Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 10·5 min read
KReF: Training-Free Retrieval for Long-Term Time-Series Forecasting and Predictive Uncertainty
TL;DR
KReF is a novel training-free retrieval framework for long-term time-series forecasting that utilizes historical data to create predictive distributions. It outperforms traditional trained models in accuracy and predictive uncertainty across multiple benchmarks.
✦ Why It Matters
Consider implementing KReF for your time-series forecasting tasks to enhance predictive accuracy without extensive training.
Key Takeaways
- Explore KReF for training-free time-series forecasting solutions.
- Utilize historical data retrieval to improve predictive accuracy.
- Implement handcrafted statistics for better data embedding.
- Evaluate KReF's performance against traditional models in your projects.
How It Works
KReF operates by treating historical lookback-future pairs as a basis for constructing an empirical predictive distribution. It preprocesses the data and embeds lookbacks using either handcrafted statistics or frozen random Fourier features.
By retrieving similar historical pairs, KReF assigns similarity weights that define predictive masses, quantiles, and point forecasts. This method allows for the construction of a probability-integral-transform map, adapting interval boundaries based on validation-selected rates.