WARA AI TRICS: Time Series Challenge
WARA AI TRICS aims to bring real-world time series problems to the academic community, recognizing their cross-domain relevance in modern industrial, telecommunication, and financial systems. To foster methodological innovation in this space, we present the first WARA AI TRICS Open Challenge for PhD students. This challenge is open to PhD student teams within WASP, with all required computational infrastructure and compute resources provided by WARA AI TRICS. Participating teams will have a 3-month development period. Dates will be announced shortly and the winners will be announced at the Winter Conference.
Traditional time series forecasting models often operate primarily on the historical observations of a target signal. However, in complex real-world dynamics, a single time series rarely contains enough information to explain all structural shifts, tail events, or localized anomalies. A sudden deviation, peak, or drop in a target signal is frequently driven by underlying contextual regimes, interactions with related signals, or external multimodal information.
This challenge investigates whether incorporating such auxiliary context can improve downstream time series prediction. Participants are encouraged to develop frameworks that combine primary target signals with heterogeneous contextual information, including related time series, cross-signal interactions, and multimodal data. The objective is to evaluate whether and to what extent this additional context improves predictive performance, particularly under non-linear dynamics, extreme events, regime changes, and other difficult prediction regimes.
To demonstrate the applicability and generality of this paradigm, participants will tackle solutions across three distinct, real-world domain case studies:
- Quantitative Finance (SEB): Modeling structural asset price deviations using an auxiliary regime variates dataset containing macroeconomics, rates, and volatility signals.
- Telecom Systems (Ericsson): Forecasting critical network Key Performance Indicator (KPI) behavior by capturing spatial, temporal, and contextual interactions across neighboring cells, concurrent KPIs, alarms, logs, and topology.
- Industrial Machines Monitoring (IPercept): Anomaly detection in repetitive industrial robot motion using time series foundation models. A robotic system executes a structured motion pattern under controlled, repeatable conditions. Participants build models that learn normal motion from a large unlabeled corpus and detect mechanical degradation patterns in a held-out evaluation window.