kwuking
commented
3 months ago According to NeuralForecast,
Exogenous variables can provide additional information to greatly improve forecasting accuracy. Some examples include price or future promotions variables for demand forecasting, and weather data for electricity load forecast.When I tried to use Time-LLM with the NeuralForecast implementation, I encountered the following exception:
Exception: TimeLLM does not support historical exogenous variables.Does this mean that Time-LLM does not use external variables for forecasting?
I also noticed that Time-LLM has been evaluated using datasets like ETT, Weather, and ECL, all of which contain multiple external variables to assist with forecasting. Is the issue due to the NeuralForecast implementation not supporting these variables, or does Time-LLM inherently not use and benefit from external variables? Thanks!
Thank you for your interest. Currently, Time-LLM uses a channel-independent method for processing (a method that has been widely adopted since PatchTST), which transforms multivariate time series into univariate time series. Therefore, for handling multivariate time series forecasting and covariate prediction, I recommend looking into my iTransformer, which focuses on joint prediction for multivariate time series.
According to NeuralForecast,
When I tried to use Time-LLM with the NeuralForecast implementation, I encountered the following exception:
Does this mean that Time-LLM does not use external variables for forecasting?
I also noticed that Time-LLM has been evaluated using datasets like ETT, Weather, and ECL, all of which contain multiple external variables to assist with forecasting. Is the issue due to the NeuralForecast implementation not supporting these variables, or does Time-LLM inherently not use and benefit from external variables? Thanks!