Title: An Amortized Predictability-Aware Training Framework for Time Series Forecasting and Classification

Abstract:

Time series datasets across diverse fields frequently suffer from noise, where training samples may include patterns with low predictability that stray from the standard data distribution. This deviation can result in unstable training processes or cause models to converge toward suboptimal local minima. Consequently, addressing the negative impact of these low-predictability samples is essential for tasks involving time series forecasting (TSF) and time series classification (TSC). Although numerous deep learning architectures have demonstrated strong performance, there is a notable lack of approaches that actively identify and penalize such samples to enhance performance from a training standpoint. To address this limitation, we introduce the Amortized Predictability-aware Training Framework (APTF), a versatile solution applicable to both TSF and TSC. APTF incorporates two innovative components designed to direct the model’s attention toward high-predictability samples while still allowing for effective learning from lower-predictability instances: (i) a Hierarchical Predictability-aware Loss (HPL) that dynamically detects low-predictability samples and incrementally increases their loss penalties as training progresses, and (ii) an amortization model that corrects for prediction errors stemming from model bias, thereby amplifying the efficacy of the HPL. The source code for this framework is publicly accessible at https://github.com/Meteor-Stars/APTF.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC