skforecast-ai

About skforecast-ai

The forecasting assistant that shows its work. Give skforecast-ai a time series: it profiles the data, chooses the model with deterministic rules, validates it and returns the forecast together with the skforecast script that produced it. Every decision comes from a rule you can read, the same data always gives the same result, and the script runs on its own with plain skforecast. An optional LLM explains every decision; it never makes them.

36-hour LightGBM forecast of hourly bike sharing users, with its 80% prediction interval against the held-out hours, and the answer of ask() that explains its metrics. Real skforecast-ai outputs, from the animation on ai.skforecast.org.

Installation

pip install skforecast-ai

The LLM layer is optional. To use it, install the extra ([bedrock] for AWS Bedrock):

pip install "skforecast-ai[llm]"

Requires Python 3.10 or newer. More options in the installation guide; to install from source, see the Contribution Guide.

Quick example

Try it without installing anything: Your first forecast runs in Google Colab, from the data to a validated forecast, with no API key.

from skforecast_ai import ForecastingAssistant

from skforecast.datasets import load_demo_dataset

Download demo dataset (monthly)

data = load_demo_dataset(verbose=False)

Profile the data, plan the model, generate the script and run it,

evaluated on the last 12 months. No LLM needed.

assistant = ForecastingAssistant() result = assistant.forecast(data=data, target="y", steps=12, test_size=12)

result.predictions.head()

pred

2007-07-01 0.851486

2007-08-01 1.048000

2007-09-01 0.998147

2007-10-01 1.147577

2007-11-01 1.110845

result.metrics

series MAE MSE MASE MAPE

0 y 0.078442 0.008933 0.802828 0.088608

print(result.code) # the skforecast script that produced this result

That single call chose a ForecasterRecursive with a Ridge estimator, its lags and window features, and ran the script that result.code returns. result.profile and result.plan hold every decision and the rule behind it. Drop test_size to train on all the data and forecast the next 12 months.

Same pipeline from the terminal

# End-to-end forecast

skforecast-ai forecast data.csv --target y --date-column date --steps 12

Only the standalone script, without running it

skforecast-ai forecast-code data.csv --target y --date-column date --steps 12 --output forecast.py

The CLI covers the whole pipeline, including ask() and the prompts of refine_plan() and create_cv(). See Using the CLI.

Ask why (optional LLM)

# pip install "skforecast-ai[llm]"

assistant = ForecastingAssistant(llm="openai:gpt-5.5")

answer = assistant.ask( "Is the MASE good, and why was this estimator chosen?", context=result, )

ask() reads the object you pass and explains it, grounded in the skforecast agent skills. It never changes the result. The dataset is never sent: a result shares only its own predictions and metrics, with a warning. Providers, credentials and local models are covered in Configuring the LLM.

From your coding agent (MCP server)

In Claude Code, install the server and its skill as a plugin:

/plugin marketplace add skforecast/skforecast-ai

/plugin install skforecast-ai@skforecast-ai

Then ask in plain language: *"Forecast the next 12 months of data/sales.csv and tell me how accurate it is."* The agent brings the language model and calls skforecast-ai as tools; every decision still comes from the same deterministic rules, every result comes with the script that produced it, and responses never carry rows of your data. The setups for Cursor, VS Code, Codex and Claude Desktop are in MCP server for coding agents.

A real Claude Code session with the server and its skill, from the agent check of skforecast-ai 0.4.0, abridged: the first three calls share a row, and an ellipsis marks every cut in a response or in the answer. The whole session shows every call, every message of the agent and its complete answer.

In VS Code, this button adds the server and asks for the directory with your data (it needs uv):

Features

A deterministic engine

• [Data profiling](https://ai.skforecast.org/stable/user-guides/agentic-forecasting-step-by-step.html): frequency, missing values, exogenous and categorical variables, and the significant lags from the partial autocorrelation.
• A plan with a rule for every decision: forecaster and estimator, lags, window and calendar features, prediction intervals, metric and cross-validation.
• The code you see is the code that ran: forecast() and backtest() execute the same script that forecast_code() and backtest_code() return. Inspect it, version it or run it with plain skforecast.
• [Model selection](https://ai.skforecast.org/stable/user-guides/agentic-forecasting.html) with compare(): every candidate is backtested with the same data and cross-validation, and the ranking is a plain sort of the metric. For a single series, a seasonal naive baseline is ranked alongside them, so you also see whether a model beats the simplest reasonable forecast.
• Reproducible: the same input always gives the same profile, plan, script and predictions.
• Python or terminal: the CLI runs the same pipeline from a CSV file or URL.
• Coding agents: an MCP server gives Claude Code, Cursor and other MCP clients the same pipeline as tools, with a skill that teaches them to use it. Responses never carry rows of your data.

The engine chooses among the forecasters of skforecast: recursive and direct, multi-series and multivariate, statistical (ARIMA) and foundation models.

An optional LLM layer

• Explains a forecast, a backtest, a comparison, a plan or a script in plain language with ask(), or answers a general forecasting question.
• Refines lags and window features from your domain knowledge with refine_plan(prompt=...).
• Translates a deployment scenario into a backtesting strategy with create_cv(prompt=...).
• Validated before it runs: every LLM output is checked by a Pydantic model; if it is not valid, you get the deterministic result and a warning.
• Any provider through pydantic-ai: OpenAI, Anthropic, Google, Groq, AWS Bedrock, Ollama for local models, or any OpenAI-compatible endpoint.

Methods

All methods work without an LLM except ask(). A prompt is always explicit: nothing is sent to an LLM unless you ask for it.

| plan() | Builds the plan: lags, window and calendar f