Yohei Nishimura

I am a PhD candidate in Quantitative Marketing at Wisconsin School of Business, University of Wisconsin-Madison. I will be on the job market in 2027 2026. My research interests span AI, quantitative modeling, and applied & empirical research, with a particular emphasis on work that carries real-world, industrial impact. I am broadly open to exploring how these interests might align with different research environments and institutional missions.

My research develops and applies quantitative methods at the intersection of quantitative modeling, machine learning, and Bayesian statistics to address substantive problems in marketing. I build economic models that integrate unstructured data—images and text—processed through deep-learning pipelines (computer vision and NLP) into economically interpretable frameworks, and employ Bayesian methods for estimation, prediction, and optimization.

Prior to my PhD, I accumulated 15+ years of industry experience in product management, software engineering, and business development, which informs my research with practical perspectives on digital platforms and marketing operations.

  • Position: Business PhD Candidate
  • Location: Madison, WI, USA
  • Citizenship: Japan
  • Email: ynishimura　@　wisc.edu

Journal Publications

Leveraging Generative AI for Visual Content in Digital Advertising

Marketing Science (forthcoming)

with R. Daviet

Abstract
  • Generative artificial intelligence (AI) for image synthesis has the potential to transform the digital advertising industry. However, a wide range of uncertainties persists regarding its integration into traditional advertising processes, including finding effective implementations, training methodologies, and achievable performance gains. Specifically, two core challenges limit its practical adoption: a search problem of finding high-performing visuals in a vast creative space, and an alignment problem of ensuring brand and campaign compatibility. This paper proposes a novel end-to-end framework that combines a generative AI with two predictive Bayesian neural networks to identify high-performance and brand-acceptable visuals. We develop a cost-effective Bayesian active learning approach solving simultaneously the dual objectives of performance and alignment. We test the framework in a live advertising campaign for an outdoor activities company. Our system generated a portfolio of visuals achieving a higher mean click-through rate and more consistency (lower variance) than creatives from both a professional human designer and a competing AI model optimizing purely for aesthetics. This research provides a validated methodology that bridges the gap between the theoretical potential of generative AI and its practical application, offering a cost-effective solution to the critical search and alignment problems in creative design.

AI-Human Hybrids for Marketing Research: Leveraging LLMs as Collaborators

Journal of Marketing (Vol 89, 2025), Winner of AMA/Marketing Science Institute/H. Paul Root Award 2025

with N. Arora and I. Chakraborty

Abstract
  • The authors' central premise is that a human-LLM (large language model) hybrid approach leads to efficiency and effectiveness gains in the marketing research process. In qualitative research, they show that LLMs can assist in both data generation and analysis; LLMs effectively create sample characteristics, generate synthetic respondents, and conduct and moderate in-depth interviews. The AI-human hybrid generates information-rich, coherent data that surpasses human-only data in depth and insightfulness and matches human performance in data analysis tasks of generating themes and summaries. Evidence from expert judges shows that humans and LLMs possess complementary skills; the human-LLM hybrid outperforms its human-only or LLM-only counterpart. For quantitative research, the LLM correctly picks the answer direction and valence, with the quality of synthetic data significantly improving through few-shot learning and retrieval-augmented generation. The authors demonstrate the value of the AI-human hybrid by collaborating with a Fortune 500 food company and replicating a 2019 qualitative and quantitative study using GPT-4. For their empirical investigation, the authors design the system architecture and prompts to create personas, ask questions, and obtain responses from synthetic respondents. They provide road maps for integrating LLMs into qualitative and quantitative marketing research and conclude that LLMs serve as valuable collaborators in the insight generation process.

Working Papers

Job Market Paper: Better Matches or Better Actions? Decomposing Welfare Gains on a Two-Sided Platform

Solo Project

Abstract
  • Every two-sided platform faces a basic design question: what information should each side see about the other? Platforms routinely foreground easy-to-measure individual metrics such as demographics while the match-specific synergy that drives outcomes remains unmeasured. Quantifying the welfare cost of this gap is challenging because observed outcomes confound two channels: a pair's inherent fit and the strategic actions taken after matching. Since post-match actions are themselves endogenous to matching, the platform cannot recognize whether its primary lever for welfare improvement lies in reorganizing partner formation or in guiding behavior after the match. Using proprietary data from a micro-influencer marketing platform, this paper asks whether the welfare loss in the decentralized market is driven more by which pairs form, or by how matched pairs act afterward. I develop an empirical framework that jointly models two-sided selection and endogenous content adjustment and simulate a centralized counterfactual that decomposes welfare gains into match reallocation and post-match adjustment. The decomposition is striking: content-adjustment optimization accounts for 96.1% of the net welfare gain, while reshuffling who matches with whom contributes a mere 3.9%. The dominant source of welfare loss is influencer over-adjustment; absent any benchmark for how much adaptation a campaign actually requires, influencers adapt their visual style too aggressively and erode net engagement. This implies that platforms should pair improved matching with explicit, pair-specific guidance leading the appropriate level of post-match action.

Working in Progress

Generative Priors for Conjoint Analysis: Leveraging LLMs for Data-Scarce Estimation

with N. Arora

Abstract
  • Reliable individual-level part-worth estimation in conjoint analysis typically requires many choice tasks per respondent, limiting field feasibility. We propose using large language models (LLMs) to construct personalized informative priors for individual-level Bayesian estimation. Diverse synthetic personas are simulated by an LLM to yield preliminary part-worths, and a lightweight MLP is trained to map persona text embeddings to part-worth vectors; at estimation time, each real respondent's embedded profile produces a personalized prior that anchors inference from the first response. In simulations spanning 27 conditions (noise level × description depth × question count), the method reduces MSE relative to standard hierarchical Bayes by 30-39% on average, with significant gains in every condition. A key requirement is that synthetic training embeddings lie on the same manifold as real persona embeddings, for which we develop a sampling procedure.

Separating Negotiation Costs from Bargaining Power: Evidence from a B2B Platform

with M. Ishihara

Abstract
  • We study price formation on a business-to-business wholesale platform that offers sellers two channels to transact with retailers: a posted price (expressed as a discount rate off the manufacturer's suggested retail price) and a private chat-based negotiation channel. We first document a set of stylized facts: despite the availability of buyer-specific negotiation, the vast majority of sellers employ uniform posted prices, and the few who adjust prices do so through discrete discount tiers rather than continuous, buyer-specific bargaining. Motivated by these facts, we propose a structural model that separately identifies buyers' negotiation costs and bargaining power — two parameters that are typically confounded in observed transaction prices. Our identification strategy exploits three features unique to this platform: (i) the simultaneous observability of posted-price purchases and negotiated purchases for the same product, (ii) clickstream data that reveal buyers' outside options through their browsing of competing sellers, and (iii) detailed chat logs that allow us to reconstruct the sequential bargaining process.

Education

Doctor of Philosophy in Business, specialization in Quantitative Marketing

2023 - Present

University of Wisconsin-Madison, Madison, WI, USA

Master of Science, Computer Sciences

2021 - 2023

University of Wisconsin-Madison, Madison, WI, USA

Bachelor of Engineering, Computer Science

2019 - 2021

Teikyo University, Tokyo, Japan

Bachelor of Art, Economics

2003 - 2007

The University of Tokyo, Tokyo, Japan

Selected Professional Experience

Software Engineer (Full Stack), Machine Learning Engineer

Oct 2021 - Feb 2022

Aoyama Art, Inc., Tokyo, Japan (Remote, Part-time)

  • Launched and maintained the media part of titel.jp.
  • Constructed new databases and an infra layer API to connect the database to the front end by a clean architecture and RESTful API.
  • Installed an article recommendation system. Built the recommended article component as a template, and also created new views and handlers. Using a naive Bayesian classifier, the same classified articles are displayed as recommended articles.

Senior Product Manager / Project Leader

Jul 2020 - Jun 2021

Mercari Inc., Tokyo, Japan

  • Responsible for introduction of business users into Mercari service: "Mercari Shops".
  • Developed 10 sets of public APIs with five backend engineers. Drove product vision, go-to-market strategy, and design discussions.
  • Launched the new service with engineers, designers, other PdMs, lawyers, and financial specialists.

Chief Operating Officer / Software Engineer / Product Manager

Dec 2017 - Jun 2020

Michael Inc., Tokyo, Japan

  • Responsible for products' management on CARTUNE (Apps / Web) and CARTUNE Parts Market.
  • Launched a new media service named "CARTUNE MAGAZINE" using Python and Google's services with two engineers and developed the UI/UX design by myself.
  • Monthly active users grew from 20,000 to 2,700,000 in two years.
  • Michael Inc. was acquired by Mercari Inc. in Oct 2018.

Project Leader / Product Manager

Apr 2016 - Nov 2017

Donuts Ltd., Tokyo, Japan

  • MixChannel renewal from a short movie app to a live streaming app as a project leader with 30+ people.
  • Developed the marketing strategy, and released TV commercial.
  • Planned annual P/Ls and product strategies.

Skills

Versatile methodological toolkit spanning structural econometrics, deep learning, and Bayesian statistics for quantitative marketing research.

Structural Modeling & Econometrics

  • Discrete choice models
  • Counterfactual simulation & policy evaluation
  • Selection bias correction

Deep Learning (Vision & Text)

  • Computer vision (image embeddings, visual feature extraction)
  • Natural language processing (text embeddings, LLMs)
  • Generative AI (image generation, synthetic data)
  • Variational auto-encoder

Bayesian Statistics

  • Markov Chain Monte Carlo methods
  • Bayesian optimization / Active learning

Additional Methods

  • Reinforcement learning / Multi-Armed Bandits
  • Causal inference / Quasi-experimental methods
  • Conjoint analysis

Programming & Tools

  • Python (PyTorch, NumPy, Pandas)
  • Julia
  • R
  • SQL / BigQuery
  • GCP / Cloud infrastructure