The Algorithmic Preference Drift: A General Equilibrium Framework on Endogenous Utility and the Atrophy of Consumer Sovereignty

How to cite

Lopez, Francisco Jacobo Murillo. 2026. "The Algorithmic Preference Drift: A General Equilibrium Framework on Endogenous Utility and the Atrophy of Consumer Sovereignty." AEA RCT Registry. AEARCTR-0018976. https://doi.org/10.1257/rct.18976-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Francisco Jacobo Murillo Lopez*

* Primary investigator

Registry ID
AEARCTR-0018976
DOI
https://doi.org/10.1257/rct.18976-1.0
First registered
Last updated
Status
In development
Trial dates
2026-07-01 – 2026-12-23
Countries
Private
JEL codes
D11, D12, D91, C93, G51, O33
Registry record
http://www.socialscienceregistry.org/trials/18976

Abstract

This study investigates how predictive Artificial Intelligence (AI) and automated financial technologies (Fintech) influence consumer decision-making, price memory, and financial autonomy. While traditional economic theory assumes that consumers possess stable, independent preferences, modern digital platforms utilize sophisticated algorithms that continuously learn from past behaviors to personalize user experiences. This research introduces a framework of "induced endogenous preferences" to examine whether these adaptive systems subtly reshape consumer utility functions and impact financial sovereignty. To empirically evaluate this phenomenon, we will conduct a 6-month, double-blind Randomized Controlled Trial (RCT) involving 2,400 active Fintech users in Aguascalientes, Mexico. Participants will interact within a controlled, custom-developed digital marketplace ("Sandboxed Fintech Wallet") where they will be randomly assigned to one of three interaction regimes: (1) a manual control group with high behavioral friction and maximum price salience; (2) a dynamic nudging group exposed to personalized predictive recommendations; or (3) a full automation group where the AI autonomously manages budget allocations and recurring transactions. Throughout the study, high-frequency data will be captured using real-time application telemetry combined with bi-weekly micro-surveys deployed via the Experience Sampling Method (ESM) to track shifts in expenditure patterns, attention intensity, psychological agency, and price memory accuracy. To rigorously test consumer resilience and adaptive capacity under economic stress, two controlled exogenous price shocks (simulated inflation and cost volatility) will be sequentially introduced into the market environment. Finally, an experimental decoupling phase will remove all algorithmic enhancements to assess behavioral hysteresis and long-term cognitive dependency. By deploying advanced econometrics and causal machine learning—including Fixed Effects panel models, Random-Intercept Cross-Lagged Panel Models (RI-CLPM), and Causal Forests—this project seeks to determine whether heavy reliance on algorithmic optimization yields true economic efficiency or, conversely, leads to a systematic erosion of consumer agency and financial self-efficacy. The findings aim to provide critical empirical evidence for the design of consumer protection policies and regulatory frameworks in the rapidly evolving digital economy

Intervention

1. Algorithmic Architecture and Treatment Engineering
The "Sandboxed Fintech Wallet" mobile application interfaces with a centralized secure server (UAA Dedicated Private Cloud) that dynamically modifies the application's API responses based on the individual's block assignment:
* Arm B (Dynamic Predictive Nudging): Operates on a hybrid recommendation system engine. The Content-Based Filtering (CBF) module processes a dynamic text-vector of item features, while the User-Based Collaborative Filtering (UBCF) module calculates cosine similarities across the panel's multi-period transaction lags. It systematically computes the predictive optimization vector $\theta_{it-1}$ to alter the information architecture and display customized behavioral prompts.
* Arm C (Full Automation): Actively implements automated financial routing algorithms. It isolates the agent's cash flow constraints and automatically processes optimal budget allocations, predictive automated savings, and one-click purchase checkouts based on the system's loss-minimization function.

2. High-Frequency Experience Sampling Method (ESM) Sequence
The application forces a strict data capture loop embedded within the user interface at every transaction point or scheduled bi-weekly wave ($T = 48$ measurement points). The data collection flow executes the following sequential steps:
1. Passive Telemetry Extraction: Real-time log capture of transaction values, exact timestamps, account balance visibility status, and the three raw components of the Shadow Liquidity Index ($SLI_{it}$): balance check frequency, screen-time on the budget dashboard, and alert open rates.
2. Psychometric ESM Chatbot Trigger: A micro-survey interface ($<90$ seconds response window) is deployed, enforcing a strict internal question order to eliminate cognitive contamination:
* Block A (Financial Agency - E_{ps}): Captures within-subject variations in internal financial locus of control and perceived self-efficacy utilizing a 2-item quick-response adaptation of the Generalized Self-Efficacy Scale (1–7 Likert Scale).
* Block B (Price Memory - Delta P_{error}): Prompts the user to input the estimated baseline price of a standardized reference market basket without checking current balances.
* Block C (Post-Purchase Regret): Deploys Marquina’s Post-Purchase Regret Scale items to monitor latent post-consumption dissatisfaction.

3. Exogenous Shock Parametrization and Safety Protocols
* Macroeconomic Shock 1 (Week 12): The server modifies the catalog database pricing array, inducing a sudden, forced 20% price increase on all essential non-durable goods (simulated inflation). The database tracks how the household consumption matrix mutates as the budget constraint tightens, evaluating changes in the Kullback–Leibler Divergence index ($D_{KL}$).
* Macroeconomic Shock 2 (Week 20): Introduces real-time price dispersion and artificial search frictions on substitute items to rigorously test the accuracy of the baseline price memory metric $\Delta P_{error}$.
* Hysteresis Decoupling (Weeks 21–24): All machine learning recommendation and automation script executions are terminated on the server for Arms B and C. The application user interface defaults to the unified manual regime (Arm A) to isolate the persistence of cognitive dependencies and evaluate behavioral hysteresis.

4. Data Protection and Anonymization Pipeline
All transaction telemetry and psychometric responses are encrypted in transit and at rest using the AES-256 cryptographic protocol. Direct personal identifiers and device MAC addresses are permanently separated at the registration gateway and replaced by a unique, irreversible alphanumeric hash token. External transactional validation emulates the structure of the Mexican Comprobante Fiscal Digital por Internet (CFDI) to guarantee data validity without storing tax or identity-sensitive information. Upon final data consolidation in December 2026, the master cross-reference identity key will be permanently destroyed under ISO/IEC 27001 compliance guidelines.

Experimental design

This study implements a randomized controlled trial (RCT) with a longitudinal panel design to investigate how different levels of financial technology interaction affect consumer choice, budget allocation, and cognitive financial attention. The experiment is embedded within a customized sandbox mobile fintech application developed in collaboration with the Universidad Autonoma de Aguascalientes (UAA). The sample consists of undergraduate and graduate students who are followed over a multi-week period divided into consecutive measurement waves.

Participants are randomly assigned to one of three balanced parallel experimental arms at the beginning of the study. The first arm, Arm A, serves as the unified manual control group, where users interact with a standard digital wallet interface and must execute all financial planning, expenditure tracking, and transfers manually without any computational assistance or external prompts. The second arm, Arm B, introduces predictive behavioral nudges, where the application platform dynamically modifies the information architecture and deploys personalized, text-based alerts based on the individual's recent multi-period transaction history to guide budgeting decisions. The third arm, Arm C, introduces a high-automation regime, where the application software automatically processes optimal budget allocations, processes pre-scheduled automated savings, and handles micro-purchases based on cash flow constraints, reducing the required manual interaction to a minimum.

The data collection architecture combines high-frequency passive telemetry logged continuously by the application server with primary data from an Experience Sampling Method (ESM) micro-survey infrastructure. The application forces data capture loops at specific transaction points and during scheduled bi-weekly waves. The study incorporates sequential exogenous changes in the application environment to rigorously evaluate consumer resilience, price memory accuracy, and behavioral adjustments. Toward the final stage of the experiment, a decoupling phase is introduced where automated features and nudges are deactivated, reverting all groups to the manual regime. This specific phase isolates the persistence of cognitive dependencies and evaluates whether past interaction with automated financial tools induces behavioral hysteresis once those tools are removed.

Design details

Randomization method
The allocation of participants to the three parallel experimental arms is executed automatically through a centralized server-side computer routine upon initial deployment of the mobile application. The randomization process utilizes a pseudo-random number generation algorithm embedded within the backend registration gateway database hosted on the Universidad Autonoma de Aguascalientes Dedicated Private Cloud. When a participant completes their profile registration and logs into the application for the very first time, the system triggers an automated script that draws a random assignment value. This assignment permanently links the user's unique alphanumeric hash token to one of the three experimental conditions: the manual control group, the predictive nudging group, or the full automation group. This digital randomization process is performed entirely in the office backend without any human intervention or public lottery, ensuring a strictly blind, exogenous, and balanced treatment allocation across the longitudinal panel.
Randomization unit
The unit of randomization in this experiment is the individual user. There are no clusters, schools, firms, or experimental sessions used as aggregated groups for treatment assignment. Every participant who enters the study is treated as an independent statistical unit and is randomized single-handedly by the server-side routine upon their first login to the mobile application. The treatment architecture, the machine learning recommendation configurations, and the automated asset routing scripts operate exclusively at this individual level throughout all forty-eight measurement waves of the longitudinal panel.
Clusters
There are 0 clusters planned for this experiment. The trial uses a pure individual-level randomization design where each participant is treated as an independent statistical unit. No aggregate grouping units, such as schools, firms, or classrooms, are utilized for the treatment assignment. The entire planned sample consists of individual users who are randomized single-handedly by the application server.
Observations
The planned number of observations for this experiment is 14,400 individual-wave data points. This total is calculated based on a target sample size of 300 individual participants who are followed longitudinally across 48 distinct measurement waves. Because the study does not use a clustered design, each observation unit corresponds to a single individual's telemetric and psychometric data bundle captured during a specific high-frequency wave. The longitudinal panel structure is designed to yield 48 repeated observations per subject, ensuring sufficient statistical power to capture within-person behavioral and cognitive variations over the course of the treatments and macroeconomic shocks.
Arms
The planned sample size is balanced equally across the three experimental arms, with a target allocation of 100 individual participants per arm, resulting in a total sample size of 300 individual subjects. The first arm, which is the manual control condition, contains 100 individual participants who execute all financial planning and tracking manually. The second arm, which is the predictive behavioral nudging treatment group, contains 100 individual participants who receive personalized text-based alerts based on their transaction history. The third arm, which is the high-automation treatment group, contains 100 individual participants who have their optimal budget allocations and savings processed automatically by the application software. Because this trial does not utilize a clustered design, these sample counts refer strictly to independent individual users, generating a balanced longitudinal panel matrix of 4,800 observations per arm across the forty-eight measurement waves of the study.
Primary outcomes
Primary Outcomes (End Points)The experiment measures four primary, high-frequency outcome variables at the individual-wave level (i,t) to evaluate the impact of algorithmic interaction on consumer behavior and cognition: 1. Shadow Liquidity Index (SLI_it) A continuous, objective behavioral metric quantifying digital financial attention intensity. It is constructed from passive device telemetry using the following logarithmic transformation: Formula: SLI_it = ln [ (Balance Check Frequency_it * Alert Opening Rate_it) / Budget Dashboard Screen Time_it ] 2. Expenditure Divergence Index (D_KL) A directional statistical distance metric quantifying the structural departure of the subject's observed consumption basket at time t (Q_it) from their baseline pre-treatment expenditure profile (P_i). It is computed via the Kullback–Leibler Divergence formula: Formula: D_KL ( Q_it || P_i ) = Sum_k [ Q_it(x_k) * ln( Q_it(x_k) / P_i(x_k) ) ] Note: Symmetry and scale stability are validated post-hoc via the Jensen–Shannon Distance. 3. Financial Agency Erosion (E_ps) A latent psychometric construct measuring variations in internal financial locus of control and perceived financial self-efficacy. It is captured intra-subject using a 2-item agile scale (1–7 Likert Scale) triggered in real-time via the Experience Sampling Method (ESM) chatbot immediately following transactions. 4. Price Memory Error (Delta P_error) A continuous cognitive accuracy metric assessing price anachronism and monetary desynchronization. It is calculated as the absolute standardized difference between the subject's unprompted ESM-reported price estimate (P_hat_it) and the actual current market price (P_it) of a reference consumption basket, scaled by the baseline price standard deviation (sigma_P): Formula: Delta P_error = | (P_hat_it - P_it) / sigma_P | Evaluation: This endpoint will be strictly evaluated immediately following the introduction of the sequential exogenous price shocks in weeks 12 and 20.
Secondary outcomes
The experiment tracks three secondary outcome variables at the individual-wave level (i,t) to capture downstream behavioral and cognitive adjustments following algorithmic exposure: The first secondary endpoint is Post-Purchase Regret, abbreviated as PPR(it). This is a continuous psychological scale metric that measures the retrospective utility degradation and cognitive dissonance experienced by the consumer after a transaction has occurred. The second secondary endpoint is the Intertemporal Consumption Variance, abbreviated as ICV(it). This is an objective behavioral metric designed to measure the smooth or volatile nature of household consumption across waves. It tracks the standard deviation of total weekly expenditures relative to the individual's long-run moving average. The third secondary endpoint is the Baseline Price Elasticity Estimate, abbreviated as E_est(it). This is an empirical behavioral parameter that measures the subject's operational responsiveness to the exogenous inflation shocks. It evaluates the percentage change in the volume of essential non-durable goods purchased divided by the percentage change in price observed during the shock waves.

Cited by

No citation data available for this registration yet.

Keywords