Dear Yair. Those people are quite on the money! Suppose you got a multivariate VAR model and you want to predict just one of the variables therein. Well, the VAR model cannot predict just one, but the whole set of variables in that model. Therefore, it is easy to forecast your target variable, but it is actually predicted together with all those predictors. If at least one of those predictors is not well predicted, then all the other will not either. And this example can be modified for any other models you can use, but the message will be the same: you cannot predict the future with precision with historical data (not even with univariate models).

When using NeuralProphet for multivariate time series forecasting, the decision to include future regressors depends on the specific model configuration and the requirements of your forecasting task. You can incorporate historical regressors (past data) and future regressors (future data) to predict the target variable. Regarding your question about predicting the period from Jan 2024 to Dec 2024, NeuralProphet can generate forecasts based solely on historical data. It utilizes advanced neural network architectures to capture complex temporal patterns from historical data and project them into the future. Therefore, it is not strictly necessary to have future indicators (regressors) to predict the target products. However, including future regressors can improve the accuracy of the forecasts, particularly when external factors or drivers impact the target variable. In practical terms, obtaining future data for regressors can pose a challenge. One approach is to use forecasted values for the regressors obtained from other forecasting models.

OK. I agree with Hassan. I was actually referring to the goal of forecasting with precision. Because in practice, you can forecast everything almost with nothing. BUT to predict just one variable with good precision, that's quite another matter. And with neural networks, you must consider the associated computational costs together with the precision benefits. Finally, let's consider the "efforts" made in the literature. There are quite a lot of papers offering new advanced tools or trying to predict specific variables in many fields, but their efforts to compare their proposals with respectable alternatives are usually weak.