Skip to content

Using AI assistants with skforecast

Skforecast provides machine-readable context files so that AI coding assistants (ChatGPT, Claude, Gemini, GitHub Copilot, Cursor, and others) can generate accurate, up-to-date code for time series forecasting.

Quick start: provide the context to an LLM

If your AI assistant can access web pages, paste the following URL into the chat:

https://skforecast.org/latest/llms-full.txt

If your assistant cannot browse URLs, open the file and paste or upload its contents instead.

This single file contains the core API reference, forecaster APIs, workflow examples, and best practices. It gives AI assistants enough context to help you with skforecast while reducing hallucinations about deprecated methods or wrong parameter names.

Example prompt:

Using the context from https://skforecast.org/latest/llms-full.txt, create a ForecasterRecursiveMultiSeries with LightGBM, fit it on a DataFrame with 3 series, and run backtesting with conformal prediction intervals.

IDE integration

If you clone or open the skforecast repository in an AI-enabled IDE that supports repository context files, the assistant can load the relevant context from the repo:

IDE / Tool File loaded How
VS Code + GitHub Copilot .github/copilot-instructions.md Used as repository instructions by Copilot
Claude Code AGENTS.md Read from the project root
OpenAI Codex / Aider AGENTS.md Read when the tool supports the AGENTS.md convention
Cursor AGENTS.md Can be used as project context when supported or configured

These files contain the same core content: project structure, all forecasters, code style, and testing conventions.

What's included in the context

The AI context covers:

  • Choice of forecaster — when to use ForecasterRecursive vs ForecasterRecursiveMultiSeries vs ForecasterRnn vs ForecasterFoundation, etc.
  • All 9 forecaster types — constructors, fit(), and predict() methods including parameters and defaults.
  • Model selection — backtesting_forecaster, bayesian_search_forecaster and other hyperparameter optimization methods, TimeSeriesFold, OneStepAheadFold, and their multi-series variants.
  • Statistical models — Arima, Sarimax, Ets, Arar wrapped by ForecasterStats.
  • Deep learning — ForecasterRnn with create_and_compile_model, LSTM/GRU architectures.
  • Foundation models (zero-shot) — FoundationModel + ForecasterFoundation with Chronos-2, TimesFM 2.5, Moirai-2, TabICL, TabPFN-TS, TFC-T0, Nori, and TS-ICL backends.
  • Feature engineering — CalendarFeatures, RollingFeatures, custom features, and exogenous variables (all built into skforecast, no extra dependency).
  • Feature selection — RFECV, SelectFromModel for lags, window features, and exogenous variables.
  • Drift detection — RangeDriftDetector and PopulationDriftDetector for production monitoring.
  • 17 specialized workflow skills — step-by-step guides for common tasks, loaded on-demand by advanced AI agents.

Workflow skills

Skforecast includes 17 modular skills — self-contained guides that AI agents can load on demand when a user asks about a specific topic. Each skill covers a complete workflow with decision trees, code examples, and common pitfalls.

Skill What it covers
choosing-a-forecaster Decision guide: "I have X situation → use Y forecaster"
autocorrelation-and-lag-selection ACF/PACF analysis with skforecast.stats to choose candidate lags
feature-engineering CalendarFeatures, RollingFeatures, custom features, exogenous variables
forecasting-single-series End-to-end forecasting with ForecasterRecursive: data prep, fit, predict, backtest, intervals
forecasting-multiple-series Global model with ForecasterRecursiveMultiSeries: encoding, dict input, multi-level predictions
foundation-forecasting Zero-shot forecasting with ForecasterFoundation: Chronos-2, TimesFM 2.5, Moirai-2, TabICL, TabPFN-TS, TFC-T0, Nori, TS-ICL
baseline-forecasting Simple baselines with ForecasterEquivalentDate (equivalent-date / seasonal-naive / moving-average) to benchmark ML models
metric-selection Choosing point, probabilistic, and multi-series aggregated metrics
backtesting-configuration TimeSeriesFold parameters: refit, gap, fold stride, initial train size
hyperparameter-optimization Grid, random, and Bayesian search with TimeSeriesFold and OneStepAheadFold
feature-selection RFECV, SelectFromModel: optional trim of lags, window features, and exog
prediction-intervals Bootstrapping, conformal prediction, quantile regression, interval calibration
statistical-models ForecasterStats with Arima, Sarimax, Ets, Arar: auto-ARIMA, seasonal config
deep-learning-forecasting ForecasterRnn with Keras: create_and_compile_model, LSTM/GRU architectures
drift-detection RangeDriftDetector and PopulationDriftDetector for production monitoring
troubleshooting-common-errors Frequent mistakes AI assistants make with skforecast and their corrections
complete-api-reference Full method signatures and availability matrix for all forecasters

These skills are bundled into llms-full.txt. AI agents that support the Agent Skills spec (such as GitHub Copilot in VS Code) can also load them individually from the skills/ directory.

The table follows the same reading order as llms-full.txt: decide → analyse inputs → build → benchmark → evaluate and tune → refine → specialist paths → operate → reference.

Context files overview

File Audience Description
llms-full.txt Any LLM user Complete context: API + 17 workflow skills
llms.txt LLMs with web search Public index with links to all documentation sections
.github/copilot-instructions.md Contributors (VS Code) Auto-injected into GitHub Copilot
AGENTS.md Contributors (Claude Code, Codex, Aider) Standard agent context file

The context files are auto-generated from maintained source files (tools/ai/llms-base.txt, llms.txt, tools/ai/ai_context_header.md, and skills/) to ensure they stay in sync with the library. They are regenerated on every release.

Tips for better results

  1. Always provide the context URL — Without it, LLMs may hallucinate methods that don't exist or use outdated API names (e.g., ForecasterAutoreg instead of ForecasterRecursive).
  2. Be specific about your forecaster — Mention which forecaster you're using. Parameter names and defaults differ across forecasters.
  3. Mention the version — Say "skforecast 0.24.0" so the LLM doesn't mix advice from older versions.
  4. Validate the output — AI-generated code is a starting point. Use backtesting or an appropriate holdout evaluation to verify model performance.