google-research/timesfm
TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
Python · ★ 33,963 · Fork 3,267 · 在 GitHub 查看
在手机上浏览源码、订阅每日 AI 趋势日报与项目架构深度解读:获取 CopoHub 客户端(iOS & HarmonyOS)→
GitHub Trending 上榜记录
google-research/timesfm 已在 GitHub Trending 每日榜单上榜 15 天,最高排名第 1,首次上榜 2026-02-20,最近一次上榜 2026-09-05。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2026-09-05 | #13 | 31,178 | +342 |
| 2026-09-04 | #9 | 31,015 | +1,618 |
| 2026-09-03 | #2 | 30,672 | +1,626 |
| 2026-09-02 | #2 | 29,645 | +326 |
| 2026-06-21 | #6 | 24,699 | +433 |
| 2026-06-20 | #2 | 24,490 | +1,510 |
| 2026-06-19 | #1 | 24,043 | +1,516 |
| 2026-06-18 | #1 | 23,032 | +858 |
| 2026-06-17 | #6 | 21,820 | +712 |
| 2026-04-04 | #3 | 14,366 | +916 |
| 2026-04-03 | #3 | 14,060 | +912 |
| 2026-04-02 | #3 | 12,984 | +1,195 |
| 2026-04-01 | #3 | 11,946 | +495 |
| 2026-02-21 | #10 | 8,945 | +404 |
| 2026-02-20 | #10 | 8,820 | +438 |
同期上榜的项目
- mattpocock/skills – Skills for Real Engineers. Straight from my .agents directory.
- affaan-m/ECC – The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
- DietrichGebert/ponytail – Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
- NousResearch/hermes-agent – The agent that grows with you
- fmtlib/fmt – A modern formatting library
- anthropics/skills – Public repository for Agent Skills
TimesFM 项目深度剖析与商业评估报告
TimesFM 项目深度剖析与商业评估报告
1. 核心定位与商业落地价值
TimesFM 解决的根本痛点是时间序列预测领域的“任务碎片化”困境。传统预测方案(如 Prophet、ARIMA、LSTM 变体)要求每个场景单独训练模型,样本量不足时精度骤降,且难以应对分布漂移。TimesFM 以预训练-微调范式将时序预测从“手工作坊”推向“工业化底座”,其核心价值主张可拆解为三层:
零样本预测能力:200M 参数模型在零样本场景下即可达到或超越专精模型的精度(论文在 Monash 基准上对比了 26 个基线),显著降低了企业试错成本。 长上下文与多尺度适配:16k 的 context 长度使其能捕捉年度季节性(96 个季度数据点)、宏观经济周期等长程依赖,这是传统统计模型和短窗口 Transformer 难以企及的。 Google 生态的“特洛伊木马”效应:已原生嵌入 BigQuery ML 和 Google Sheets,这意味着企业无需迁移数据即可通过 SQL 调用 SOTA 模型。对于已深度使用 GCP 的客户,这构成了极强的粘性——预测能力成为数据仓库的“即插即用”功能,而非独立的 MLOps 项目。
竞品差异化护城河:相较于 Amazon Chronos(基于 T5 编码器)和 Moirai(统一多粒度训练),TimesFM 的护城河在于decoder-only 架构 + 非对称的 patch 化注意力。这种设计使其在保持线性计算复杂度的同时,能对任意长度历史进行条件化预测(无需像 Chronos 那样固定输入窗口)。此外,Apache 2.0 许可和双后端(PyTorch/Flax)支持显著降低了企业采用的合规风险。
商业落地场景:零售需求预测(安全库存优化)、云资源容量规划(结合 1k horizon 的连续分位数预测进行成本风险对冲)、金融波动率预测(10%-90% 分位数覆盖 VaR 计算)、制造设备剩余寿命预测。值得注意的是,其 inferispositive 标志位和归一化策略对非负时序(如销量、库存)做了针对性优化,这暗示了 Google 内部在供应链场景的实战沉淀。
2. 核心架构设计与技术深度
TimesFM 的架构精髓在于对 Transformer decoder 的时序领域适配,其模块拆解体现了清晰的“领域嵌入”思想:
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/google-research/timesfm)