Suppose that we are interested in estimating continuous-valued targets $y_t$ from continuous-valued observations $x_t$ over discrete time steps $t = \{1,2,3,\dots,T\}$. Could you give me some examples on real-world applications that during training we have data on both $x_t$ and $y_t$ and at test time we want to estimate $y_t$ given $x_t$?

There are applications where $y_t$ is always hidden where we can use algorithms such as expectation maximization (EM) to train the model. Also, there are applications for time-series prediction using RNNs such as the one presented here. But I'm looking for applications related to real-time estimation instead of prediction.


I think what you are referring to is called Nowcasting - i.e. like forecasting but for variables that are happening now, not in the future.

The textbook example of this is large-scale economic variables like GDP: There is a present value $y_t$ for the GDP, but it will take a long time to gather all the information required to calculate $y_t$. So instead we use a model to estimate $\hat{y}_t$ based on present data $x_t$ and historical values of $\pmb{y}$ ($y_{t-1}$, $y_{t-2}$, $y_{t-3}$, etc...).

You can simply use time series models with external regressors (for example ARIMAX, or RNN/LSTM if your data is rich enough and complex enough to warrant a deep learning approach).

A popular nowcasting approach from the last few years has been Bayesian Structural Time Series models, see this paper and the corresponding BSTS package in R.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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