Unified Forecasting Architecture

TimeCopilot simplifies the forecasting lifecycle by providing a consistent TimeCopilotForecaster interface that manages diverse model types. This allows developers to swap between traditional statistical methods (AutoARIMA, AutoETS, SeasonalNaive, Theta, Prophet) and modern foundation models (Chronos, TimesFM) without changing the underlying pipeline code. The framework handles the complexity of model initialization, hardware-aware model selection (e.g., choosing between chronos-bolt-small or tiny based on GPU availability), and cross-validation.

Pipeline Execution and Evaluation

To ensure robust performance, the pipeline utilizes rolling cross-validation across multiple windows. By calculating standard error metrics—Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE)—developers can objectively rank models on a leaderboard. Once the best-performing model is identified, the system generates probabilistic forecasts with 80% and 95% prediction intervals, providing a clear view of uncertainty alongside point estimates.

Anomaly Detection and LLM Integration

Beyond forecasting, TimeCopilot includes built-in anomaly detection capabilities. By running detection at a 99% confidence level, the system flags unusual observations, which can be visualized alongside historical data to identify data quality issues or significant events. Finally, the framework offers an optional LLM agent (supporting OpenAI or Anthropic models) that acts as an analytical layer. This agent can interpret the forecasting results, evaluate the selected model against a baseline (like SeasonalNaive), and provide natural language responses to specific analytical queries, bridging the gap between raw model output and actionable business insights.