Benchmarking Pretrained Forecasting Services for River Flow Prediction

Overview: Help us evaluate the performance, usability, and adaptability of modern forecasting services, like Nixtla, Sulie, TimesFM and Chronos, that offer pretrained time series models. You'll assess how well these services perform on river flow data, how easily they can be fine-tuned, and whether they can replace more complex in-house LSTM pipelines for operational forecasting. We aim to test at least two of the services mentioned, depending on API/model availability. This internship bridges real-world hydrological modelling with cutting-edge machine learning services, providing insights into when it’s better to build custom models versus adopt pretrained platforms.

Key challenges:

• Explore and benchmark pretrained forecasting platforms using river flow data
• Run zero-shot and fine-tuned evaluations across multiple catchments and regimes (e.g., rainfall-driven vs snowmelt)
• Assess forecast quality across multiple horizons (+1h to +24h) using standard metrics (MAE, CRPS, NSE) alongside hydrologically relevant ones such as DTW for hydrograph shape alignment and peak error for flood timing and magnitude.
• Analyse ease of integration, API usability, and support for covariates like precipitation, evaporation, upstream flow, groundwater levels
• Compare cost vs. control: when is a plug-and-play service “good enough” — and when does in-house modelling still win?

You'll learn about:

• Working with modern time series platforms
• Evaluating model performance across forecast horizons and physical regimes
• Applying domain knowledge to test hydrologically relevant prediction tasks
• Model operations: deployment, tuning, scaling, and explainability trade-offs
• Making technical recommendations grounded in usability, cost, and transparency

Ideal for: A student looking for a 3-month internship, interested in applied time series modelling, machine learning operations and benchmarking real-world ML systems. You’re curious about the tension between automation and control — and want to help figure out whether pretrained foundation models can forecast rivers as well as hand-built LSTMs.

Contact: Hans Korving and Tom Heskes