内容来源：https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/

内容总结：

谷歌研究院于2026年8月31日正式发布新一代时间序列基础模型TimesFM-3。该模型拥有3.3亿参数，基于超过1万亿时间点的真实及合成数据预训练，实现了单次前向传播即可完成高精度多变量时间序列预测，在主流基准测试中显著优于其他预测模型。

此前，TimesFM系列（包括2025年9月发布的2.5版本）仅支持单变量预测，即仅依赖单一历史序列进行推演。然而，现实场景中的预测问题大多涉及多变量交互，例如预测冰淇淋销量时，需综合考虑关联产品销量、客流量、天气、促销活动及节假日等外部因素。TimesFM-3通过原生预训练直接支持多变量预测，能够零样本联合预测多条共同演化的时间序列，捕捉变量间依赖关系，无需针对特定任务进行微调即可提升整体预测准确性。

在技术架构上，TimesFM-3延续了此前经过验证的解码器专用Transformer架构，将连续数据点分组为32步长的补丁进行高效处理，并对每条序列单独归一化以适应不同尺度。针对已知未来协变量（如已排定的促销计划），模型采用“前瞻”策略，将当前补丁与未来补丁拼接，使其在预测时能利用已知的未来信号。与旧版逐补丁生成预测的模式不同，TimesFM-3采用“连续补丁掩码”策略，在单个前向过程中一次性生成整个预测区间，大幅降低延迟、误差累积和计算成本。模型可为每个目标序列在每步预测中输出9个分位数（10%至90%），提供完整的概率性预测结果。

以冰淇淋销售预测为例，传统单变量模型只能根据历史销售勾勒周度模式，无法感知特定日期的促销计划；而TimesFM-3的多变量模式可将促销安排作为已知未来协变量输入，从历史背景中学习促销与销量提升的关系，准确预测出促销日约20%的销量增幅，从而在整月预测中获得更精确的营收预期。

评测结果显示，在Gift-Eval、FEV-Bench和Time三大公开基准中，TimesFM-3在点预测与概率预测指标上均位列所有预训练基础模型之首，超越了Chronos-2、Toto 2.0系列及前代TimesFM-2.5。即便在完全隔离序列、不使用任何协变量或交叉信息的单变量模式下，TimesFM-3也能与竞争对手持平或胜出；切换至完整多变量模式后，其平均排名进一步提升。

TimesFM-3现已开放于GitHub和Hugging Face平台，BigQuery集成将于数周内上线。该项目由Yichen Zhou、Petros Mol、Abhimanyu Das和Samet Oymak共同参与完成。

中文翻译：

2026年8月31日
Ayush Jain 和 Rajat Sen，谷歌研究部门研究科学家
我们推出TimesFM-3，这是一款最先进的时间序列基础模型，能够在单次前向传播中实现高精度的多变量时间序列预测，在各大基准测试中显著优于其他预测模型。
自2024年TimesFM首次亮相以来，我们见证了时间序列基础模型在多个领域的真实世界时间序列预测任务中的广泛采用，涵盖零售、金融、可观测性、制造、医疗保健和自然科学等行业。
在TimesFM-2.5（2025年9月发布）之前，我们的模型严格限制为单变量预测：仅利用单个时间序列的历史数据进行预测。然而，大多数真实世界的预测问题本质上都是多变量的：多个时间序列和辅助外部特征共同影响某个时间序列的未来预测。以零售连锁店的冰淇淋销量预测为例。仅凭过去的销量很难完整还原全貌。一个好的预测还应该参考相关产品（如冰淇淋蛋筒、糖浆）的销量、历史客流量，以及已知的未来事件，如天气预报、促销活动和节假日。
今天，我们推出TimesFM-3，这是我们下一代时间序列基础模型，原生为多变量预测而预训练。TimesFM-3拥有3.3亿个参数，在包含超过1万亿个时间点的真实世界和合成时间序列语料库上进行预训练。在继承前代模型的高效性和零样本泛化能力的基础上，TimesFM-3以零样本方式增加了对复杂多变量场景的稳健支持。它可以联合预测多个共同演进的时间序列，捕捉依赖关系从而提升整体准确性，而无需针对特定任务进行微调。该模型原生支持：
TimesFM-3沿用了其前代产品经过验证的仅解码器Transformer架构。与前几版一样，我们通过将连续数据点分组为32个时间步长的补丁来高效处理时间序列。然后，我们参照TimesFM-2.5的方式对每个时间序列进行归一化，以应对尺度差异巨大的时间序列。
对于目标和过去协变量序列，每个标记直接由单个补丁构建。然而，对于过去-未来协变量，TimesFM-3采用了一种巧妙的“前瞻”策略：每个标记将当前补丁与未来补丁拼接在一起，使模型能够窥见即将到来的已知信号。
补丁完成标记化后，它们经过一个输入残差块，进入主Transformer堆栈，该堆栈以二维网格形式运作：
这两种注意力机制在多层中交替进行，将时间模式与跨序列关系无缝融合。
此前的TimesFM版本一次生成一个补丁的预测，带来了延迟、误差累积和计算成本问题。TimesFM-3采用连续补丁掩码策略，在单次前向传播中生成整个预测区间。模型在观测上下文之后附加未来区间的掩码占位标记。目标和过去协变量序列在预测区间内被掩码（因为其未来值未知），而过去-未来协变量保持可见，为模型提供节假日或计划事件等已知的未来信号。通过交替的注意力层，模型同时填充所有被掩码的预测区间补丁，无需任何迭代循环。模型在预测区间的每个时间步为每个目标时间序列预测9个分位数（从第10百分位到第90百分位），提供预测不确定性的完整概率视图。
让我们重新回到冰淇淋销量示例。假设你正在制定下个月的促销排期，并希望预测销量以便提前准备。标准的单变量模型（下图中的红线）仅查看历史销量并将周模式向前投射——但它对特定日期的计划促销一无所知。TimesFM-3的多变量模式（下图中的蓝线）采取了不同的方法：通过将计划促销排期作为过去-未来协变量传入，模型从历史上下文中学习促销与销量提升之间的关系，然后将该知识应用于未来有计划促销的日子。结果是一个能够预见每个促销日约20%销量提升的预测。在下图中，促销协变量中的琥珀色色块标明了哪些日子有促销活动——蓝色预测线对每个促销日都做出了明显响应，而红色预测线则没有。整个月累计下来，这为预测营收带来了更准确的预测结果。
我们在三个综合性的公开预测基准上对TimesFM-3进行了评估：Gift-Eval、FEV-Bench和Time。在全部三个基准上，TimesFM-3在所有预训练基础模型中，无论是点预测还是概率预测指标均排名第一。下图展示了三个基准在点预测准确性和概率预测质量上的平均排名（数值越低越好）。我们与近期的基础模型进行了对比，包括支持多变量的模型（如Chronos-2和Toto 2.0系列），以及我们之前的模型TimesFM-2.5。
每个图中包含TimesFM-3的两个条目。“单变量模式”数据点表示在没有任何协变量或跨序列信息的情况下评估模型的性能，每个目标序列独立处理，与传统的单变量模型一样。即使在单变量模式下，TimesFM-3也已经达到或超越了其他竞争模型。当我们切换到完整的多变量模式时，TimesFM-3再次实现飞跃，在点预测和概率预测两方面均获得全面最佳平均排名。
我们推出TimesFM-3，这是我们TimesFM系列零样本时间序列基础模型的最新一代，在多个公开基准上取得了最先进的多变量和单变量预测性能。TimesFM-3现已可在GitHub和Hugging Face上获取，其BigQuery集成将在未来几周内上线。与此同时，你可以立即在单变量任务上试用TimesFM-2.5，熟悉BigQuery中的AI.FORECAST命令——无需机器学习专业知识即可上手。
本项目是与Yichen Zhou、Petros Mol、Abhimanyu Das和Samet Oymak的合作成果。

英文来源：

August 31, 2026
Ayush Jain and Rajat Sen, Research Scientists, Google Research
We introduce TimesFM-3, a state-of-the-art time series foundation model that enables highly accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks.
Since the debut of TimesFM in 2024, we’ve seen the adoption of time-series foundation models for real-world time-series forecasting tasks across multiple domains, such as retail, finance, observability, manufacturing, healthcare and natural sciences.
Up until TimesFM-2.5 (released in September 2025), our models were strictly limited to univariate forecasting: forecasting using only the history of a single time series. Yet, most real-world forecasting problems are inherently multivariate: where multiple time series and auxiliary external features jointly impact the future forecast of a time series. Consider forecasting ice cream sales for a retail chain. Past sales alone rarely tell the full story. A good forecast should also draw on sales of related products (e.g., ice cream cones, syrups), historical foot traffic, and known future events like weather forecasts, promotions, and holidays.
Today we introduce TimesFM-3, the next generation of our time-series foundation model that is natively pre-trained for multivariate forecasting. TimesFM-3 has 330 million parameters and is pre-trained on a real-world and synthetic time-series corpus comprising more than 1 trillion time points. Building on the efficiency and zero-shot generalization of its predecessors, TimesFM-3 adds robust support for complex multivariate scenarios in a zero-shot manner. It can jointly predict multiple coevolving time series, capturing dependencies that improve overall accuracy without requiring task-specific fine-tuning. The model natively supports:
TimesFM-3 builds on the proven decoder-only transformer architecture of its predecessors. As in previous versions, we process time series efficiently by grouping contiguous data points into patches of 32 time steps. We then apply normalization per time-series similar to that of TimesFM-2.5 in order to account for time series with vastly different scales.
For target and past-covariate series, a token is constructed directly from a single patch. However, for past-future covariates, TimesFM-3 employs a clever "lookahead" strategy: each token concatenates the current patch with future patches, allowing the model to peek at upcoming known signals.
Once the patches are tokenized, they pass through an input residual block and enter the main transformer stack, which operates as a 2D grid:
These two attention mechanisms alternate for several layers, seamlessly blending temporal patterns with cross-series relationships.
Previous versions of TimesFM generated forecasts one patch at a time, introducing latency, compounding error accumulation, and computational cost. TimesFM-3 uses the strategy of Contiguous Patch Masking to generate the entire forecasting horizon in a single forward pass. The model appends masked placeholder tokens for the future horizon alongside the observed context. Target and past-covariate series are masked in the horizon (since their future values are unknown), while past–future covariates remain visible, providing the model with known future signals like holidays or scheduled events. Through the alternating attention layers, the model fills in all masked horizon patches simultaneously, with no iterative loop required. The model predicts 9 quantiles (from the 10th to the 90th percentile) for each target time series at every horizon step, providing a full probabilistic view of the forecast uncertainty.
Let’s revisit the ice cream sales example. Imagine you are working on next month’s promotion schedule and want to forecast the sales to anticipate. A standard univariate model (the red line, below) looks at the historical sales and projects a weekly pattern forward — but it has no idea about planned promotions on specific days. TimesFM-3's multivariate mode (the blue line, below) takes a different approach: by passing in the planned promotion schedule as a past-future covariate, the model learns the relationship between promotions and sales lift from the historical context, then applies that knowledge to future days with planned promotions. The result is a forecast that anticipates a ~20% sales bump on each promotion day. In the chart below, the amber blocks in the promotion covariates highlight which days have promotions — and the blue forecast visibly responds to each one, while the red forecast does not. Over the full month, this adds up to a more accurate forecast for projected revenue.
We evaluated TimesFM-3 on three comprehensive public forecasting benchmarks: Gift-Eval, FEV-Bench, and Time. On all three benchmarks, TimesFM-3 is the top-ranked model in terms of both point and probabilistic forecasting metrics among all pre-trained foundation models. The plots below show average rank across tasks for both point forecast accuracy and probabilistic forecast quality (lower is better) for the three benchmarks. We compare against recent foundation models including multivariate-capable models, such as Chronos-2 and the Toto 2.0 family, as well as our previous model TimesFM-2.5.
Each plot includes two entries for TimesFM-3. The "univariate mode" point shows performance when the model is evaluated without any covariate or cross-series information, treating each target series independently, just like a traditional univariate model. Even in this univariate mode, TimesFM-3 already matches or outperforms other competing models. When we switch to the full multivariate mode, TimesFM-3 takes another leap, achieving the best average rank in both point and probabilistic forecasting across the board.
We introduce TimesFM-3, the latest generation of our TimesFM family of zero-shot time series foundation models, that obtains state-of-the-art multivariate and univariate forecasting performance on multiple public benchmarks. TimesFM-3 is now available on GitHub and Hugging Face, with its BigQuery integration landing in the coming weeks. In the meantime, you can try TimesFM-2.5 immediately on your univariate tasks to familiarize yourself with the AI.FORECAST command in BigQuery - no ML expertise required.
This project is joint work with Yichen Zhou, Petros Mol, Abhimanyu Das and Samet Oymak.