2
$\begingroup$

Lets say we want to predict a single target variable and we have 10 regressors/features. Assume we would like to predict 30 days ahead (daily predictions up to 30 days ahead) and our data is a daily time-series where we have NO future regressor values.

In this case, the only way we can use a discrete model such as ARIMAX is to lag the regressors by our forecast horizon (in our case 30 days) in order to be able to pass some regressor values for future predictions. However this means we are using outdated information to predict the future - instead of using yesterday's regressor information, we are using information from 30 days ago.

From my understanding continuous time models such as Gaussian Processes, Ornstein-Uhlenbeck (which is supposedly the continuous time version of an AR process), CARMA process etc, would allow us to use ALL the data that we have to predict n-steps ahead as they are continuous time models.

Is my logic sound or is there something I am missing here?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

I disagree that we are using outdated information to predict the future. At time $t$, if you want to predict $y_{t+30}$, you cannot use information from $t+1,t+2,\dots,t+29$ because you do not have it. You are using the freshest information available: that from time $t$. Therefore, I do not think that this is a problem. Consequently, I do not think it needs a solution in terms of a continuous time model. Even with a continuous time model, you will not have future data at time $t$, so you will have to keep the 30 day lag to be able to produce a forecast.

If you wish to predict not only 30 but 1, 2, ..., 30 steps ahead, you may create separate models for each forecast horizon. This way you may always use the most recent data.

Improve this answer
$\endgroup$
4
  • $\begingroup$ I understand your point, however if we assume that current time is $t$ and if our regressors are available up to $t$, then when we shift them by 30 lags, to predict tomorrow $t+1$, we are using regressor values up to 29 days ago and the other 29 days are reserved for predicting $t+2$ until $t+30$. This means that predicting tomorrow will not use regressor data from yesterday which is probably the most relevant one. Unless I am missing something here? $\endgroup$
    – MilTom
    Commented Feb 8 at 10:54
  • $\begingroup$ @MilTom, I did not understand that you want to predict 1, 2, ..., 30 days ahead. You wrote 30, so I understood it literally, or perhaps "in a narrow sense". You can build separate models for different prediction horizons. This is not uncommon. E.g. the model for predicting 1 step ahead could very well include the first lag (and more, if needed). $\endgroup$
    – Richard Hardy
    Commented Feb 8 at 11:11
  • $\begingroup$ My bad, I was not clear enough. Yes, I want to predict daily up to 30 days ahead. Different models for varying prediction horizon makes sense $\endgroup$
    – MilTom
    Commented Feb 8 at 11:17
  • $\begingroup$ @MilTom, there you go :) $\endgroup$
    – Richard Hardy
    Commented Feb 8 at 11:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.