Skip to content

Installation

skforecast-ai needs Python 3.10 or newer and is available on PyPI. The core package runs the whole forecasting pipeline offline, with no API key; the LLM reasoning layer and the foundation model backend are optional extras.

Python PyPI

To try it before installing anything, open Your first forecast in Google Colab: its first cell installs the package. Open In Colab

Install

Profiling, planning, forecasting, backtesting, comparison and the CLI. No LLM, no network access at run time.

pip install skforecast-ai

Adds ask() and the LLM-guided variants of refine_plan() and create_cv(). Enough for OpenAI, Anthropic, Google, Groq, Ollama and any OpenAI-compatible endpoint.

pip install "skforecast-ai[llm]"

The LLM layer plus the AWS client that Bedrock needs.

pip install "skforecast-ai[bedrock]"

Adds the backend of Chronos-2, the default model of ForecasterFoundation, so it can run in forecast(), backtest() and compare(). Scripts that use it can be generated without it.

pip install "skforecast-ai[foundation]"

To install a specific version, pin it (pip install skforecast-ai==0.4.2). The development version, which may be unstable, installs from GitHub:

pip install git+https://github.com/skforecast/skforecast-ai@main

To work on the code, see the Contribution Guide.

Check the installation

The snippet below downloads a small monthly dataset, evaluates a model on its last 12 months and prints the metrics. It runs in deterministic mode, so it needs no LLM and no configuration.

from skforecast.datasets import load_demo_dataset
from skforecast_ai import ForecastingAssistant

data = load_demo_dataset(verbose=False)
assistant = ForecastingAssistant()
result = assistant.forecast(data=data, target="y", steps=12, test_size=12)
print(result.metrics)

If the installation works, it prints one row of metrics:

  series       MAE       MSE      MASE      MAPE
0      y  0.078442  0.008933  0.802828  0.088608

The command line interface is installed with the package:

skforecast-ai --version

Check the LLM configuration

With the LLM extra installed, check-llm reports how the provider, the model and the credentials resolve, without calling the model. It exits with an error when something is missing, for example the API key:

export OPENAI_API_KEY="sk-..."
skforecast-ai check-llm --llm openai:gpt-5.5

The same check is available in Python as assistant.check_llm(). Configuring the LLM covers every provider and its credentials.

Optional dependencies

Extra Adds Use it for
llm pydantic-ai, the only LLM abstraction the package uses OpenAI, Anthropic, Google, Groq, Ollama and OpenAI-compatible endpoints
bedrock pydantic-ai with its Bedrock support, and boto3 Amazon Bedrock
mcp mcp, the Model Context Protocol SDK The MCP server for coding agents
all Every provider-specific dependency and the mcp extra (full is an alias) Several providers in the same environment
foundation chronos-forecasting, the backend of Chronos-2 Running ForecasterFoundation with its default model

The core dependencies (skforecast, pandas, pydantic, statsmodels, typer and rich) are listed with their versions on PyPI.

Without the foundation extra, compare() leaves ForecasterFoundation out of the automatic candidates with a MissingBackendWarning. Other foundation models (TimesFM, Moirai, TabICL, ...) need their own backend; skforecast names the package to install when it is missing.

Groq

The groq extra (pip install "skforecast-ai[groq]") is kept for backwards compatibility. Groq support is already included in the llm extra.

Next steps

  • Your first forecast

    From raw data to a validated forecast and the script that produced it, with no LLM.

    Your first forecast

  • Ask the assistant

    Connect an LLM to explain results, add domain knowledge to the plan and describe the deployment.

    Ask the assistant