unit8co/darts

Features

• Time Series Forecasting - Provides a comprehensive framework for predicting future numerical values using statistical, machine learning, and deep learning models.
• Covariate Integration - Generates future covariates using pre-defined index encoders, custom functions, or transformers to improve prediction accuracy.
• Time Series Anomaly Detection - Provides tools for identifying outliers and unusual patterns in temporal data using scorers and aggregators.
• Automated Model Selection - Automatically determines the best ARIMA parameters for univariate series using information criteria.
• Deep Learning Forecasting - Fits univariate or multivariate base models to time series data to predict future values using deep learning.
• Dynamic Covariate Integration - Integrates external time-varying data streams and known future events into forecasting models to increase accuracy.
• Global Forecasting Models - Fits a single model across several different target time series to leverage shared patterns across the dataset.
• Hierarchical Reconciliation - Implements hierarchical time series reconciliation to ensure consistency between aggregate and disaggregate forecast totals.
• Hierarchical Series Management - Includes a hierarchical time series manager for organizing nested series and reconciling forecasts across aggregation levels.
• ML-Based Forecasting Models - Provides various deep learning architectures, including MLP-based encoder-decoders, for high-performance time series forecasting.
• Multivariate Forecasting - Predicts multiple interdependent variables simultaneously using vector autoregressive moving average models.
• Probabilistic Forecasting - Generates uncertainty intervals and value ranges instead of single point estimates using parametric distributions or conformal prediction.
• Statistical Forecasting - Implements traditional ARIMA models combining autoregressive lags, differencing, and moving average windows.
• Time Series Machine Learning Frameworks - Offers a unified framework for training and evaluating a wide range of statistical and deep learning models for time series forecasting.
• Global Forecasting Models - Allows a single model to be trained across multiple different time series to leverage shared patterns.
• Global Modeling - Supports global time series modeling to train a single model across multiple different series for improved generalizability.
• Time Series Toolkits - Ships as a comprehensive Python toolkit for forecasting, anomaly detection, and preprocessing of univariate and multivariate time series.
• Time Series Transformations - Provides tools for cleaning and transforming raw temporal data through filtering, scaling, and missing value imputation.
• Hierarchical Forecast Reconciliation - Adjusts forecasts across nested levels of aggregation to ensure that child series sum up to their parent totals.
• Forecasting Backtesting - Simulates historical performance by sliding a prediction window across the data to validate accuracy before live deployment.
• Stacking Ensembles - Combines multiple individual model predictions using a secondary regression model to learn optimal weights for the final forecast.
• Gradient Boosted Tree Forecasting - Applies gradient boosted tree algorithms to predict future values using lagged target and covariate values.
• Model Fine-Tuning - Adjusts the weights of pre-trained foundation models to improve performance on specific temporal datasets.
• Hyperparameter Optimization - Tunes model settings using grid search or external optimization frameworks to maximize predictive performance.
• Time Series Tabularization - Transforms time series data into a tabular format of lagged features to enable the use of standard regression algorithms.
• Time Series Backtesting - Evaluates model performance by simulating predictions on historical data to validate accuracy before live deployment.
• Baseline Benchmarking - Generates simple benchmark predictions to evaluate the performance of complex forecasting models.
• Forecast Aggregation - Merges forecasts from multiple different models to improve overall accuracy through naive or learned weighting.
• Intermittent Demand Forecasting - Provides specialized forecasting for sparse or infrequent data using Croston method variants.
• Linear-Based Forecasting - Predicts future values of univariate or multivariate series using a linear-based model with covariates.
• Linear Decomposition Models - Predicts future values of univariate or multivariate series using a linear decomposition approach with covariates.
• MLP-Based Forecasting - Predicts future values by mixing historical data, known future inputs, and static contextual information using an MLP architecture.
• Model Transferability - Allows trained models to be transferred to different input series at prediction time for zero-shot or few-shot forecasting.
• Multi-Rate Sampling Models - Predicts future values using multi-rate sampling and multi-scale interpolation to reduce computational cost.
• Neural Basis Expansion Analysis - Predicts future values using neural basis expansion analysis with generic or interpretable architectures.
• Random Forest Forecasting - Predicts future values of a target series using random forest regression based on historical lags.
• Reduction-Based Forecasting - Converts time series data into tabular feature-target pairs to apply standard regression or classification algorithms.
• RNN-Based Forecasting - Uses RNN encoders and fully connected networks with LSTM or GRU architectures to produce predictions.
• Temporal Convolutional Networks - Predicts future values using a dilated Temporal Convolutional Network that incorporates past covariates.
• Temporal Fusion Transformers - Predicts future values using a Temporal Fusion Transformer architecture to provide interpretable results.
• Zero-Shot Forecasting - Enables predictions on unseen time series using pretrained foundation models without dataset-specific training.
• Regression Model Wrapping - Transforms compatible regression models into time series forecasting models by tabularizing the series data.
• Transfer Learning - Trains models on large datasets and applies the learned weights to forecast new, unseen time series.
• Covariate Encoders - Automatically generates future-known features by applying encoders and transformers to the temporal index during model creation.
• Time Series Imputation - Interpolates or fills gaps in time series data where values are represented as NaNs.
• Time Series Manipulation - Performs operations like splitting, slicing, arithmetic, and stacking to modify series attributes and markers.
• Fourier-Series Seasonality - Models complex seasonal patterns using a state space framework and Fourier series.
• Dynamic Time Warping - Compares two temporal sequences that vary in speed or time to determine how closely they align.
• Forecasting Performance Metrics - Calculates accuracy and error metrics to compare the effectiveness of different forecasting models.
• Time Series Filtering - Applies Kalman, Gaussian Process, or Moving Average filters to remove noise and infer underlying states.
• Forecasting and Analytics - Forecasting and anomaly detection library for time series.
• Time Series - Library for easy time series manipulation and forecasting.
• Time Series Analysis - User-friendly library for forecasting and anomaly detection.
• Data Validation and Anomaly Detection - Library for time series forecasting and anomaly detection.
• Time Series Analysis - Unified library for time series forecasting and neural networks.
• Analysis Toolkits - Unified library for forecasting and anomaly detection tasks.

Star history