Behavioral Insights: The findings demonstrate that transaction-derived histories contain strong temporal regularities that vary systematically with the level of destination aggregation, age, gender, and temporal scale. The study concludes that while human mobility contains persistent structure, the strength of this structure is contingent on whether one analyzes specific locations or broader economic activity categories.

Limitations
The author notes that the data is from 2011 and may not reflect contemporary payment habits or online purchasing trends. Additionally, the data represents only transactions observed through the bank's infrastructure, not a complete record of all human movement. The aggregation of merchant categories into eight broad types also influences the results at the business-type level.

Technical Summary: Demographic Differences in the Predictability of Consumer Mobility

More like this