awslabs/gluonts

Features

• Deep Learning Forecasting - Implements deep neural network architectures like CNNs and RNNs specifically for univariate and multivariate time series prediction.
• Probabilistic Forecasting - Predicts future time series values as probability distributions instead of point estimates to quantify uncertainty.
• Calendar Feature Extraction - Generates seasonal, holiday, and lag features from timestamps to improve forecasting accuracy.
• Feature Assembly - Combines static and dynamic features into tensors and applies embedding layers to categorical data.
• Hierarchical Time Series Modeling - Implements aggregation matrices to maintain consistency between leaf and aggregate levels in tree-structured time series.
• Probabilistic Models - Provides a toolkit for representing time series data as probability distributions and calculating their log-densities.
• Time Series Model Training - Implements training procedures for fitting deep learning models to historical temporal sequences.
• Temporal Test Window Generators - Creates overlapping or non-overlapping temporal windows from datasets to evaluate model performance over time.
• Time Series Data Engineering - Provides a comprehensive pipeline for preparing raw temporal data and transforming it into tensors for machine learning models.
• Time Series Feature Engineering - Transforms raw target series and features into latent representations using neural network encoders.
• Time Series Forecasting - Provides the core capability to generate time series forecasts as probability distributions, quantiles, or discrete samples.
• Forecast Evaluation - Includes a dedicated suite for measuring forecast accuracy and probabilistic coverage using specialized metrics like quantile loss.
• Probabilistic Forecast Evaluators - Measures forecast accuracy using specialized probabilistic metrics like quantile loss and continuous rank probability scores.
• Zero-Shot Forecasting - Generates time series predictions for new datasets using pretrained models without requiring task-specific training data.
• Time Series ML Toolkits - Implements a comprehensive set of tools for scaling, transforming, and pipelining temporal data for machine learning models.
• Dataset Metadata Schemas - Defines schema types and properties for describing time series targets, features, and prediction lengths.
• Hierarchical Dataset Conversions - Transforms hierarchical time series structures into standardized dataset formats for probabilistic forecasting.
• Time Series Data Loading - Reads time series datasets from files or dictionaries and organizes them by frequency and dimension.
• Gaussian Time Series Distributions - Represents data using Gaussian probability distributions to calculate log-density and quantiles for forecasting.
• Time-Series Aggregations - Calculates higher-level totals from leaf-level data using summation matrices for hierarchical consistency.
• Hierarchical Time Series Organizations - Represents tree-structured time series data using aggregation matrices to link leaf levels to higher levels.
• Temporal Dataset Splitting - Divides time series datasets into training and testing subsets using fixed temporal offsets or calendar dates.
• Distribution Statistics - Calculates statistical properties such as cumulative distribution, log-density, and quantiles for time series events.
• Baseline Forecasting Models - Generates reference predictions using constant values or means to evaluate the performance of complex models.
• Categorical Distributions - Represents categorical distributions to compute log-densities and draw random samples across batch shapes.
• Distribution Domain Bijections - Employs bijective mappings to transform neural network outputs into valid probability distribution domains.
• Distribution Parameter Projections - Maps dense layer outputs to the required shapes and arguments to construct probabilistic distributions.
• Causal Convolutions - Uses 1D causal convolutions to ensure outputs depend only on past inputs, preventing temporal information leakage.
• Dataset Statistics Analyzers - Calculates descriptive statistics, including target means and feature distributions, for time series datasets.
• Deep Learning Architectures - Assembles complex time series models by combining modular deep learning blocks like CNNs, RNNs, and MLPs.
• External Forecasting Library Integrations - Wraps third-party forecasting libraries like Prophet and R forecast into a unified probabilistic interface.
• Baseline Forecast Generators - Produces static forecasts using a fixed set of samples or single constant values as a benchmark.
• Probabilistic Accuracy Metrics - Computes the Continuous Rank Probability Score to measure the accuracy of probabilistic forecasts.
• Incremental Model Updating - Enables refining existing predictors by updating model weights with new data without restarting the training process.
• Lazy-Loaded Time Series Batching - Implements a lazy-transformation pipeline to efficiently prepare time series batches for deep learning backends.
• Binned Distribution Loss Calculators - Computes loss for probabilistic forecasts by applying label smoothing to the cross-entropy of bin probabilities.
• Gated Convolutional Layers - Uses gated mechanisms and dilated convolutions to capture long-range temporal dependencies in time series data.
• Mixture Model Estimation - Combines different probability distributions into a single mixture using weighted probabilities to represent complex patterns.
• Input Parameter Specifications - Provides declarative definitions of expected shapes and data types for model inputs to ensure tensor consistency.
• Model State Serialization - Saves trained model architectures and weights to disk for persistence and deployment.
• Non-Parametric Forecasting - Implements non-parametric forecasting by sampling from historical observations using exponential or uniform kernels.
• Predictor Serialization - Deno restores a saved predictor from a file and assigns it to an available GPU or CPU.
• Dirichlet Distributions - Represents a Dirichlet distribution to calculate log-densities and draw random samples for probabilistic modeling.
• Recurrent Neural Networks - Builds RNN, LSTM, and GRU blocks with bidirectional processing to model sequential dependencies in temporal data.
• Zoneout Regularization - Implements the zoneout regularization mechanism for recurrent neural network cells to prevent overfitting.
• Invertible - Implements invertible residual networks to map tensors between different probability distributions.
• Local Model Training - Supports training separate, personalized models for each individual time series within a dataset.
• Baseline Forecasting - Produces simple baseline predictions using means or moving averages to establish a performance floor.
• Probabilistic Prediction Analyzers - Computes quantiles, means, and medians from predicted distributions to analyze future value estimates.
• Uncertainty Interval Measurements - Assesses the quality of uncertainty intervals using Mean Scaled Interval Score and coverage metrics.
• Sequence Decoders - Transforms encoded representations into forecasted time steps to predict future quantiles using multilayer perceptrons.
• Statistical Forecasting - Wraps statistical models like AutoARIMA and AutoETS to perform forecasting within a unified interface.
• Apache Arrow Processing - Utilizes the Apache Arrow columnar memory format for high-performance data processing and streaming.
• Multivariate Series Alignment - Aligns multiple univariate time series by timestamps and applies padding to create multivariate datasets.
• Standard Dataset Loaders - Downloads and loads standardized time series datasets from remote or local storage for model benchmarking.
• Dataset Preparation - Provides utilities to load raw time series data and partition it into training and testing sets.
• Laplace Time Series Distributions - Represents time series data using a Laplace distribution to calculate cumulative distribution functions and log-densities.
• Time Series Data Normalization - Includes normalization techniques such as min-max scaling and average absolute value scaling for time series inputs.
• Time Series Management - Organizes collections of time series data using specific access patterns for model training.
• Time Series Transformations - Converts start-date fields into frequency-based periods to align time series observations.
• Parallel Workload Distribution - Distributes forecasting tasks across multiple CPU cores to increase data throughput for large datasets.
• Bijective Transformations - Implements invertible mathematical mappings to ensure probabilistic consistency during data transformations.
• Degenerate Distributions - Represents a degenerate distribution to calculate log-density and quantiles for fixed values.
• Empirical Distributions - Represents a distribution based on samples to compute cumulative distribution functions and quantiles.
• Low-Rank Multivariate Gaussian Distributions - Represents a multivariate Gaussian distribution using a diagonal and low-rank matrix to reduce complexity.
• Binned Distributions - Defines probability distributions using discrete bin centers and probabilities to approximate continuous values.
• Distribution Function Calculators - Calculates cumulative distribution functions, means, and quantiles relative to binned distributions.
• Generalised Pareto Distributions - Implements the Generalised Pareto distribution to calculate cumulative distribution functions and random samples.
• Beta Distributions - Computes log-density and statistical moments for Beta-distributed probabilistic time series data.
• Logit-Normal Distributions - Represents a logit-normal distribution to compute log-densities, quantiles, and random samples for forecasting.
• Dirichlet Multinomial Distributions - Represents a discrete multivariate probability distribution to calculate log-densities and draw random samples.
• Power Transformations - Provides Box-Cox power transformations to stabilize variance and normalize univariate time series data.
• Interval and Quantile Conversions - Translates prediction interval levels into quantile levels to standardize uncertainty measurements.
• Dynamic Linear Modeling - Implements linear dynamical systems as distributions to represent time series via latent states and emission coefficients.
• Time Series Analysis - Probabilistic time series modeling and deep learning.
• Time Series Analysis - Toolkit for probabilistic time series modeling.
• Analysis Toolkits - Probabilistic modeling and forecasting library.

Star-Verlauf