본문으로 건너뛰기
버전: docs v25.02

FCST Release Note

v2.1.2​

Feb. 25. 2025

Bug Fixes Fix library version error

  • Change library version: prophet from 1.1.4 to 1.1.6
  • Delete version info of lightgbm

Compatibility: ALO 2.5.1 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11

v2.1.1​

Jun. 24. 2024

Bug Fixes Fix importing error

  • Fix tsfresh import error in GPU infrastructure

Compatibility: ALO 2.5.1 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11

v2.1.0​

Apr. 11. 2024

New Features  

Evaluation Report newly developed

  • You can summarize and visualize the analysis results through the Evaluation Report.
  • Wrapper can load datasets and configs by interoperating with alo's assets.

Improvements Function and parameter improvements

  • Logging for memory and runtime is updated.
  • Label encodier is added as categorical encoder.
  • metrics.py is developed.
  • Timestamp is added to inference output.

Compatibility: ALO 2.3.1 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11

v2.0.0​

Mar. 14. 2024

New Features  

Develop Preprocess for biz forecasting

  • Through code optimization, the computational time has been reduced by 1/16 compared to the previous version, and individual preprocessing method can now be turned off.
  • Preprocessing functionality has been integrated into the train/inference assets, leading to difficulties in code management and slow preprocessing speeds. To address this feedback, refactoring has been conducted, and preprocessing has been separated into a separate asset.  

Improvements Refactor code architecture

  • Moved functionality unrelated to the model from the nbeats.py model file to forecasting.py.
  • Excluded feature_importance, which was executed in the deep learning model, to enhance speed (Deep learning XAI will be updated in the future).
  • Move HPO, train, and cross-validation functions to base_model.py to improve the structure for common use when adding new models in the future.
  • Changed HPO function from SKOPT to Optuna for reduced execution time

Compatibility: ALO 2.3.0 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11

v1.3.0​

Dec. 29. 2023

New Features

Support weekly time scale and regularizing period function

  • In addition to the existing monthly, daily, and hourly time scales, now it supports the forecasting in weekly time scale. By the way, what if your weekly time series was made on an irregular time basis (that is, the data was measured on Monday this week but on Tuesday last week etc.)? Don't worry. The AI content automatically makes it regular : )

Feature expansion

  • Feature expansion with the tsfresh library came back. Now you can expand x features for a better forecasting result.

Support national holidays in DL model

  • Now you can make the DL model consider national holidays, as well.

Support Enable/disable function

  • You can tweak the operation of the AI content by enabling or disabling various functionalities like HPO, feature expansion, and many reporting functions. The details are well defined in the configuration file.

Checking Memory

  • Now it better manage the memory consumption thanks to various Python memory optimization methods like memory returning.

Feature importance

  • The feature importance reports and model performance reports are now being summarized in a single report file as well. It is benefitial especially when you forecast a bunch of time serieses at once.

Compatibility: ALO 2.1.0

v1.1.0​

Nov. 22. 2023

New Features

Supports multivariates learning

  • You can make a single forecasting model which learned the overall characteristics of a set of multiple time serieses. You have to choose the DL model to use this functionality.

Supports multiple time series forecasting

  • Please do not be confused. Unlike the multivariates learning, you will get a set of forecasting models each of them corresponds to the respective time series among the entire time series set. Both of the GAM model and DL model support this.

Supports the interpolation of missing values in time index

Supports the time series whose time index is irregular

  • ex. Weekly data measured on an arbitrary week day

Compatibility: ALO 2.1.0

v1.0.0​

Oct. 16. 2023

New Features

Dynamic regression-based time-series forecasting

  • The service additionally uses a set of independent variables to predict the future target variable

A variety of time scale support

  • Forecasting either in hourly, daily, or monthly time scale

**Easy of use **

  • Users can make an accurate forecast for most use cases with just a simple set of inputs - the input time-series and only 6 parameters (forecasting period, time scale, target country code, and the column names of time stamp, target variable, and holiday variables)

NBEATS model has been added

  • NBEATS (Boris O. et al. N-BEATS: Neural Basis Expansion Analysis For Interpretable Time Series Forecasting, ICLR (2020)) is a SOTA deep learning model introduced in 2020 and has beaten its competitors statistical models for the first time in history. Forecasting AI content users now can choose either NBEATS or Prophet model; NBEATS generally provides a better forecasting accuracy than Prophet especially when the training set is not enough. But, this benefit is at the cost of execution speed; NBEATS is a deep learning model, where as Prophet is not.

Hight-speed automatic HPO

  • The Baysian Optimization-based HPO process runs in multiple threads now and dramatically shorten the execution time

Compatibility: ALO 1.0.0