Hi , This is Harsha I am planning to do multivariate forecasting .I am having 60 variables out of which 20 variables are not known in future but in history I have everything and all are important .How to leverage finetuning of TimeGpt in such a scenario?

Hello! You can follow this tutorial on finetuning TimeGPT. In the

df

argument, make sure that your dataframe has all the columns of your exogenous features. For the features to be used, you have to pass their future value over the forecast horizon in the

X_df

argument (see an example here). For now, TimeGPT only supports features that are known both in the past and in the future. For your features with unknown future values, you will need to forecast them in order to use them (as shown here). I hope this helps!

If not with timegpt may I know any other models where I can train with all historical data and while doing inference only with known varaibles I should predict target varaible . Please let me know if any models you know

You can look at our complete list of models in neuralforecast that support exogenous features. You are looking for models that support historical (H - information only available in the past) and future (F - information known in both the past and the future) exogenous variables. So models like BiTCN, NHITS, NBEATSx, TSMixerx would be good options.