Summary
Fahd Mirza walks through a live local installation and demonstration of TimesFM-3, Google Research’s latest time-series foundation model — a 330-million-parameter decoder-only transformer pre-trained on over one trillion real and synthetic time points. The model performs zero-shot forecasting with no fine-tuning required, and the video covers the full setup process using UV on an Ubuntu system with a consumer-grade GPU, pulling the model from Hugging Face.
The demo centers on a synthetic dataset representing daily active users of a dating app around Sydney Pride season. Mirza runs two forecasts: a univariate pass using only historical user counts, and a multivariate pass with a binary covariate flag marking Pride event days — a signal the model can see into the future. The result is visually clear: the blind forecast misses the surge entirely, while the covariate-informed forecast tracks the actual spike closely. Mean absolute error drops from roughly 3,000 users per day to around 1,000 — a threefold improvement from a single additional data column.
The video explains the code in plain terms, covering CSV loading, train/test split, model initialization, and dual-forecast measurement, making it accessible to practitioners without a deep time-series background. TimesFM-3’s ability to incorporate known future events like promotions and holidays as covariates is positioned as its key practical advantage over conventional univariate forecasting approaches.
📺 Source: Fahd Mirza · Published September 05, 2026
🏷️ Format: Hands On Build
