ai-ml-timeseries

# Time Series Forecasting - Production Patterns **Scope note:** This skill covers forecasting system construction and evaluation. It is not part of the LLM-build or LLM-training stack — route LLM lifecycle, prompting, or provider questions to [ai-llm](../ai-llm/SKILL.md). **July 2026 posture:** define a cutoff timestamp before modelling, start with strong baselines, prefer horizon-aware validation over IID thinking, treat known-future covariates explicitly, and verify fast-moving tooling against current official docs before recommending it. This skill is the implementation guide for **forecasting systems**: - timestamp integrity, frequency checks, and point-in-time feature design - local, global/panel, and hierarchical forecasting workflows - leakage-safe backtesting, horizon-wise evaluation, and business-loss alignment - probabilistic forecasting, calibration, and interval quality - time-series foundation models (TSFMs) and zero-shot benchmark patterns - forecasting-specific handoff, fallback, and lineage requirements Use this skill for **forecasting depth**. Use sibling skills for general data science, generic LLM strategy, or full production operations. ## When To Use This Skill Activate this skill when the user asks for: - building or reviewing a forecast model - choosing between local, global/panel, hierarchical, or foundation-model approaches - creating lag, rolling, calendar, or known-future covariate features - designing a rolling-origin backtest or fixing tem...