[Feat] TimeGPT Forecasting

[Yibei990826]

Description

These are feature requests from our users, compiled through individual meetings, with the numbers indicating how many times each feature was mentioned:

Forecast Model

• Improve performance in time series with clear cyclicity or seasonality, compared against models like LSTM and XGB (4).
• Multivariate time series modeling (3).
• Enhanced accuracy for industry-specific datasets (2).
• Improved handling of nonlinear forecasts.(1)

Exogenous Variables

• Capability to input only historical exogenous variables or forecast them based on text, rules, or specific indications (3)
• Automatic selection of relevant exogenous variables (2).
• Enhanced interpretability with feature importance insights (2).
• Support for exogenous variables with different frequencies (e.g., hourly, daily, weekly) (1).

Finetune

• Greater control over finetuning options, allowing deeper customization based on specific rules (6).
• Ability to persist or reset the model post-finetuning (4).
• Option to initiate finetuning through text prompts (1).
• Indicator to show performance improvement after finetuning (1).
• Control over how many layers to finetune (1).
• Automatic selection of the best parameters for the task (1).
• Lower data requirements for minimum data points needed (1).
• Assistance with choosing the optimal loss function (1).

Use case

No response