Foundation Models in Forecasting: Are We There Yet? Lessons from the Trenches

Dr. Irena Bojarovska

Generative AI & Synthetic Data
Python Skill Intermediate
Domain Expertise Intermediate

I. The Promise vs. The Reality (10 min)

  • The hype cycle: A brief overview of the current time-series foundation model landscape.
  • The "zero-shot" myth: Why "out-of-the-box" performance often hits a ceiling in e-commerce due to a lack of domain context (discounts, commercial activities).
  • Between two extremes: We experimented with stable market-level KPIs and noisy article-level demand — to see where time-series foundation models succeed or fall short versus tuned classical models.

II. Lessons from the Trenches: What Works and What Doesn’t (15 min)

  • The setup: We run daily and weekly forecasts across 20+ markets at two horizons (5 and 16 weeks), evaluating KPIs like GMV, items sold, cancellation rate, plus article-level forecasts for established best-sellers and newly launched items.
  • The "win" column: Scenarios where foundation models reduced MLOps overhead and delivered surprisingly strong zero-shot baselines (lower WAPE (weighted absolute percentage error) across 52 rolling validation rounds).
  • The "fail" column: Why foundation models still struggle with rigid non-linear business constraints and "rare event" sensitivity (e.g., cold-starts for new SKUs).
  • The “stability” column: How foundation model forecasts evolve as the target date approaches: do predictions converge and uncertainty shrink smoothly, or do they jump around unpredictably?

III. Summary: The 2026 Roadmap (5 min)

  • Are we there yet? Our final conclusion on whether foundation models are ready to be the primary engine or remain a strong supporting model.
  • Final advice: A checklist for teams deciding where to start their foundation model journey.

Dr. Irena Bojarovska

Irena Bojarovska is an Applied Scientist at Zalando SE, focusing on time‑series forecasting and demand prediction across 24+ markets.

Originally from Macedonia, she earned a BSc and an MSc in Applied Mathematics and Computer Science in Russia and a PhD in Applied Harmonic Analysis from TU Berlin. She began her industry career as an analyst at Air Berlin and, since 2017, has worked on causal inference for marketing, automation, demand forecasting, hierarchical reconciliation, and time‑series foundation models at Zalando. Outside work she leads a math circle for children at Lyzeum 2 and enjoys spending time with her family.