Time Series Forecasting - Production Patterns

Define the prediction cutoff before modelling: only information available then can enter features or baselines. Compare models on the decision horizon against the forecast currently driving decisions.

Scope Boundaries

• General EDA, tabular modelling, experiment design, or reusable DS workflow -> ai-ml-data-science
• Deployment architecture, monitoring stack, release gates, incident playbooks -> ai-mlops
• Generic LLM lifecycle, prompting, or provider selection -> ai-llm
• RAG and search systems -> ai-rag

Workflow

1. Define the forecast target, horizon, cutoff timestamp, granularity, and business-loss shape first.
2. Route generic DS workflow or full production-ops work to the adjacent skill when forecasting depth is not the main need.
3. Choose the baseline, feature pattern, and model family from the decision tree.
4. Run leakage-safe backtests, compare horizon-aware metrics, and add interval or calibration checks when decisions need uncertainty.
5. Before selecting a package or checkpoint, open its official docs from data/sources.json and check the installed API, supported covariates, context/horizon limits and weight licence. For TSFM comparisons, load model roles and lookup steps; record the benchmark snapshot next to any quoted rank. For MAPIE code, load probabilistic patterns and compare the example to the installed release's docs.