TimesFM is a pretrained decoder-only transformer foundation model for time-series forecasting, developed by Google Research. The repository provides multiple model versions (v1.0, v2.0, v2.5, v3.0) with implementations in PyTorch, Flax/JAX, and MLX (Apple Silicon). It supports zero-shot univariate and multivariate forecasting with covariates, quantile outputs, and various preprocessing options (RevIN normalization, linear detrending, symmetric averaging). The codebase includes a high-level forecaster API, CLI scripts for CSV forecasting and system checks, fine-tuning examples (LoRA/DoRA), and comprehensive benchmarks (GIFT-Eval, TIME, FEV). While primarily an application library, it also contains research-oriented components like benchmark runners and notebooks, making it a hybrid project with a strong application focus.
AssetOpsBench is a unified open framework for building, orchestrating, and evaluating domain-specific AI agents in industrial asset operations and maintenance (Industry 4.0). It provides a suite of MCP (Model Context Protocol) servers that expose domain tools for IoT telemetry, work order lifecycle, failure mode analysis, time-series forecasting/anomaly detection, vibration diagnostics, and utility functions. The framework includes multiple agent runners (plan-execute, Claude, OpenAI, Deep, Stirrup, OpenCode, direct LLM) that consume these tools, and a comprehensive evaluation pipeline with LLM-as-judge and deterministic structured-answer scorers. The system is designed to run reproducible benchmark scenarios with ground truth, producing leaderboard reports for agent comparison. It is accepted at KDD 2026 and powers multiple competitions. Architecturally, the system is a collection of Python packages with 14 CLI entry points defined in pyproject.toml. Each MCP server is a standalone FastMCP application using stdio transport, backed by CouchDB for persistence. Agent runners share a common trajectory model and observability layer (OpenTelemetry with file-based export). The evaluation pipeline is pluggable with a scorer registry. The codebase also includes a significant data/research component: a time-series feature extraction library (130+ extractors), a model catalog with foundation models (TinyTimeMixer, sktime wrappers), synthetic data generation for vibration signals, and showcase notebooks demonstrating end-to-end model lifecycle. The hybrid nature makes it both a software benchmark framework and a research platform for industrial AI.
pytorch-forecasting is a deep learning library for time series forecasting, providing implementations of state-of-the-art models such as Temporal Fusion Transformer, N-BEATS, N-HiTS, DeepAR, PatchTST, TiDE, TimeXer, and many others. Built on PyTorch and PyTorch Lightning, it offers a complete pipeline from data preprocessing (TimeSeriesDataSet, encoders, samplers) through model training (with GPU support, callbacks, hyperparameter tuning via optuna) to prediction and evaluation (point, quantile, and distributional metrics). The library supports multivariate forecasting, static and time-varying covariates, categorical features, and multiple targets. A major v2 refactoring is underway, introducing a new data module layer and simplified package interface, though the v2 API is explicitly marked as experimental and unstable. The codebase is well-tested with automated integration tests across all models and metrics, and employs a composition pattern for metrics (supporting weighting, aggregation, and distribution losses) and a layered architecture for models (base classes, package containers, building blocks).
Time-MoE is a library/framework for training and deploying time series foundation models using a Mixture-of-Experts (MoE) architecture. It provides tools for training the TimeMoE model from scratch or fine-tuning pretrained versions, evaluating on benchmark datasets, and generating forecasts. The codebase is organized into modules for models, datasets, training utilities, and I/O operations, with a clear separation of concerns. Key features include efficient handling of large sequential datasets via sharded binary files, a custom MoE layer with load balancing loss, and integration with Hugging Face Transformers for model management and generation.
ThymeBoost is a Python library for time series forecasting that applies a gradient boosting approach to decompose and predict trend, seasonality, and exogenous components. Unlike traditional additive models, it iteratively fits these components on residuals across multiple rounds, allowing flexible mixing of estimators (e.g., ARIMA for trend, Fourier for seasonality, OLS for exogenous). The library exposes a clean API through the `ThymeBoost` class with methods for fitting, predicting, optimizing hyperparameters, ensembling, and online learning. It also includes an experimental `AutoForecast` class and a set of unit tests for trend models. While it is packaged as a reusable library (application), the repository includes Jupyter notebooks and sample datasets, giving it a hybrid character that also supports data exploration and research use cases. The codebase demonstrates solid software engineering with abstract base classes for each component type, operator overloading for composing models, and a well-structured boosting loop. However, it shows signs of single-developer maintenance: incomplete error handling, commented-out code, and experimental features. The dependencies are reasonably current, but Python 3.7 compatibility may limit adoption on newer runtimes.