Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting Google AI发布TimesFM-3：一款3.3亿参数的零样本多变量时间序列预测基础模型

TL;DR

• Google Research released TimesFM-3, a 330M parameter time series foundation model pretrained natively for multivariate forecasting on over 1 trillion time points
• Unlike previous univariate versions, TimesFM-3 forecasts multiple related series simultaneously in a single forward pass with zero-shot capability
• The model uses alternating causal temporal and full variate attention mechanisms with Contiguous Patch Masking to produce the entire forecast horizon at once
• TimesFM-3 ranks #1 among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard for both point and probabilistic metrics
• The model weights are licensed non-commercially; only the repository code is Apache-2.0, while TimesFM 2.5 remains the production-usable option

Why It Matters

TimesFM-3 represents a significant architectural shift in time series forecasting by moving from univariate to native multivariate modeling, which better reflects real-world forecasting scenarios where multiple interdependent variables influence outcomes. For AI practitioners, this means the ability to capture cross-series correlations without task-specific fine-tuning, potentially reducing engineering overhead and improving forecast accuracy across domains like retail, energy, and supply chain.

Technical Details

• Architecture: Decoder-only transformer backbone with patch-based tokenization (32-step patches), normalized per series to handle scale variations
• Attention Mechanisms: Two alternating attention types—causal temporal attention (horizontal, strictly causal within series) and full variate attention (vertical, captures cross-series correlations at each time step)
• Input Types: Supports three zero-shot input modalities—multiple targets (joint forecasting with point and quantile outputs), past covariates (historically known variables), and past-future covariates (known future values like promotion calendars)
• Contiguous Patch Masking: Training strategy where masked placeholder tokens are appended for the entire horizon; targets and past covariates are masked while past-future covariates remain visible, enabling simultaneous horizon prediction
• Probabilistic Forecasting: Each target receives 9 quantiles (10th-90th percentile) at every horizon step
• Scale: 330 million parameters pretrained on 1 trillion+ time points of real and synthetic series

Industry Insight

The non-commercial licensing of TimesFM-3 weights creates a strategic dilemma: while it sets a new benchmark for multivariate forecasting accuracy, production deployments must rely on the older TimesFM 2.5 (Apache-2.0) or negotiate commercial licensing, potentially slowing enterprise adoption. Organizations should prioritize benchmarking TimesFM-3 on their specific use cases to understand the multivariate advantage before committing to production infrastructure, and consider contributing to open-source multivariate forecasting research to help shift licensing models toward more accessible commercial terms.

TL;DR

• Google发布TimesFM-3，3.3亿参数的原生多变量时间序列基础模型，单次前向传播即可联合预测多个相关序列
• 模型在超1万亿时间点上预训练，支持多目标、历史协变量和过去-未来协变量三种零样本输入，无需任务微调
• 采用交替因果时序注意力与全变量注意力的双注意力机制，结合连续补丁掩码实现全视界并行预测
• 在GIFT-Eval、fev-bench和TIME三大基准上均获预训练基础模型最高平均排名，覆盖点估计与概率预测指标
• 权重采用非商业许可，仅允许基准测试，生产部署仍需使用Apache-2.0授权的TimesFM 2.5

为什么值得看

TimesFM-3突破了时间序列预测长期依赖单变量自回归的局限，首次以原生多变量架构实现零样本联合预测，大幅降低实际业务中多序列协同建模的门槛。其在多个权威基准上的领先表现，标志着时间序列基础模型从"单点预测"向"系统级预测"的关键跃迁。

技术解析

• 多变量原生预训练：TimesFM-3首次放弃单变量设计，直接在1万亿+真实与合成时间点上预训练，支持同时预测多个目标序列，并输出每个目标9个分位数（10%-90%），实现概率预测。
• 双注意力架构：模型将连续32步划分为补丁，通过交替的因果时序注意力（水平方向，严格防止信息泄露）和全变量注意力（垂直方向，捕捉跨序列相关性）建模复杂依赖关系。
• 连续补丁掩码（Contiguous Patch Masking）：借鉴TiRex训练策略，在预测视界内填充掩码占位符，目标与历史协变量被掩码，而过去-未来协变量（如促销日历）保持可见，使模型在一次前向传播中并行生成全视界输出，避免逐补丁解码的延迟与误差累积。
• 三类零样本输入：支持多目标联合预测、仅历史已知的协变量（如过往客流量）、以及未来值已知的协变量（如排期事件），无需任务特定微调即可适配多种实际场景。
• 许可限制：代码为Apache-2.0开源，但3.0权重采用非商业许可，禁止生产部署，商业用户需继续使用TimesFM 2.5。

行业启示

• 多变量时间序列建模进入基础模型时代：TimesFM-3证明原生多变量架构在零样本场景下的有效性，未来时间序列预测将更多借鉴NLP/视觉基础模型范式，从"为每个任务训练模型"转向"预训练+零样本适配"。
• 协变量整合成为核心竞争力：对过去-未来协变量的原生支持，使模型能直接利用业务日历、促销计划等先验信息，这对零售、能源、供应链等强依赖外部因子的行业具有直接应用价值。
• 许可策略影响落地节奏：非商业许可虽保护模型资产，但可能延缓工业界在生产环境的验证与迭代，建议企业评估TimesFM 2.5的持续优化路径，或关注后续许可政策的调整。