Gurobi’s Jerry Yurchisin: a prediction is not yet a decision

Jerry Yurchisin, Senior Data Science Strategist at Gurobi Optimization, works with organizations to apply mathematical optimization to business decisions. Before joining Gurobi, he worked on government and commercial projects at Booz Allen Hamilton. His interest in optimization grew while he studied statistics and operations research at the University of North Carolina at Chapel Hill. He sees the two disciplines as answering different questions: statistics and machine learning help predict what might happen, while optimization determines what to do with those predictions.

Yurchisin finds that many data science teams build sophisticated prediction models but have little involvement in the decisions that follow. Their work often ends with a report or dashboard, leaving managers to determine the next action. He encourages data scientists to extend their involvement into decision-making by modeling business objectives, available choices, and constraints. In one graduate project, he used statistical methods to predict when water infrastructure would deteriorate, then applied optimization to develop a replacement schedule.

Yurchisin believes optimization is more accessible than many practitioners realize. Introductory resources often emphasize advanced mathematics and complex problems, even though simpler models can deliver value. His first professional project used optimization to help the U.S. Coast Guard evaluate crew arrangements and reduce deferred maintenance. The model required no advanced techniques but provided useful operational insights.

At Gurobi, he has helped develop introductory optimization courses for data scientists who already understand Python and machine learning. The courses start with realistic business problems, letting participants build working models before studying more advanced techniques. He also encourages practitioners to revisit optimization problems they abandoned years ago because they took too long to solve. Improvements in computing hardware and optimization software mean that models considered impractical five or ten years ago may now be solvable.

Yurchisin believes generative AI can help experienced practitioners build optimization models faster by translating requirements into mathematical formulations and generating code. However, language models may overlook earlier constraints, while optimization models must satisfy every hard constraint to produce a feasible solution. This matters when decisions involve regulations, budgets, or capacity limits.

Yurchisin encourages operations research practitioners to participate actively in AI discussions. He argues that machine learning, generative AI, simulation, and optimization serve different purposes and work best together. He sees opportunities for greater use of optimization in financial services, where regulatory and operational constraints make decisions more complex.

Benjamin Baer, Founding Member of the DecideWise community, warns that HARK’ing, or Hypothesizing After the Results are Known, can undermine enterprise decision intelligence. It occurs when analysts discover patterns in data and present explanations developed afterward as if they were the original hypotheses. He argues that building automated rules around untested correlations creates false confidence, increases model drift, and wastes resources on unreliable signals. Baer recommends separating exploratory analysis from validation, testing findings against independent datasets, and documenting hypotheses, expected outcomes, and success metrics before deployment. He also advocates using simulations and stress tests to determine whether decision rules remain reliable under changing conditions.

Organized by Marcus Evans, this two-day conference examines how pharmaceutical companies use generative AI, agentic AI, and real-world evidence to improve commercial forecasting and strategic planning. Sessions cover AI-powered forecasting, rare disease and product lifecycle planning, scenario modeling, and data governance. Speakers from Novartis, UCB, BioMarin, CSL, Ferring Pharmaceuticals, and Lundbeck will share case studies through panels, roundtables, and workshops.

The meeting takes place at Moscone Center South and the San Francisco Marriott Marquis, bringing together more than 6,000 researchers, industry experts, students, and employers to explore operations research and analytics. The in-person program includes technical presentations, poster sessions, technology showcases, and exhibitor workshops, with a dedicated AI poster session on November 2. A career fair and community events offer opportunities to recruit specialized talent and connect with researchers and practitioners.

📊 Report Spotlight: Top Strategic Technology Trends for 2027 (Gartner)

Gartner's report identifies 10 technologies shaping enterprise IT across three areas: nontextual AI, autonomous agent governance, and resilient infrastructure. It forecasts that large quantitative models will replace LLMs in 25% of strategic forecasting and operational modeling tasks by 2030, while speech reasoning models will handle over 70% of enterprise voice interactions. Organizations that prioritize semantic context for AI agents could improve accuracy by up to 80% and cut costs by up to 60% by 2027. Gartner also predicts that AI agent abuse will contribute to 25% of enterprise security breaches by 2028.

Join the conversation: Can superintelligent AI be trusted to police itself?

Join the DecideWise discussion on whether superintelligent AI can safely monitor and control other superintelligent AI systems. Explore the idea that equally capable AI could prevent dangerous behavior, the risks of relying on AI to police AI, and whether such oversight would be enough to protect society. Share your perspective on who or what should remain in control.

Mastercard Services CTO Nirav Mehta explains how Decision Intelligence Pro uses AI models to evaluate transactions in milliseconds, preventing about $13 billion in fraud through early October 2026. Mastercard's Data Commercialization Platform reduced location-data delays from 8-10 weeks to hours and processing time from weeks to under 10 minutes. Mehta also discusses machine-to-machine payments, unresolved identity challenges in agentic commerce, and how smaller engineering teams can work alongside AI agents.