The Fifth Elephant 2026 Winter Edition

Submission type:
Talk/session proposal - 30-40 mins

Describe your session in 2 paragraphs.

Ad inventory forecasting sounds straightforward until an advertiser asks: “How many impressions can I buy for a specific page type, ad position, device, geography, and audience next month; and can you guarantee that delivering those impressions won’t compromise existing campaign commitments?” For retail media networks, answering this question requires much more than predicting traffic. Every new campaign competes for a finite pool of impressions, with overlapping targeting, different priorities, pacing strategies, and existing reservations. This session introduces Foresight, an in-house inventory forecasting and decision intelligence platform being developed for Roundel, Target’s Retail Media Network. Inspired by Google Ad Manager’s forecasting capabilities, Foresight is designed to support three critical business decisions: pre-launch media planning to determine how much inventory can be confidently sold; in-flight campaign optimization to identify delivery risks and simulate corrective actions; and incremental opportunity identification to uncover additional inventory that can support advertiser growth and revenue.

The technical challenge spans three interconnected problems: forecasting future inventory supply, modeling competition for that inventory, and estimating campaign deliverability under real-world constraints. Foresight combines autoregressive and classical machine learning models to forecast impression supply across dimensions such as page type, position, device, and app/web surface, with targeting adjustments for geography and audience segments. Beyond supply forecasting, the system is being designed to account for overlapping campaign demand, priorities, pacing, and reservations to distinguish inventory that is genuinely available from inventory that could be delivered only by impacting other campaigns. Its architecture separates computationally intensive offline forecasting and demand modeling from low-latency, on-demand scenario evaluation, enabling planners to test changes to targeting, campaign duration, priority, or impression goals. We have already achieved forecast accuracy comparable to Google Ad Manager for more than 40% of traffic, and are extending the platform toward scalable, competition-aware forecasting that supports both new and existing campaigns. The session will share the modeling decisions, architectural trade-offs, and practical lessons involved in building a forecasting system that moves beyond predicting impressions to enabling better inventory allocation and monetization decisions.

Takeaways from the session:

1. A practical framework for decomposing ad inventory forecasting into baseline supply forecasting, targeting adjustment, and reservation/availability logic.

2. Lessons from building an internal forecasting engine that combines autoregressive models, classical ML, and pattern-specific modeling classes to reproduce ad-server-grade forecasts at scale.

Which audiences will benefit most from your session?

This session will be useful for data scientists, ML engineers, analysts, ad-tech practitioners, retail media teams and product/data leaders who work with constrained inventory, marketplace supply, campaign planning, or decision systems. It will also benefit people interested in applied forecasting where the hard part is not only model choice, but data definitions, granularity, sparse slices, low-latency serving needs, business constraints, and explainability.

Bio:

Yogesh Soniwal is a seasoned data scientist with 12+ years of industry experience across medical engineering and retail media domains. He currently works at Target, where he leads Roundel’s foundational capabilities team, solving problems across campaign measurement, agentic insight generation, and forecasting.

Draft slides:
Not available yet. I can share a short draft deck before the event with anonymized examples, architecture diagrams, model comparison results, and synthetic data illustrations.

What don’t you know yet, or what do you want help with?

I would like critique on the right modeling abstraction for this problem. Should granular ad inventory forecasting be treated primarily as a hierarchical time-series problem, a constraint optimization problem, a simulation problem, or a hybrid of all three? I am also looking for examples from people who have handled sparse targeting slices, audience overlap, retail-event seasonality, high-throughput forecasting systems, and availability forecasting under existing reservations or campaign priorities.

match making tags

I can help with: experience, technique, critique
I need help with: ideas, experience, technique, critique
I’d like to meet: practitioners, domain experts, tool builders, dataset owners

Topic/domain tags:
forecasting, retail media, ad-tech