Alibaba Qwen Team Unveils CoCo: Multimodal Time-Series AI Forecasting

Alibaba researchers behind Qwen have introduced Context as Covariates (CoCo), a hybrid forecasting framework combining Large Language Models with Time Series Foundation Models. By converting unstructured news and textual reports into numerical covariates to guide quantitative forecasting models, CoCo achieves up to a 12% error reduction over standalone LLMs and time-series architectures across financial, economic, and operational benchmarks.

What Is Alibaba's Context as Covariates (CoCo) Framework?

Real-world time-series datasets in finance, supply chain, weather, and security are heavily influenced by qualitative events—such as breaking news, earnings calls, economic reports, and social announcements. While traditional Time Series Foundation Models (TSFMs) like Chronos2 excel at quantitative history, they lack the ability to read or reason about qualitative text. Conversely, Large Language Models (LLMs) like GPT-4o or Qwen models possess deep textual reasoning but struggle with precise numerical calculations.

To bridge this gap, Alibaba researchers introduced Context as Covariates (CoCo). Instead of forcing an LLM to generate raw numerical predictions or feeding raw text directly into numerical models, CoCo uses an LLM to distill unstructured textual context into numerically grounded forecast and confidence covariates. These covariates are then ingested by a dedicated TSFM backbone to produce highly accurate predictions.

How Does CoCo Bridge Large Language Models and Time Series Foundation Models?

Connecting LLMs with TSFMs required solving two core challenges: the lack of paired multimodal training datasets and the alignment between textual reasoning and numerical output.

Alibaba solved these challenges through a novel two-stage training approach:

• Synthetic Multimodal Data Generation: The team developed a scalable pipeline that uses LLMs to automatically annotate real-world time-series data with realistic textual context.
• Reinforcement Learning Alignment: The LLM (specifically Qwen3-4B-Instruct-2507) is fine-tuned using Group Relative Policy Optimization (GRPO). The reward function is directly tied to how much the generated covariates improve the downstream TSFM's forecasting accuracy.
• TSFM Fine-Tuning: The quantitative backbone (Chronos2) is subsequently fine-tuned on the covariates emitted by the aligned LLM.

When evaluated against state-of-the-art baselines across financial markets, economic indicators, weather forecasts, and traffic telemetry, CoCo achieved up to a 12% Mean Absolute Error (MAE) improvement over top-tier LLM forecasters and TSFMs.

Why Is Multimodal Time-Series Forecasting a Breakthrough for Enterprise Decision-Making?

For enterprise technical leaders and business strategists, traditional time-series models have always suffered from a blind spot: unexpected macro events. A numerical model trained on past stock prices or shipping volumes cannot anticipate the impact of a newly published regulatory policy or severe weather advisory until the data reflects it post-facto.

By establishing a formal framework for incorporating real-time textual signals into numerical baselines, CoCo allows predictive systems to react dynamically to news before historical curves update. This enables predictive risk management, automated inventory rebalancing, and enhanced algorithmic trading strategies.