<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>TimesFM on C.CUI's Log</title><link>https://cuicaihao.github.io/tags/timesfm/</link><description>Recent content in TimesFM on C.CUI's Log</description><generator>Hugo</generator><language>en-AU</language><lastBuildDate>Sun, 08 Nov 2026 10:30:00 +1100</lastBuildDate><atom:link href="https://cuicaihao.github.io/tags/timesfm/index.xml" rel="self" type="application/rss+xml"/><item><title>Time Series Foundation Models - Google TimesFM (Part 2): Synthetic Retail Experiments and Forecast Analysis</title><link>https://cuicaihao.github.io/posts/2026-11-08-timesfm-retail-controlled-experiments-and-multi-model-benchmark/</link><pubDate>Sun, 08 Nov 2026 10:30:00 +1100</pubDate><guid>https://cuicaihao.github.io/posts/2026-11-08-timesfm-retail-controlled-experiments-and-multi-model-benchmark/</guid><description>Using four five-year synthetic retail series across 13 non-overlapping 28-day rolling windows, this article compares TimesFM 3.0 univariate zero-shot forecasting, known-covariate integration, and statistical and tree-based baselines under both expanding-history and fixed 104-week training windows. Incorporating future-known business schedules consistently lowers forecast error on regular and pulse-driven scenarios (Grocery, Wine, Garden), whereas in extreme zero-inflated demand (Luxury), single-stage models still fail to detect sparse purchase events.</description></item><item><title>Time Series Foundation Models - Google TimesFM (Part 1): Model, Algorithms, and Engineering Practice</title><link>https://cuicaihao.github.io/posts/2026-11-07-dissecting-google-timesfm-architecture-and-algorithms/</link><pubDate>Sat, 07 Nov 2026 10:30:00 +1100</pubDate><guid>https://cuicaihao.github.io/posts/2026-11-07-dissecting-google-timesfm-architecture-and-algorithms/</guid><description>Benchmark leaderboards list dozens of time-series foundation models, but production selection requires more than an average rank. This article dissects Google TimesFM from version 1.0 to 3.0: input patching and local instance normalization remain consistent, while decoding shifts from a 128-step output head to 3.0&amp;rsquo;s parallel CPM filling, covariate handling evolves from external XReg linear regression to native multivariate attention, and open checkpoints split into commercial and non-commercial tracks.</description></item><item><title>Should We Build a Foundation Model in a Forest of Trees? A Pragmatic Guide to Time Series Foundation Models</title><link>https://cuicaihao.github.io/posts/2026-11-01-time-series-foundation-models-vs-gbdt-practical-guide/</link><pubDate>Sun, 01 Nov 2026 07:00:00 +1100</pubDate><guid>https://cuicaihao.github.io/posts/2026-11-01-time-series-foundation-models-vs-gbdt-practical-guide/</guid><description>&amp;lsquo;Foundation models for time series&amp;rsquo; actually encompasses two fundamentally different approaches: repurposing pretrained language models via adaptation layers (where ablations reveal removing the LLM leaves accuracy unchanged while speeding up training by orders of magnitude), and native time series foundation models pretrained from scratch on physical temporal data (which demonstrate real promise in cold-start and covariate-driven forecasting). Here is an engineering reality check and practical decision guide.</description></item></channel></rss>