Analytics & FORECASTING 2025

September 29 – 30, 2025 | Geffen Hall, Columbia Business School, 645 W. 130th Street| In-Person & Livestream
9am-5pm on 9/29; 9am-12:30pm on 9/30

EVENT CONTENT IS NOW AVAILABLE.

Call for Content is Open!

We are currently accepting academic submissions from MSI academic members for this event. Find out more here.

Overview

Synthetic data—artificially generated data that mirrors real-world patterns—is gaining traction as a tool for enhancing market research while addressing privacy and compliance challenges. It is already being used in concept testing, product development, packaging and pricing studies, market sizing, and consumer trend research.

For market researchers, its value lies in delivering scalable, rapid, and privacy-compliant insights, particularly when engaging with niche or hard-to-reach segments. It also supports continuous, always-on research through digital twins and real-time simulations.

In forecasting, synthetic data helps to mitigate issues arising from data sparsity and signal loss due to privacy restrictions and cold start challenges. However, its ability to model complex human decisions and emotional nuances remains uncertain. Challenges persist regarding data quality, ethical use, bias reinforcement, and regulatory inconsistency. There is also a critical need for industry-wide standards and validation frameworks. Significant questions remain: Can synthetic data replicate the richness and unpredictability of real human behavior sufficiently to support high-stakes forecasting and decision-making? Where is the line between synthetic realism and synthetic bias, and how do we manage it? Is synthetic data a bridge—or a barrier—to more effective and reliable forecasting models?

MSI and the ARF are excited to co-produce the ANALYTICS & FORECASTING conference for the first time. It will be held at Columbia University’s Business School on September 29-30. This is an opportunity to engage in discussions about the opportunities and limitations of modeling in market research applications and to showcase innovative work from academia and the business sector.

Speakers

AGENDA

MONDAY, SEPTEMBER 29

TUESDAY, SEPTEMBER 30

MONDAY, SEPTEMBER 29

TUESDAY, SEPTEMBER 30

Register

Click here for a list of nearby hotels.

This is a member only event and there is no charge for members. Click here to learn more about membership.

8:00-9:00am	Registration & Breakfast
	Using Generative AI for Marketing Intelligence
9:00-10:00am	Opening Remarks & Plenary Sessions
	The Next Era of Marketing Mix While annual planning remains essential for any business, it is no longer realistic to assume that planning and measurement can be done only once a year for brands. Managing the Marketing Mix and planning for the upcoming year has shifted from an annual task to an on-demand capability, available now and in the near future. To succeed, you must be accurate and actionable, as well as flexible and agile. The mix of art and science is becoming a crucial part of Marketing Mix and measurement overall, with both becoming more interconnected to support the alignment of multiple overlapping metrics. How can brands use the latest technology and top-tier data and modeling to turn Marketing Mix from an annual project into a core capability? Jonathan Dizney – Mix Director, Circana Troy Noble – VP, Global Measurement Product Lead, Circana
	Predicting Purchase Intent in Customer Interactions with AI Assistants: Context-Specific Small Language Models vs. LLMs This study explores how to predict purchase intent during consumer-initiated interactions with AI assistants—both within and outside the purchasing process. As AI assistant adoption rises, identifying purchase intent becomes essential for targeted advertising, especially since many interactions occur outside the traditional purchase funnel. Both methods accurately predicted purchase intent, but the graph-based model performed better for monetizable keywords, demonstrating that specialized “small language models” can outperform LLMs for niche tasks. These findings provide marketers with practical guidance on when and how to advertise via AI assistant channels. Wendy Moe – Dean’s Professor of Marketing, University of Maryland and Amazon Scholar
10:45-11:15am	Morning Break
	Digital Twins for Synthetic Market Data
11:15am–12:45pm	Plenary Sessions
	Twin-2K-500: A Dataset for Building Digital Twins The rise of large language models (LLMs) like GPT has sparked interdisciplinary interest in using “silicon samples” to emulate human responses, affecting both research and practical applications. If these simulations prove reliable, they could improve experimental design, theory development, and customer insights. An increasing number of companies now offer LLM-based tools for these purposes. A high-quality dataset is presented with over 500 questions and 2,000 respondents, demonstrating strong test-retest reliability and replication of known effects. The main focus is on building digital twins, which initially predict human behavior with up to 88% accuracy in test-retest measures. Although encouraging, there are limitations, such as issues with non-normative behavior, representational diversity, and the dataset’s U.S. focus. Still, this resource seeks to accelerate both academic and commercial use of LLM simulations. Olivier Toubia – Glaubinger Professor of Business, Columbia Business School
	Predicting Behaviors with LLM-Powered Digital Twins of Consumers Digital twins of consumers have emerged as a promising approach to simulate consumer thinking, feeling, and decision-making. Grounded in the psychological theory, which conceptualizes behavior as a function of both personal traits and contextual factors, this research proposes and validates a dual-component framework for constructing Large Language Model (LLM)-based consumer digital twins. Fine-tuning on consumer-specific data, including user-generated content, allows the model to internalize individual traits, preferences, and cognitive and behavioral patterns. At the same time, retrieval-augmented generation (RAG) dynamically incorporates information specific to consumer context at inference. By aligning LLM adaptation techniques with foundational psychological theories about behavior, this method enables psychologically grounded simulations of individual-level consumer behavior at a scale. This research contributes to the literature on generative AI, synthetic agents, and digital twins in consumer research and, at the same time, offers a new methodology for theory-driven modeling and privacy-compliant personalization in practice. Shane Wang – Professor of Marketing, Pamplin College of Business, Virginia Tech
12:45 – 1:45pm	Lunch
	Extracting Consumer Preferences
1:45 – 3:50pm	Plenary Sessions
	Panel Synthetic Data: Pros and Cons for Market Research Synthetic data enables researchers to quickly create large, privacy-safe datasets, simulate rare consumer behaviors, and test hypothetical product scenarios. It can also help balance samples and cut data collection costs. However, synthetic data might not fully reflect real-world consumer complexity, risk introducing biases, and be difficult to validate. Insights based solely on synthetic datasets may overfit or fail to generalize, making careful testing against real data essential. Hear a variety of perspectives about the benefits and drawbacks of using synthetic data. Rajan Sambandam – President, TRC Insights Ayelet Israeli – Marvin Bower Associate Professor, Harvard Business School Neeraj Arora – Arthur C. Nielsen, Jr. Chair, Marketing Research and Education, University of Wisconsin-Madison More speakers TBD Moderator: Oded Netzer – Vice Dean of Research and Arthur J. Samberg Professor of Business, Columbia Business School
	Replicas & Realities: The Strategic Role of Consumer Twins Digital consumer twins are unlocking new possibilities in innovation analytics, from spotting early winners to tailoring concepts for targeted segments. Drawing on Colgate-Palmolive’s CPG experience, this session explores where these models deliver the most value, where they fall short, and how they can shape a more effective learning strategy. Katy Qian – Data Scientist, Colgate-Palmolive Company Helen Wolf, Ph.D. – Senior Director, Global Consumer Experience Insights, Colgate-Palmolive Company
	Semantic Targeting This research presents a methodology for real-time, personalized targeting by creating dynamic customer representations that integrate diverse behavioral data. The approach combines LLM-based item embeddings with attention mechanisms to address data sparsity and fragmentation. Evaluated on real-world digital advertising data, the attention-based model outperforms simple averaging, especially for long-term or abstract predictions, producing stable and generalizable customer embeddings. This work advances marketing research by adapting LLMs to customer data and provides practitioners with a scalable, flexible foundation for targeting systems that adapt to evolving consumer behavior. Isamar Troncoso – Assistant Professor of Business Administration, Harvard Business School
3:20-3:50pm	Afternoon Break
3:50-5:00pm	Plenary Sessions & Closing Remarks
	Optimal Product Design Synthesis: Pairing Generative Models with Adaptive Preference Measurement This research introduces a preference measurement framework that combines generative AI with adaptive survey design to guide product development based on individual consumer preferences. Using Stable Diffusion to generate realistic product images on-the-fly and high-dimensional Bayesian optimization to learn preferences efficiently, the method goes beyond traditional approaches like conjoint analysis. It captures nuanced aesthetic preferences—colors, textures, shapes—and enables automatic generation of designs aligned with consumer tastes. The framework can reduce design costs, accelerate development, and create products better matched to consumer preferences across industries such as fashion, automotive, and creative goods. Ryan Dew – Assistant Professor of Marketing, The Wharton School, University of Pennsylvania
	LLM Time Machines: Valuing Brands Over Time This research explores using Large Language Models (LLMs) like ChatGPT to estimate brand value over time, especially for free digital services such as social media platforms. Traditional methods are expensive and static, failing to reflect dynamic changes in brand equity. By benchmarking LLM-generated valuations against annual data from incentive-compatible discrete choice experiments, LLMs can accurately mimic human choices and valuations. This approach allows for retrospective brand value estimation and provides predictive insights into future trends, highlighting shifts linked to major events and changes in consumer behavior. The method gives marketers a scalable, practical tool to monitor and forecast brand equity, improving strategic brand management. Felix Eggers – Professor, Copenhagen Business School
	Closing Remarks Scott McDonald, Ph.D.– President & CEO, ARF
5:00 – 6:30pm	Cocktail Reception

8:00-9:00am	Registration & Breakfast
	Simulating Consumers
9:00-10:40am	Opening Remarks & Plenary Sessions
	From Noise to Signal: When Synthetic Data Adds Real Value Information theory, decision science, and structural analysis offer a principled approach to determining where synthetic data provides the greatest value in market research. By quantifying tradeoffs across analytical accuracy, economic impact, representativeness, and privacy, Dynata can prioritize use cases where synthetic data meaningfully enhances insight, fills gaps, or supports inference. They will illustrate this framework with examples showing how entropy, mutual information, and scenario modeling guide the selective deployment of synthetic data for segmentation, rare-event simulation, and feasibility forecasting. Alan Briancon, Ph.D.– Vice President of Research & Data Science, Dynata
	Using LLMs for Market Research This paper examines how LLMs can help researchers and practitioners understand consumer preferences. Focusing on the distributional nature of LLM responses, multiple outputs are generated per survey question and present two key findings. First, willingness-to-pay estimates derived from LLM responses are realistic and comparable to human studies. Second, that fine-tuning LLMs with prior survey data improves alignment with human responses for existing and new product features, though not for entirely new categories or segment differences. These results highlight when and how LLMs can meaningfully support product development and market research. Ayelet Israeli – Marvin Bower Associate Professor, Harvard Business School
	Synthetic Data in Action: Ipsos’ Human-Centered Approach to AI Research  Synthetic data—especially through tools like PersonaBots and digital twins—is transforming market research by enabling faster project turnaround times, greater efficiency, and a powerful new way to access insights. In partnership with Stanford, Ipsos is building a ground-truth database to validate these AI-generated responses. While synthetic participants offer real-time interaction and can simulate hard-to-reach segments, a human-in-the-loop, ethically grounded approach is necessary to ensure accuracy and accountability. The potential is significant, but only with transparent methods for validation and thoughtful strategies to blend human and synthetic respondents. Stephanie BannosRyback – Head of Business Transformation & AI Change Management, Ipsos Xufeng Wang, Ph.D. – Partner & Chief Data Scientist, Data Labs, Ipsos
10:40-11:00am	Morning Break
11:00am–12:10pm	Plenary Sessions & Closing Remarks
	Simulating Real Consumers: How Accurately Can LLMs Mimic Individual Respondents in Market Research? This study explores whether Large Language Models (LLMs) can emulate specific individuals for market research. Unlike previous work that focused on group-level preferences or stylized personas, detailed socio-economic, psychographic, and personality data from real people are integrated into an LLM to create synthetic counterparts. The results demonstrate that synthetic respondents can replicate individual consumers with high accuracy. However, systematic biases still exist. The findings suggest that LLM-based synthetic personas cannot yet replace traditional market research but can serve as scalable, fast, and insightful tools when used carefully. Leonard Kinzinger, M.Sc. – Technical University Munich (TUM)
	Closing Remarks

Overview

Speakers

AGENDA

Register

SponsorS

Related Content

How to Calculate Reach and Frequency Using Virtual IDS

Predicting Attention to Advertising Through Machine Learning

ARF Attention Measurement Validation Initiative Phase 2