When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting

TL;DR

This paper evaluates the benefits of online adaptation for edge time-series forecasting under distribution drift. The authors identify sources of comparison bias and propose a leakage-free evaluation protocol. They find that the warmup budget and learning rate selection significantly impact the measured adaptation benefit.

Why It Matters

The paper's findings have practical significance for edge computing applications, such as smart meter and building sensor data analysis, where online adaptation can improve forecasting accuracy. The proposed evaluation protocol can help developers make informed decisions about online adaptation strategies. The results can also inform the design of more efficient and effective edge computing systems.

Key Findings

• warmup budget affects static baseline performance
• learning rate selection impacts optimizer quality
• Adam outperforms SGD+m in most evaluated settings