GETTING STARTED
CAPABILITIES
- Forecast
- State of the art anomaly detection
- Online (Real-Time) Anomaly Detection
DEPLOYMENT
TUTORIALS
- Anomaly detection
- Exogenous variables
- Training
- Fine-tuning
- Validation
- Uncertainty quantification
- Special topics
- Computing at scale
USE CASES
Changelog
Complete list of changes for each version of the Nixtla client.
Changelog Overview
Below you’ll find the complete list of changes for each version of the Nixtla client. Expand any version to see details about new features, improvements, changes, or deprecations, along with links to full release notes.
Feature Enhancements
Online anomaly detection
We introduce the online_anomaly_detection method, which allows you to define a detection_size on which to look for anomalies.
Feature Enhancements
Persisting fine-tuned models
You can now run an isolated fine-tuning process, save the model, and use it afterward in all of our methods:
Fine-tune the model
Save the model
Use it in forecast, cross_validation, or detect_anomalies
zstd compression
All requests above 1MB are automatically compressed using
zstd, which helps when sending large data volumes or with slower connections.
Refit argument in cross_validation
Set refit=False to fine-tune the model only on the first window in
cross_validation. This significantly decreases computation time.
Feature Enhancements
Integer and custom pandas frequencies
The client now supports integer timestamps and frequencies, and custom pandas timestamps
(including CustomBusinessHour).
Usage method
You can programmatically retrieve your current API call count and limits using the new usage method.
Historic exogenous in cross validation
The cross_validation method now accepts the hist_exog_list parameter, enabling definition of historical exogenous features.
Feature Enhancements
Fine-tune depth
Specify the fine-tuning depth through the finetune_depth parameter in
forecast and cross_validation.
Feature Enhancements
V2 API endpoints
The client now uses V2 API endpoints, providing lower latency.
orjson serialization
Payload serialization now uses orjson for performance improvements, especially with exogenous features.
Historical exogenous features
Historical exogenous features (hist_exog_list) are supported in the forecast method.
Feature contributions
Activate feature contributions by setting:
feature_contributions = True
in the forecast method.
Feature Enhancements
Cross validation endpoint
The cross_validation method now performs a single API call instead of individual calls per window.
Changes & Deprecations
Important:
The nixtlats package has been deprecated in favor of the nixtla package.
Changes & Deprecations
Deprecation of TimeGPT class
Replace TimeGPT with NixtlaClient. Also note:
- Parameters renamed:
token→api_key,environment→base_url - Method renamed:
validate_token→validate_api_key - Update environment variables to match new parameter names.
Changes & Deprecations
Renamed fine-tuning parameter:
- From
finetune_stepstofewshot_steps
(This change was later reverted for compatibility reasons).
Feature Enhancements
Enhanced fine-tuning capability with new parameters:
finetune_lossfinetune_steps
Feature Enhancements
Implemented num_partitions parameter for improved resource optimization.
Feature Enhancements
New model: timegpt-1-long-horizon
Cross-validation support for multiple windows
Improved retry behavior
Using parameters:
max_retriesretry_intervalmax_wait_time
Environment tokens are handled automatically
Introduced a FAQ section