ECIS 2025: what inventory planning can learn from language models

SOLISTIQ Team 5 min read

© Qais Tweissi, CC BY-SA 4.0

Large language models have shown that a single pre-trained model can solve many tasks. In our ECIS 2025 paper we carry this idea over to inventory planning and measure success where it counts: by the cost in the warehouse.

Before large language models, almost every language task needed its own model: one for translating, one for summarizing, one for sorting emails. Today a single pre-trained model does all of that, often without additional training.

In inventory planning, by contrast, many companies still build, maintain and regularly retrain a separate model for each product. Our paper “What Can We Learn from LLMs? Building a Foundation Model for Inventory Management” asks what happens when that changes. It appeared in 2025 in the Business Analytics track of the European Conference on Information Systems (ECIS), one of the major international conferences in information systems. It was written by Magnus Maichle, Nikolai Stein and Richard Pibernik.

Why information systems

Our work at NeurIPS and ICLR addresses the machine learning community. ECIS speaks to a different audience. Here the question is less whether a method works and more what it changes in companies and why it often does not reach them yet.

That fits our starting point. Many companies, especially mid-sized ones, still hardly use AI in inventory planning. Not because the methods are missing, but because the effort is too great: preparing data, setting up models per product, maintaining and monitoring them. A foundation model starts exactly there.

What we borrowed from language models

We adopted three ideas. The first is a unified architecture: a transformer, like the one behind language models, processes the sales history of every product, whether fast mover or spare part. The second is a single training run for all products instead of one per product.

The third is called cross-learning. The model learns patterns across the entire assortment, for example how seasonality or intermittent demand typically behave. What it has learned from one product helps it with the next.

This leads to perhaps the most important property: the model can also forecast products it has never seen and for which there is hardly any history. The technical term is zero-shot forecasting. For new products and slow movers, where per-product models reach their limits for lack of data, this makes a big difference.

Measured by cost, not by forecast

How we measure mattered just as much to us. In research, a forecasting model is usually judged by how close it comes to actuals. We measured instead what wrong inventory costs: overstock that ties up capital, and stockouts that cost revenue and customers. In the paper these costs are called inventory distortion costs.

We tested on publicly available sales data from the retail chain Walmart. There our model outperformed several state-of-the-art methods on exactly these costs. Because the data is public, the results can be verified. And because we test later model generations on the same held-out period, we still see today how much better each new generation gets.

From the first version to Q-Foundry

The first version of the model was built in March 2024 in the Data Driven Decisions research group at the University of Würzburg. The ECIS paper describes this first iteration. Today it has become Q-Foundry, pre-trained on the sales histories of around 2.5 million products with more than 1.5 billion data points.

The core idea has remained: instead of building a separate model for every product, a single pre-trained model forecasts every product without retraining, including new and data-poor items.

Above all, we learned that the analogy to language models holds, but not completely. Sales figures are not words. They are often zero, fluctuate with prices and the calendar, and show only what was sold, not what would have been sold without an empty shelf. A foundation model for inventory planning has to take that seriously. And in the end it has to be measured not by the most beautiful forecast, but by the best order.

How Q-Foundry delivers forecasts for every service level and how it recognizes when it is uncertain is the subject of our articles on NeurIPS 2024 and ICLR 2025. Want to know what a foundation model can do for your assortment?
Talk to us.

Further reading

  • Maichle, M. J., Stein, N. & Pibernik, R. (2025). What Can We Learn from LLMs? Building a Foundation Model for Inventory Management. ECIS 2025 Proceedings, 11. aisel.aisnet.org
  • Bommasani, R. et al. (2021). On the Opportunities and Risks of Foundation Models. arXiv:2108.07258
  • Brown, T. et al. (2020). Language Models are Few-Shot Learners. NeurIPS 2020. arXiv:2005.14165
  • Ansari, A. F. et al. (2024). Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research. arXiv:2403.07815
  • Das, A., Kong, W., Sen, R. & Zhou, Y. (2024). A Decoder-only Foundation Model for Time-series Forecasting. ICML 2024. arXiv:2310.10688
  • Makridakis, S., Spiliotis, E. & Assimakopoulos, V. (2022). M5 Accuracy Competition: Results, Findings, and Conclusions. International Journal of Forecasting 38(4), 1346–1364. doi.org/10.1016/j.ijforecast.2021.11.013
  • Bertsimas, D. & Kallus, N. (2020). From Predictive to Prescriptive Analytics. Management Science 66(3), 1025–1044. doi.org/10.1287/mnsc.2018.3253

Another look behind the scenes.