Suppose you have a general regression model, which behaves like a black-box to you. All you have is a $\ \ \text{train}(X,Y) \ \ $ function, where you input your predictor matrix $X \in \mathbb R^{N\times M}$ and target vector $Y \in \mathbb R^N$, as well as a $\ \ \text{predict}(x) \ \ $ function which takes a row-vector $x \in \mathbb R^{1\times M}$ and produces the prediction $y \in \mathbb R$.

Consider further the task of time-series prediction, i.e. there is a multivariate time series $\{\mathbf x_0, \mathbf x_1,\ldots\, \mathbf x_t\}$, where $\mathbf x_k \in \mathbb R^{1\times M}$, and you want to predict the outcome $y_{t+n} \in \mathbb R\ \ $(here $n>0$, and often $n=1$).

Two common methods to achieve this are

  • autoregressive approach: choose an order $q$ and take several samples $\big\{(\mathbf x_{t-q+1},\ldots, \mathbf x_t),y_{t+n} \big\}$ and feed it into your regression model.

  • autoregressive–moving-average approach: further add earlier prediction results, i.e. use data $\big\{(\mathbf x_{t-q+1},\ldots, \mathbf x_t),(y_{t-r},\ldots, y_{t+s}),y_{t+n} \big\}$, where $s<n$ (--not necessarily, but I would assume it to be reasonable).

Beside these obvious approaches, do you have other suggestions/references for using a general regression method for time-series prediction? Further comments on the above methods are also welcome.


Your Answer

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

Browse other questions tagged or ask your own question.