This message was deleted.

Hi Kai, I'll do my best to answer your questions the way I understand them, so let me know if I missed something. First, all models used lagged data of the target variable. You pass a window of data to predict over a certain horizon. For example, input the last 24 hours of data to predict the next 6 hours of data. If by lagged data, you mean other exogenous variables, like another time series, or calendar events, like holidays, then some models in neuralforecast also support them. The advantage of using a feature like holidays, is that you know when they will happen in the future, so it can help make better predictions (if your series is influence by holidays of course). You can also pass dynamic exogenous variables to some neuralforecast models. The tricky part, like you say, is that you have to forecast them separately to inform future predictions, so maybe prediction errors will be magnified. As for your leakage question, as long as you only use past information to predict the future, then there is no leakage. That's why you must keep your data ordered in time. I hope I understood your questions correctly and that it helps!

Hi Marco, thank you so much for your comprehensive and prompt response. I apologize I didn’t make it clear. I understand the concept of autoregression (if i understand this correctly) in NeuralForecast, which uses a window of historical/lagged (input_size) data to forecast future steps (horizon). However, I’ve noticed that someone created an additional feature which is the lagged data of the target variable itself. For instance, I’ve seen someone use the past 168 hours of electricity prices to forecast the next 72-hour prices, while also creating a 24-hour lag of the price as a covariate. Does the use of a 24-hour lag in this context make sense to you and if NeuralForecast supports this kind of 24-hour lag? And would this lead to future information leakage in cross validation? I’m keen to understand this better. Thank you once again for your time and insights!

I don't see any problem in doing that, and there should be no leakage 🙂

Great! Thank you so much! Have a good day!