Self-Adaptive Scale Handling for Forecasting Time Series with Scale Heterogeneity

A new machine-learning module designed to handle time series data that vary dramatically in scale has been proposed by researchers, with tests on financial sales data from Ant Group platforms showing consistent performance gains when integrated into popular forecasting models [1][2]. The work, submitted for publication on 18 June 2026, addresses a persistent challenge in time series forecasting (TSF): most research assumes scale-homogeneous data, where different series share similar numerical ranges [1][2]. In industrial settings such as financial product sales, however, individual time series can differ by orders of magnitude, a condition the authors term scale heterogeneity [1][2]. Because these series often exhibit similar temporal patterns, forecasting them jointly is desirable, but existing scaling techniques introduce trade-offs. Global normalization can compress low-scale signals, while window-based scaling can destroy semantic discriminability and amplify inverse-scaling errors [1][2]. The proposed solution, called the self-Adaptive Scale-handling (AS) module, learns adaptive scale factors tailored to each input [1][2]. It comprises two components: Scale Calibrating (SC), which uses neural networks to calibrate prior mean scaling factors, and Scaling Selection (SS), which decides whether to apply the calibration or retain the original factor to avoid over-calibration [1][2]. The design aims to preserve semantic discriminability while reducing inverse-scaling errors [1][2]. Experiments were conducted on real-world fund sales datasets sourced from Ant Fortune and Alipay [1][2]. Ant Fortune and Alipay are both operated by Ant Group, the financial technology affiliate of Alibaba Group [5][7]. Ant Group was originally established in 2004 as Alipay, the payment arm of Alibaba, and has since grown into one of the world’s largest fintech companies, with Alipay reported to have over 1.3 billion users as of 2020 [5]. The group’s planned initial public offering, which would have valued the company at US$313 billion, was suspended by Chinese regulators in 2020 [5][6]. The researchers report that the AS module integrates seamlessly into popular TSF models and consistently improves their performance on these datasets [1][2]. The code and dataset have been made publicly available [1][2].