The best items sell out first, so a recommender that learns from what is left mistakes them for unwanted. Judging each choice against what was actually available at the time fixes it.
Many digital platforms recommend products from a demand-depleted catalog, such as crowdfunding projects, job postings, and house listings, where the most appealing products attract the most demand and disappear first. Standard negative sampling treats these products as rejected by later users, causing the phenomenon that we call the availability bias in recommender systems. This bias concentrates on the best products, persists as data grows, and can invert rankings, leading to misguided model selection and irrelevant recommendations. Because it enters through the eligibility of the candidate set itself, existing debiasing corrections and negative-sampling schemes, which operate within a candidate set taken as given, cannot remove it. We propose the active-set estimator, which scores each choice only over the reconstructed set of alternatives open at that moment, and prove that it recovers the true utilities even though closure is caused by the choices being modeled. On a large education-crowdfunding marketplace, training on the whole catalog instead of the active set costs the strongest methods about three-quarters of their accuracy. Under the correct protocol, the best-performing method captures 87% of donation dollars in a donor’s top ten recommendations, worth about $22 million per test window. The size of the bias follows the share of closed items in the candidate set: on a second demand-depleted platform, where sales close fewer than one listing in a hundred, it is negligible, and on a third platform, where demand sustains availability, the bias reverses sign. These results provide platform operators with a consistency-backed estimator, the measured cost of ignoring availability, and design principles for any market in which demand removes the most attractive products.
Keywords: Recommender Systems, Availability Bias, Negative Sampling, Choice Modeling, Demand-Depleted Platforms, Crowdfunding
NFT buyers act on a small set of strategic intents that recur across users and evolve within each user. Modeling those intents improves recommendation accuracy over co-occurrence and attribute-matching baselines.
The rapid growth of interest in non-fungible tokens (NFTs) has generated increasing demand for effective asset discovery systems in NFT marketplaces. Enhancing asset discovery can help improve user experience, stimulate market participation, and ultimately strengthen the NFT ecosystem. Unlike conventional goods, NFTs are unique digital assets whose value is primarily driven by social consensus rather than intrinsic attributes. As a result, user engagement with NFTs is governed by strategic intents, leading to intent-dependent relational patterns among NFTs that evolve within individual users over time and remain consistent across users with shared intents. Traditional recommendation approaches, which rely on co-occurrence statistics or intrinsic attribute matching, fail to capture this intent-driven complexity. We propose an Intent-Aware Theory-Anchored Recommendation (IARL) model for NFT recommendation. IARL models relationships among NFTs under distinct latent intents, capturing both their temporal evolution within individual users and their consistency across users. This intent-aware relational modeling aligns with real-world purchasing dynamics and enables more accurate user preference representations. Extensive experiments on a real-world, large-scale NFT transactions dataset demonstrate IARL's significant performance enhancement over state-of-the-art baseline methods, validating its effectiveness in the NFT recommendation scenario.
Keywords: Blockchain Technology, NFTs, Recommender Systems, Sequential Recommendation, Hypergraph Neural Networks, Purchase Intent, Disentangled Theory-Anchored Recommendation
AI-extracted consumer constructs can reproduce survey findings and still fail convergent and discriminant validity tests. A directed-response data collection procedure restores validity across five measurement methods.
The abundance of unstructured consumer data makes Generative Artificial Intelligence (AI) an attractive methodological tool for consumer research, yet its validity for Natural Language Processing (NLP) applications remains insufficiently examined. This paper addresses this critical gap by examining contemporary NLP methods: AI, Deep Learning (DL), and Machine Learning (ML), through three interconnected studies. In Part 1, we juxtapose findings derived from a multi-method study encompassing experimental and field data (analyzed using NLP methods). Despite striking similarities in the results, we identify potential validity threats stemming from method bias and conceptual overlaps in NLP-extracted constructs. Part 2 applies construct validity assessment to compare these NLP methods. The results reveal significant challenges in achieving convergent and discriminant validity when measuring consumer psychological constructs. In Part 3, we introduce a novel directed-response data collection procedure that substantially reduces measurement error while enhancing construct validity. We demonstrate this approach by measuring the variables in Part 1 through five distinct methods (self-report, expert ratings, AI, DL, and ML), establishing a more robust methodological framework for consumer behavior research employing NLP techniques. Our findings highlight the necessity of methodological caution when implementing NLP tools for consumer research, while providing a practical pathway for their valid application.
Keywords: Artificial Intelligence, Construct Validity, Natural Language Processing, Research Methodology, Consumer Behavior, Multi-method Research, Large Language Models
A Heckman-style correction for deep credit-risk models, so that a model trained only on approved loans stays valid for the applicants a lender turned down.
ChatGPT-polished loan applications carry less soft information and raise credit misallocation. Lenders respond by weighting hard information more, which partly offsets the loss.
We document systematically how various dimensions of writing quality matter for different borrower groups on a dominant online credit platform, and would be reshaped by the use of large language models (LLMs). Using human assessments, we find that LLMs (e.g., ChatGPT) significantly enhance the writing and perceived quality of loan applications. We further build proprietary BERT-based multimodal models of lenders' decision-making that accommodate lenders' constraints, and introduce "deep" Heckman correction for sample selection bias. We demonstrate that ChatGPT adoption decreases soft information conveyed and increases credit misallocation, likely due to convergence in writing; when lenders respond to borrowers' LLM adoption, they rely more on hard information, mitigating misallocation. We leverage generative modeling to characterize the counterfactual equilibrium with endogenous borrowers' LLM adoption and lenders' responses. Our findings provide insights into the evolving role of soft information and potential impacts of GenAI on lending.
Keywords: Generative AI, Heckman Correction, Lending, LLMs, Misallocation, Multimodal Data
Expert forecasts are formed in sequence, so a well-chosen subset can beat the full-crowd consensus. A ranking network that uses each forecast's position in the record selects that subset and improves on the all-report average across half a million earnings forecasts.
As expert opinions are formed sequentially, each rests on a distinct information basis. The admissible history may support learning and correction or transmit common errors. Moreover, a better-chosen subset may outperform the full crowd, yet consensus methods fail to account for these features. This motivates our study of Sequential-Opinion Selective Aggregation (SOSA), the decision problem of choosing which dated opinions should represent the crowd when collective wisdom is formed from a sequential record. Existing methods either score opinions in isolation, ignore chronology when drawing on group context, or decouple the ranking from the aggregation decision it serves. To address SOSA, and premised in forecast combination theory, we develop an end-to-end system, the Temporal Dependence-Aware Ranking Network (TDRN). It uses date-valid report content, locates each opinion on the completed path, and learns a single quality score from opinion-level and aggregate-level feedback. The decision rule retains a prespecified subset of the highest-scoring opinions and averages them equally. In rolling out-of-sample tests on over half a million annual earnings-per-share forecasts for firms listed in the Chinese market from 2010 to 2023, TDRN achieves the strongest overall opinion-quality ranking and selective-aggregation performance compared with a range of competitive benchmarks. Its selected consensus improves on the all-report average under both fixed and proportional capacities. Matched comparisons show that chronology changes a limited but consequential part of the selected set. Its incremental value declines as selection capacity expands. These findings show how machine augmentation can use chronology to allocate representation among human opinions, improving crowd wisdom without replacing expert forecasts or obscuring the aggregation rule.
Keywords: Wisdom of Crowds, Forecast Combination, Learning to Rank, Analyst Forecasts, Selective Aggregation
Firms whose inclusion claims exceed what employees report experiencing see higher turnover, with the largest effect among minority employees.
This study examines how firms' emphasis on inclusion within their corporate social responsibility (CSR) initiatives influences employee turnover, with particular attention to inclusionwashing—the discrepancy between a firm's publicly communicated commitment to inclusion and employees' perceived inclusiveness of their workplace. We contribute to the CSR literature by demonstrating the workforce consequences of misaligned inclusion messaging. Drawing on three independent datasets, we combine firm-level inclusion disclosures with employee-level sentiment data to construct an Inclusion Perception Gap (IPG), which serves as a measure of inclusionwashing. We then assess the relationship between the IPG and employee turnover. Our analyses show that firms with larger IPGs experience significantly higher turnover, with especially pronounced effects among minority employees. We further corroborate these findings through an online behavioral experiment conducted on Prolific. The findings suggest that, despite outward signaling of commitments to inclusion, firms suffer internal consequences when employee experiences diverge from external narratives. Firms that overstate their inclusiveness risk higher employee turnover. For researchers, we provide a framework to study inclusionwashing and its workforce consequences while validating Glassdoor as a reliable source of employee sentiment data.
Keywords: Inclusionwashing, Corporate Social Responsibility (CSR), Employee Turnover, Behavioral Experiments, Econometrics, Multi-Method
Stockout risk spreads across substitute products within a store and across stores selling the same product, so availability behaves as a networked process.
Stockouts are typically studied as endpoint outcomes that affect demand, yet little is known about whether stockout risk itself propagates across interconnected retail units. This study investigates stockout contagion in grocery retail by combining (i) a hidden Markov model to infer weekly stockout probabilities from point-of-sale data and (ii) dynamic spatial panel models to quantify propagation across economically connected products and stores. The framework addresses two key empirical challenges: stockouts are unobserved in transactional data, and correlated stockout realizations may reflect shared shocks rather than structured dependence. Conditioning on promotional activity, supplier disruption indicators, demand controls, fixed effects, and explicitly defined networks, we isolate structured propagation patterns that persist beyond shared exposure. Using IRI scanner and consumer panel data, we document economically meaningful stockout contagion both within stores across substitute products and across stores for identical products. Within-store propagation is strongest when substitution networks are defined broadly at the category level rather than by a single closest substitute. Across stores, contagion is widespread under diffuse networks and remains strong when links are restricted to behaviorally proximate locations, while chain-based operational linkages generate more heterogeneous effects. Decomposition of spatial impacts indicates that stockout risk spreads primarily through network connections across stores and products, rather than remaining localized. On average, a one-percentage-point increase in the stockout probability of a focal SKU raises the stockout probability of connected products or stores by approximately 0.06 to 0.30 percentage points, with substantial heterogeneity across categories. Categories characterized by longer stockout duration exhibit stronger propagation, indicating that persistence amplifies contagion. These findings suggest that availability risk in retail operates as a dynamic networked process rather than as isolated SKU-store events, providing a scalable empirical framework for identifying and managing systemic stockout risk using standard transactional data.
Keywords: Retail Operations, On-shelf Stockouts, Contagion Effect, Empirical Study