NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Existence Precedes Value: Joint Modeling of Observational Existence and Evolving States in Time Series Forecasting
TL;DR
Real-world time series data often suffer from missing observations, complicating accurate forecasting. Timeflies is a new framework that jointly models future observability and value estimation in time series.
✦ Why It Matters
Engineers can improve time series forecasting accuracy by incorporating observability modeling into their approaches.
Key Takeaways
- Implement Timeflies for better forecasting in incomplete time series data.
- Use the Shadow benchmark to evaluate forecasting models.
- Adopt the OVJE metric for comprehensive predictability assessment.
- Focus on modeling future observability alongside value estimation.
Full Summary
Time series data in real-world applications frequently have gaps due to issues like sensor inactivity and irregular sampling, making forecasting difficult. Traditional methods typically assume future observation times are known, which is unrealistic in many scenarios.
Timeflies is introduced as a unified framework that reformulates forecasting into a joint problem of predicting both the occurrence of future observations and their values. It employs an observation stream and a value stream, interconnected through modules for reliability-aware embedding, observation-guided dependency modeling, and joint prediction.
A new benchmark, Shadow, combines natural missingness from public datasets with real-world industrial data, and the Observation-Value Joint Entropy (OVJE) metric is proposed for evaluation. Extensive experiments demonstrate that Timeflies consistently outperforms existing forecasting methods, emphasizing the need to model future observability.
This advancement can significantly enhance forecasting accuracy in applications with incomplete data.