Ant Group’s Time Series AI Practices: AntFlux Engine and Real‑World Applications

Ant Group presents its practice in time‑series AI, focusing on the AntFlux intelligent engine and its deployment in various business scenarios.

Value of Time Series

Time‑series data is the second most important data type after structured data, with higher value than images, video, text, or audio. It appears in many domains such as finance, weather forecasting, traffic prediction, supply‑chain management, and autonomous driving.

Ant Group Business Applications

User‑service layer: fine‑grained wealth and insurance advisory using micro‑level user behavior and asset allocation.

Business‑operation layer: liquidity forecasting and resource management for consumer finance, wealth, and online‑business units.

Cloud‑computing infrastructure: AI‑driven elastic capacity to improve resource utilization.

Types of Time Series

Asynchronous series (e.g., user behavior logs) with irregular intervals, modeled by neural point processes.

Synchronous series with regular intervals, typically modeled by global models covering all sequences.

AntFlux Platform

AntFlux is an industrial‑grade time‑series algorithm platform that provides advanced algorithms, large‑scale computation, and end‑to‑end workflow support.

Insight: time‑series analysis, anomaly detection, feature generation.

Forecaster: synchronous modeling with a rich algorithm library and automatic model selection.

ATS: asynchronous time‑series modeling.

AIStudio Components: drag‑and‑drop components for model building and deployment.

WorkFlow: orchestrated time‑series pipelines.

Research, Community, and other supporting modules.

Time‑Series AI Technologies

Statistical models : Holt‑Winter exponential smoothing and ARIMA, which are single‑series, require strong priors, and struggle with cold‑start and large‑scale maintenance.

Deep models (2018‑present) : TCN, N‑BEATS, DeepAR, Autoformer, Informer, ETSformer, NSformer, PatchTST, Corrformer, offering unified modeling, better feature extraction, but limited cross‑scenario transfer.

Large‑scale time‑series models (2023‑present) :

Time‑LLM: reprogrammed large language models for univariate forecasting using prompt‑based statistical descriptors and patch reprogramming, achieving state‑of‑the‑art performance on long‑ and short‑term tasks.

iTransformer: variable‑centered tokenization for multivariate series, enabling efficient attention across series and strong performance on benchmark datasets.

Memory‑Augmented State‑Space Model (EMSSM): external memory to capture long‑range dependencies, improving multi‑step prediction.

SLOTH: hierarchical time‑series model that fuses top‑down and bottom‑up information and integrates prediction with downstream decision optimization.

Key Model Highlights

Time‑LLM leverages model reprogramming to align time‑series statistics with LLM inputs, requiring no fine‑tuning and delivering lightweight yet accurate forecasts.

iTransformer treats each variable as a whole, improving representation for multivariate data and achieving competitive results.

Memory‑augmented models store past hidden states in an external key‑value memory, enabling long‑horizon predictions such as holiday sales.

SLOTH combines hierarchical attention and convolution to capture multi‑level patterns, achieving top performance in hierarchical forecasting tasks.

Business Applications of AntFlux

Time‑series insight: anomaly detection, situational awareness, hotspot mining.

Time‑series forecasting: elastic capacity for green computing, fine‑grained fund management, risk pricing, cost control, supply‑chain optimization.

Time‑series decision: predictive planning and control to automate business execution and improve cost efficiency.

Q&A Highlights

Time‑LLM currently supports univariate series; future work will extend to multivariate with prompt designs that capture inter‑series relationships.

External textual information can still be aligned via prompts to assist forecasting.

Open challenges in time‑series forecasting include better utilization of covariates, multimodal fusion, model scaling benefits, and interpretability of deep models.

iTransformer handles multivariate data by variable‑centered modeling; multimodal handling is addressed by Time‑LLM.

In summary, Ant Group’s AntFlux platform integrates a wide range of time‑series AI technologies—from classic statistical methods to cutting‑edge large language model reprogramming—providing comprehensive solutions for insight, forecasting, and decision making across diverse industry scenarios.