Machine learning forecasting system deployed on Google Cloud at IATA
One global ML model produces approximately 5,700 bidirectional forecasts
Reduced reliance on manual forecast production and single-person execution
Short-term demand forecasts for traffic outlooks and internal research
Automated monthly pipeline—repeatable operations and less manual intervention
IATA operationalized a Google Cloud-based air travel machine learning forecasting model for internal research, traffic outlook development, and publication workflows.
IATA operationalized a Google Cloud-based air travel machine learning forecasting model for internal research, traffic outlook development, and publication workflows.
From prototype to operational forecasting
IATA—the International Air Transport Association—brings together aviation industry data and analysis to help the sector understand traffic trends. Through reports such as its Monthly Passenger and Cargo Market Analysis, Quarterly Chartbook, and biannual Global Outlook for Air Transport, IATA provides a view of evolving demand across the global air transport market.
Traditional forecasting is usually driven by macro-economic indicators, such as GDP, trade data, long-range projections. What was missing was a pure data-driven approach: something flexible enough to capture short-term patterns that you simply cannot see in those numbers. Machine learning provides that.
Serafeim Loukas
Principal Data Scientist, IATA
Short-term air travel forecasting is especially valuable because demand can shift quickly. Existing economic analysis and industry expertise at IATA largely rely on econometric models that focus on long-term demand using macro-economic input which often update slowly and change infrequently. While the economic models remain central to IATA's outlook work, the opportunity was to complement that expertise with a repeatable, data-driven signal focused on the next three to 12 months.
One way to address this opportunity was machine learning (ML). An ML system can learn directly from historical passenger data at monthly frequency, adapt to recent patterns, and incorporate forward-looking signals such as published airline schedules and booking trends. The question was not whether to build it, but how to build it, so it would actually run reliably month after month, without relying on any single person to execute it manually.
Operationalizing the workflow on Google Cloud
The system IATA built on Google Cloud is deliberately lean and fully automated. Every month, a scheduled Cloud Scheduler job is triggered, activates a Cloud Run Function, and publishes a message to Pub/Sub. The ML pipeline, orchestrated through Gemini Enterprise Agent Platform with Kubeflow, comes to life without a single manual step.
The first stage is data retrieval. The pipeline is authenticated securely using credentials stored in Google Cloud Secret Manager, encrypted by Cloud KMS, and pulls the latest monthly data into Cloud Storage. From there, it moves through preprocessing and advanced feature engineering, incorporating not only historical passenger volumes by route but also forward-looking airline schedule data, booking signals, and event and holiday calendars to help the model anticipate seasonal shifts.
At the core is a forecasting model, trained globally across all 5,700 bidirectional country pairs simultaneously.
Google Cloud is a strategic part of our multi-cloud approach. It enables us to deploy the validated system in a scalable, automated environment, eliminating key bottlenecks such as manual errors and dependence on a single person being available to run the forecast. That ability to operationalize the solution reliably and efficiently was the real reason for moving to the cloud.
Serafeim Loukas
Principal Data Scientist, IATA
The model is trained in bidirectional traffic flows between countries. Once retrained on the latest data, the new model version is registered in the Gemini Enterprise Agent Platform’s Model Registry, ensuring full version control and traceability. Forecast outputs—monthly Revenue Passenger Kilometers (RPK) projections up to twelve months ahead—flow to a business-facing dashboard used by IATA's industry analysis team.
The practical impact is immediate: an IATA economist opens the dashboard and finds a fresh, data-driven short-term forecast already there: consistent, auditable, and ready to be layered into the analysis.
A first step, and a blueprint
We are very excited to use IATA’s first machine learning-based demand forecasting model as an internal tool to complement our existing economic and industry analysis. This short-term forecast model, deployed on Google Cloud in a fully automated manner, will be instrumental for us to evaluate the industry outlook for the next 3-12 months.
Bojun Wang
Senior Economist in Modelling and Forecasting, IATA
For IATA, the significance of this deployment extends well beyond the forecast itself. This is IATA’s first machine learning forecasting deployment on Google Cloud, and that carries real internal weight. The project demonstrated that IATA could move from a proof of concept to a production-grade, automated ML pipeline: operationally robust, version-controlled, and running without manual intervention across 5,700 country pairs every month.
The outputs are designed for internal use and research, giving IATA a data-driven short-term signal to complement its existing macro-economic analysis and publication workflows. The system does not replace economists' judgment. It provides a cleaner instrument for exploring near-term developments, testing assumptions, and supporting informed analysis.
The roadmap is deliberately forward-looking. IATA intends to use this deployment as an operational foundation for future time-series forecasting use cases on Google Cloud.
With the pipeline, model life cycle, and dashboard layer in place, the organization can use this implementation as a blueprint for additional forecasting models for internal use and research.
In terms of next steps, one avenue under consideration is using generative AI to process and synthesize external signals as inputs for future model versions. The architecture is also frontend agnostic: as IATA's tooling evolves, the forecasting layer beneath it remains stable and portable.
The International Air Transport Association (IATA) is the trade association for the world’s airlines, representing over 370 airlines accounting for some 85% of global air traffic.
Industry: Travel
Location: Switzerland
Products: Gemini Enterprise Agent Platform, Cloud Storage, Cloud Scheduler, Cloud Run functions, Pub/Sub, Secret Manager, Cloud KMS
