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

The article introduces Ant Group's time‑series AI practice, highlighting the significance of time‑series data as the second most valuable data type after structured data and its pervasive presence in domains such as finance, weather, traffic, and autonomous driving.

It outlines five core sections: the value of time‑series, AI techniques for time‑series, the algorithm platform, business applications, and a Q&A session.

Two main categories of time‑series are described: asynchronous series with irregular intervals, modeled by neural point processes, and synchronous series with regular intervals, typically handled by global models.

Ant Group developed an industrial‑grade time‑series algorithm platform called AntFlux , which provides advanced algorithms, large‑scale computation, and end‑to‑end services for research, development, and production.

The evolution of time‑series AI techniques is traced from classic statistical models (Holt‑Winter, ARIMA) to deep models (TCN, N‑BEATS, DeepAR, Autoformer, Informer, ETSformer, NSformer, PatchTST, Corrformer) and finally to large models (Time‑LLM, iTransformer, LLM‑Time). Statistical models suffer from limited reuse, strong priors, and cold‑start issues, while deep models improve feature learning but still lack cross‑scenario transfer.

Ant Group's proprietary models address specific challenges: APTN, DeepAR+, BiDA, EMSSM for sparse data; multi‑modal models (Time‑LLM) that reprogram large language models to handle time‑series via prompt engineering and patch reprogramming; hierarchical models (SLOTH) for naturally layered series; asynchronous models (NHPI, HYPRO, PromptTPP, LAMP) for irregular intervals; and large‑scale models (Time‑LLM, iTransformer) that achieve strong few‑shot and zero‑shot performance.

Time‑LLM leverages model reprogramming to convert a frozen LLM into a time‑series predictor by transforming inputs into textual statistics, feeding them as prompts, and mapping outputs back to numeric forecasts, achieving lightweight yet high‑accuracy results.

iTransformer treats each variable as a token, applying multi‑head attention across entire series, enabling efficient parallel prediction and superior performance on multivariate benchmarks.

Memory‑augmented state‑space models introduce external memory to capture long‑range dependencies, improving predictions for scenarios like seasonal sales.

The AntFlux platform comprises modules such as Insight (analysis, anomaly detection), Forecaster (synchronous modeling), ATS (asynchronous modeling), AI‑Studio Components (drag‑and‑drop modeling and AutoTS), WorkFlow (pipeline orchestration), and a suite of product families that support one‑stop services from data ingestion to model deployment and monitoring.

Business applications are categorized into time‑series insight (historical analysis), forecasting (future value prediction for resource planning, risk pricing, supply‑chain), and decision (optimizing actions based on predictions). A green‑computing use case demonstrates how AI‑driven elastic capacity reduces carbon footprint by aligning CPU usage with demand.

The Q&A addresses future directions, such as extending Time‑LLM to multivariate series, handling auxiliary textual information, remaining challenges in time‑series forecasting, and the capabilities of iTransformer and large models.