TiRex Time Series Forecasting Model

TiRex provides zero-shot forecasting capabilities by leveraging the xLSTM architecture. It is designed to capture long-range dependencies and temporal dynamics, making it suitable for highly irregular or complex time series.

To master forecasting with TiRex, follow these three pillars:

1. Theory: Understand the xLSTM architecture, frequency-based resampling, patching, and how zero-shot generalization is achieved.
2. Workflow: Follow the structured playbook for data preparation (scaling and adapters), configuring forecast parameters (quantiles, batching, devices), and using evaluation scripts (sMAPE, MASE, CRPS).
3. Practice: Implement forecasting by preparing data, running models in zero-shot mode using load_model, and evaluating/visualizing results.

TiRex is a zero-shot forecasting model designed to handle both long and short horizons. It leverages enhanced in-context learning to provide forecasting capabilities without requiring task-specific fine-tuning. The project includes specialized models for both forecasting (regression) and time series classification, and is built upon the xLSTM (Extended Long Short-Term Memory) architecture.

Key resources available:

• Forecasting Playbook: Theory and hands-on guides for time series forecasting.
• Benchmarks: Evaluation results using datasets and metrics such as sMAPE, MASE, and CRPS.
• Deployment: Guidance on packaging, ONNX/TensorRT export, and edge deployment.

TiRex performs time series regression by leveraging a frozen, pre-trained TiRex forecasting model as a feature extractor. The model uses xLSTM to extract embeddings from time series data (including highly irregular or complex series). These embeddings are then passed into a task-specific regression head that maps the features to continuous target values.

Key architectural concepts:

• Zero-shot feature extraction: The core forecasting model remains frozen.
• Embedding extraction: Features are extracted and aggregated from the time series.
• Regression head: A trainable layer added on top of the frozen embeddings to predict continuous values.

TiRex is a 35M parameter pre-trained time series forecasting model built on the xLSTM (Extended Long Short-Term Memory) architecture.

It is designed for Zero-Shot Forecasting, meaning it can perform time series forecasting on new datasets immediately after download without requiring any task-specific training or fine-tuning.

Key capabilities include:

• Quantile Predictions: It provides both point estimates and quantile estimates (useful for uncertainty quantification).
• Multi-Horizon Performance: It is optimized for both long-term and short-term forecasting horizons.
• Sequence Modeling: It leverages xLSTM's ability to capture long-range dependencies and complex patterns in time series data.

The classification process follows a specific sequence of transformations to convert raw time series into class predictions:

1. Hidden States Extraction: Extract hidden states from all layers of the TiRex forecasting model.
2. Sequence Aggregation: Apply mean pooling across the sequence dimension.
3. Embedding Normalization: Normalize embeddings to account for different feature spaces across layers.
4. Layer Aggregation: Concatenate the features from each layer.
5. Optional Embedding Augmentations:
   • Differenced-series embeddings: To identify subtle patterns.
   • Absolute sample statistics: To restore scale information lost during normalization.
6. Simple Classifier Prediction: A classification head (such as Random Forest, Linear Layer, or Gradient Boosting) is trained on the final embedding to map it to class labels.

TiRex provides three primary regressor implementations within the tirex.models.regression module. All regressor classes follow a consistent API pattern for training, inference, and persistence.

Available models:

• TirexLinearRegressor: Linear regression implementation.
• TirexRFRegressor: Random Forest regression implementation.
• TirexGBMRegressor: Gradient Boosting Machine regression implementation.

Each model supports the following methods:

• fit(X, y): Trains the model on the provided features X and targets y.
• predict(X): Generates predictions for the input features X.
• save_model(path): Serializes and saves the trained model to the specified path.
• load_model(path): Loads a previously saved model from the specified path.

The tirex.models.classification module provides three main classifier implementations for time-series classification tasks:

1. TirexLinearClassifier: A linear model implementation.
2. TirexRFClassifier: A Random Forest implementation.
3. TirexGBMClassifier: A Gradient Boosting Machine implementation.

All classifiers share a common interface for training, inference, and persistence.

To understand the underlying mechanics of how TiRex classifies time series, refer to the theoretical documentation. Key concepts include:

• Zero-shot protocol: Using a frozen TiRex forecasting model as a feature extractor.
• Embedding strategies: How embeddings are extracted and aggregated.
• Multivariate handling: How the model processes multivariate time series.
• Augmentation: Embedding augmentation techniques used to improve performance.
• Transformation pipeline: The conceptual flow of the classification transformation.