Building a Forecasting Pipeline with TimeCopilot

In this tutorial, we build an end-to-end forecasting workflow with TimeCopilot. We prepare a panel dataset containing real airline passenger data and a synthetic seasonal series with injected anomalies, then evaluate a diverse collection of statistical, foundation, and optional GPU-based forecasting models. We use rolling cross-validation and multiple error metrics to identify the strongest model, generate probabilistic forecasts with prediction intervals, visualize future trends, and detect unusual observations.

Finally, we explore TimeCopilot's optional LLM agent, which selects a forecasting model and translates its predictions into an accessible analytical response. We install TimeCopilot, UtilsForecast, and Matplotlib to prepare the forecasting environment. We enforce compatible NumPy and SciPy versions to prevent binary conflicts.

We then restart the Colab runtime so the updated libraries load correctly. We import the required libraries, verify the environment, and detect GPU availability. We load the AirPassengers dataset and create a second synthetic seasonal series with injected spikes.

We combine the two series into a panel dataset and set the forecasting horizon and monthly frequency. We configure a diverse collection of statistical, Prophet, and Chronos forecasting models. We select the Chronos model size according to the available hardware and optionally include TimesFM when a GPU is present.

We then initialize TimeCopilotForecaster to manage all models through one consistent interface. We perform rolling cross-validation across three windows to measure each model's forecasting performance. We calculate MAE, RMSE, and MAPE for every series and aggregate the results into a leaderboard.

We identify the model with the lowest mean RMSE for subsequent forecasting and visualization. We generate 12-month probabilistic forecasts with 80% and 95% prediction intervals. We define a reusable plotting function that displays historical values, point forecasts, and uncertainty ranges.

We apply this function to each series to compare its observed history with the predicted future trajectory. We detect anomalies across the panel and visualize the flagged observations in the synthetic series. We optionally initialize the TimeCopilot LLM agent when an OpenAI or Anthropic API key is available.

We use the agent to select a model, evaluate it against SeasonalNaive, and explain the forecast in response to a practical question. The ability to build a unified forecasting pipeline like this has significant implications for businesses and organizations that rely on accurate predictions to inform their decision-making. By automating the process of model selection, evaluation, and interpretation, TimeCopilot enables developers to focus on higher-level tasks and provides stakeholders with actionable insights.