TimeGPT-1: production ready pre-trained Time Series Foundation Model for forecasting and anomaly detection. Generative pretrained transformer for time series trained on over 100B data points. It's capable of accurately predicting various domains such as retail, electricity, finance, and IoT with just a few lines of code 🚀.
OTHER License
Bot releases are hidden (Show)
TimeGPT
class in favor of NixtlaClient
.NixtlaClient.validate_token
method in favor of NixtlaClient.validate_api_key
.short-horizon
and long-horizon
models in favor of timegpt-1
and timegpt-1-long-horizon
respectively.fewshot_steps
and fewshot_loss
in favor of finetune_steps
and finetune_loss
respectively.TIMEGPT_TOKEN
environment variable in favor of NIXTLA_API_KEY
.time_col
preserve their type in the outputs (timestamp), previously they were cast to string.NixtlaClient.weights_x
is now a list of lists if num_partitions != None
, where each element corresponds to the weights for a specific partition.Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.5.2...v0.6.0
Published by AzulGarza 4 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.5.1...v0.5.2
Published by AzulGarza 6 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.5.0...v0.5.1
Published by AzulGarza 6 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.4.0...v0.5.0
Published by jmoralez 6 months ago
We're deprecating the nixtlats
package in favor of nixtla
. Please ensure you make the following changes:
pip install nixtla
from nixtla import NixtlaClient
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.3.0...v0.4.0
Published by AzulGarza 7 months ago
Deprecation of TimeGPT
Class:
In an effort to streamline our API and align with industry best practices, we're deprecating the TimeGPT
class in favor of the new NixtlaClient
class. This change is designed to provide a more intuitive and powerful interface for interacting with our services.
Before:
from nixtlats import TimeGPT
# Initialize the TimeGPT model
timegpt = TimeGPT()
After:
from nixtlats import NixtlaClient
# Initialize the NixtlaClient
nixtla = NixtlaClient()
Renaming of Configuration Parameters:
To enhance clarity and consistency with other SDKs, we've renamed the token
parameter to api_key
and environment
to base_url
.
Before:
timegpt = TimeGPT(token='YOUR_TOKEN', environment='YOUR_ENVIRONMENT_URL')
After:
nixtla = NixtlaClient(api_key='YOUR_API_KEY', base_url='YOUR_BASE_URL')
Introduction of NixtlaClient.validate_api_key
:
Replacing the previous NixtlaClient.validate_token
method, this update aligns with the new authentication parameter naming and offers a straightforward way to validate API keys.
Before:
timegpt.validate_token()
After:
nixtla.validate_api_key()
Environment Variable Changes:
In line with the renaming of parameters, we've updated the environment variables to set up the API key and base URL. The TIMEGPT_TOKEN
is now replaced with NIXTLA_API_KEY
, and we've introduced NIXTLA_BASE_URL
for custom API URLs.
Backward Compatibility & Future Warnings:
These changes are designed to be backward compatible. However, users can expect to see future warnings when utilizing deprecated features, such as the TimeGPT
class.
finetune_steps
and finetune_loss
parameters were renamed to fewshot_steps
and fewshot_loss
. Additionally, the model parameter values changed from short-horizon
and long-horizon
to timegpt-1
and timegpt-1-long-horizon
, with an emphasis on preserving backward compatibility. In version 0.3.0, these changes are deprecated in favor of reverting to the original parameter names and values, ensuring a seamless transition for existing users.Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.2.0...v0.3.0
Published by AzulGarza 7 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.21...v0.2.0
Published by AzulGarza 8 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.20...v0.1.21
Published by AzulGarza 8 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.19...v0.1.20
Published by AzulGarza 11 months ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.18...v0.1.19
Published by AzulGarza 11 months ago
Release of new forecasting methods. Among the updates, we've unveiled the timegpt-1-long-horizon
model, crafted specifically for long-term forecasts that span multiple seasonalities. To use it, simply specify the model in your methods like so:
from nixtlats import TimeGPT
# Initialize the TimeGPT model
timegpt = TimeGPT()
# Generate forecasts using the long-horizon model
fcst_df = timegpt.forecast(..., model='timegpt-1-long-horizon')
# Perform cross-validation with the long-horizon model
cv_df = timegpt.cross_validation(..., model='timegpt-1-long-horizon')
# Detect anomalies with the long-horizon model
anomalies_df = timegpt.detect_anomalies(..., model='timegpt-1-long-horizon')
Choose between timegpt-1
for the first version of TimeGPT
or timegpt-1-long-horizon
for long horizon tasks..
You can dive deeper into your forecasting pipelines with the new cross_validation
feature. This method enables you to validate forecasts across different windows efficiently:
# Set up cross-validation with a custom horizon, number of windows, and step size
cv_df = timegpt.cross_validation(df, h=35, n_windows=5, step_size=5)
This will generate 5 distinct forecast sets, each with a horizon of 35, stepping through your data every 5 timestamps.
The new retry mechanism allows the making of more robust API calls (preventing them from crashing with large-scale tasks).
max_retries
: Number of max retries for an API call.retry_interval
: Pause between retries.max_wait_time
: Total duration of retries.timegpt = TimeGPT(max_retries=10, retry_interval=5, max_wait_time=360)
The TimeGPT
class now automatically infers your TIMEGPT_TOKEN
using os.environ.get('TIMEGPT_TOKEN')
, streamlining your setup:
# No more manual token handling - TimeGPT has got you covered
timegpt = TimeGPT()
For more information visit our FAQS section.
Questions? We've got answers! Our new FAQ section tackles the most common inquiries, from integrating exogenous variables to configuring authorization tokens and understanding long-horizon forecasts.
New Features:
Fixes:
Documentation and Miscellaneous:
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.17...v0.1.18
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.16...v0.1.17
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.15...v0.1.16
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.14...v0.1.15
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.13...v0.1.14
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.12...v0.1.13
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.11...v0.1.12
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.10...v0.1.11
Published by FedericoGarza about 1 year ago
Full Changelog: https://github.com/Nixtla/nixtla/compare/v0.1.9...v0.1.10
Published by FedericoGarza about 1 year ago
Hot Fix pydantic