Publications
Abstract
Implementing Deep Learning in Estimation of Heterogeneous Taste Parameters in Hierarchical Structural Models
Co-authored with Anastasia Lebedeva (MSBA from Simon Business School)
This paper introduces NN+S, a novel semi-parametric framework that integrates neural networks with structural models to estimate heterogeneous taste parameters with enhanced accuracy and interpretability, offering significant applications in marketing. By designing a neural network architecture that models both observed and unobserved heterogeneity, as well as intra-individual random taste shocks, NN+S captures complex preference structures while retaining structural interpretability. The NN+S model enables companies to predict consumer responses to counterfactual scenarios, such as new product introductions, personalized pricing, or changes in product attributes. Monte Carlo simulations assess NN+S's ability to recover the true taste parameters. Empirical results show that NN+S outperforms linear hierarchical Bayesian models (LHBM) and the model by Farrell et al. (2020) (FLM) on a classical margarine purchase dataset for the same individuals, but new purchases (in-sample predictions), leveraging purchase histories for precise targeting and market segmentation. For new customers (out-of-sample predictions), NN+S achieves a slightly higher hit rate than alternative approaches. Scalable via standard neural network tools, NN+S empowers marketers to optimize strategies such as dynamic pricing and product recommendations, providing a flexible and interpretable tool for demand estimation.
Published
Abstract
ConjointNet: Combining Conjoint and Consumer Panel Data Using Neural Networks
Co-authored with:
Mitchell Lovett (Simon Business School, Benjamin Forman Professor of Marketing)
Bhoomija Ranjan (Monash Business School, Senior Lecturer)
Accurate prediction of consumer preferences for new-to-market attributes is critical for successful product development and marketing strategies. This study examines the effectiveness of a novel approach that integrates conjoint survey data (SP data) with panel data on actual purchases (RP data) to improve the reliability of consumer preference predictions for new-to-market attributes (NTMA). By adopting the NN+S approach, we address the challenges associated with the interplay between RP and SP data, specifically focusing on mitigating the issues related to selecting the optimal tightness between these data types and preventing overfitting through a k-fold cross-validation procedure. We illustrate our methodology by applying it to a unique dataset and comparing its performance with an alternative approach (Ellickson et al., 2019). The proposed method demonstrates a substantial performance boost for both same-sample customers and out-of-sample customers, as well as for new-to-market attributes. The proposed method does not require an expert's opinion for selecting the linked attributes, which is an additional improvement over Ellickson et al., 2019.
In-progress
Abstract
Higher Prices under Optional Rule-Based Repricing: A Simulated Triopoly with Parallel Manual and Automated Markets
Solo project
We study how optional, rule-based automated repricing affects seller competition when human participants act as sellers in real-time simulated triopolies. Each participant operated simultaneously in two isolated markets with identical demand and payoffs. One market allowed only manual price changes, while the other also offered Amazon-style match and undercut rules with bounds. Our experiments showed that average prices and profits were higher when repricing rules were available. Observed levels in both environments remained far above the symmetric Nash benchmark implied by the multinomial logit demand specification, but automated markets lay closer on average to the joint-profit benchmark. Average prices also rose with the number of sellers who regularly activated rules, whereas the specific rule label had little additional association with market averages once automation was in use. In this paper we discuss economic channels consistent with these patterns. The paper contributes within-subject, parallel-market experimental evidence that isolates how optional rule menus shift average prices and profits relative to manual-only environments. Prior field evidence links high prices under automatic repricing to overnight resetting, meaning intermittent upward price moves. The mechanisms discussed in this paper do not rely on resetting.
Abstract
Microeconomic Foundations of Matrix Factorization: Recovering Unobserved Product Attributes from Online Ratings.
Co-authored with Sangkil Moon (Department Chair and Cullen Endowed Professor of Marketing, Belk College of Business)
This paper investigates an application of the matrix factorization (MF) approach for the estimation of unobserved product characteristics from online ratings. We illustrate our approach using online reviews from the beer industry. The results show that estimated unobserved characteristics can be associated with observed characteristics, showing potential usage in cases where observed characteristics are not available. At the same time, estimated unobserved characteristics have several advantages over observed characteristics, including relevance, interpretability, and low dimensionality. In the case of the beer industry, the incorporation of estimated unobserved characteristics improves the substitution patterns of the aggregate demand model. This paper quantifies the errors of the estimates and shows that the bias caused by the selection of reviews is negligible. This technique can also be applied for the estimation of characteristics to improve the substitution pattern estimates in industries such as video games, movies, and books.
Abstract
Signed Product2Vec: Mapping Like and Dislike Structure from Binary Game Recommendations
Solo project
This paper proposes a signed extension of Product2Vec and P2V-MAP that recovers market structure from binary product evaluations rather than from unsigned shopping baskets. Each game X enters the vocabulary as two tokens, like(X) and dislike(X), corresponding to recommend and do-not-recommend reviews. A reviewer’s set of signed evaluations is treated as a basket; because a user records at most one recommendation per game, a basket cannot contain both tokens for the same title. A neural language model trained on these baskets yields embeddings in which games that are co-recommended, co-rejected, or recommended in opposition occupy different regions of latent space. Dimensionality reduction produces a two-dimensional map with two entries per game. Proximity among like-tokens identifies shared-taste and complementary-recommendation structure; proximity among dislike-tokens identifies co-rejection neighborhoods; cross-polarity proximity identifies audience conflict; and the distance between a game’s like and dislike tokens measures how polarized its evaluators are. Applied to large-scale video-game reviews, the approach maps preference and aversion in a digital entertainment catalog and yields implications for positioning, targeting, and franchise strategy. Prior Product2Vec and P2V-MAP research embeds co-purchase in grocery and e-commerce assortments and does not consider signed, mutually exclusive review tokens.
Future
Abstract
Audit of Recommendation and Citation Biases in Consumer-Facing Large Language Models
Solo project
Large language models are becoming a default layer of product discovery: shoppers ask for advice and receive a synthesized shortlist, a named retailer, and a handful of citations without visiting review sites. This paper reports a large-scale descriptive audit of that new visibility market. We generate hundreds of first-person U.S. consumer personas spanning seven age groups and both genders, then submit the same persona-conditioned shopping request to five consumer-facing assistants (ChatGPT, Claude, Gemini, Grok, and Perplexity). Models return structured recommendations, cited sources, and a single best choice (brand, seller domain, and price). On 691 mutually schema-valid personas (3,455 best-choice observations), brand winners are fragmented—the leading brand is only about 2–3% of picks in every model—while seller attention is highly concentrated and model-specific. Amazon’s share of best-choice seller domains ranges from about 3% in ChatGPT to about 48% in Grok (spread of 45 percentage points; Cramér’s V = 0.35). Overall seller-domain distributions differ more across models than brand or price-tier distributions. In a companion study, the same five models infer sixteen socio-demographic and attitudinal traits from the persona narratives (n = 716 complete five-model records). They agree on structural facts such as homeownership and marital status (about 91–92% five-model agreement) but almost never agree on income, and they assign systematically more liberal, Democratic profiles to female personas than to male ones (mean ideology gap of 0.68 points on a 1–7 scale). The results imply that “LLM optimization” is not a single ranking problem: which retailer captures zero-click demand, and which shopper stereotypes are applied, depends on which assistant the consumer happens to use.
Egor Kudriavtcev
Empowering businesses with data-driven solutions
Contact
Connect
© 2025. All rights reserved.