TimeCopilot: Building a Forecasting Pipeline with Foundation Models and Anomaly Detection

TimeCopilot is a Python library for building an end-to-end time series forecasting pipeline. It combines classical statistical methods, pretrained foundation models and an optional LLM agent that picks the best approach and explains the result. The library targets analysts who need to compare different approaches on real data quickly, without spending time tuning each model by hand.

How the pipeline works

The authors demonstrate TimeCopilot on two datasets: a real airline passenger dataset (a classic benchmark) and a synthetic seasonal series with deliberately embedded anomalies. This choice allows an honest test of the system, including on data where it is known in advance what should be detected.

The workflow includes several stages:

• Loading and preparing panel data
• Training models via rolling cross-validation
• Evaluating across several metrics: MAE, MAPE, RMSE
• Generating probabilistic forecasts with confidence intervals
• Detecting anomalies with visualization
• Running an LLM agent to select and explain the best model

Three classes of models

TimeCopilot supports three types of models that can be compared within a single experiment.

Statistical models are classical methods: ARIMA, ETS, Theta and others. They are easy to interpret, work on small datasets with pronounced seasonality and do not require a GPU. They give a reasonable baseline on almost any task.

Foundation models are pretrained time series architectures trained on large data corpora. They are the equivalent of language models for numerical sequences: they can work out of the box, without retraining on a specific dataset.

GPU-accelerated models are neural network models (for example, N-BEATS, TiDE) for large datasets with long forecast horizons. They are optional and become relevant when statistical methods are no longer enough.

Evaluation relies on rolling cross-validation with an expanding window: each model is trained on the same history and tested on the next horizon. This reveals real generalization ability rather than mere fitting to the training data.

Anomalies and uncertainty

Instead of a single forecast line, TimeCopilot generates probabilistic predictions: a range of values with a given probability (80% or 95%). For inventory, load or budget planning this is practically useful: an organization sees the pessimistic and optimistic corridors at the same time.

A built-in anomaly detector runs alongside the forecast. On the synthetic series with artificial outliers, the system correctly separated normal behavior from abnormal behavior. Anomaly flags are overlaid directly on the series visualization, which simplifies auditing and sharing results within a team.

The LLM agent explains its choice

The most unusual part is the optional LLM agent. After all models are evaluated, it examines the cross-validation results, selects the best model for the specific series and generates a text explanation of why this model is more accurate than the others.

"The foundation model outperformed ARIMA because of nonlinear

seasonality with growing amplitude" — the agent produces this kind of explanation automatically.

This is a step toward AutoML with an auditable result: the analyst gets not just a recommendation but a rationale that can go into a report or be passed to a stakeholder without further interpretation.

What this means

TimeCopilot lowers the barrier to entry for advanced forecasting: there is no need to tune dozens of models by hand or interpret metrics yourself, as the LLM agent takes over part of that work. For analytics teams working with business metrics, demand or operational KPIs, it is a ready-made framework for a production-ready pipeline with a human-readable rationale for every decision.