1
$\begingroup$

AR, MA, GARCH and VAR models are standard for autoregressive prediction, which is the forecasting of a variable using its own historical lags for features/regressors.

Support vector regression (SVR), random forests, boosting and ANNs are typical machine learning algorithms.

How good are these at time series forecasting? Are any of them able to forecast farther into the future beyond 1-day forecast horizons by using historical lags for input features? Do any sources give a theoretical basis for machine learners' applicability to autoregression? And to assess their out-of-sample performance, given that the task is autoregressive prediction of time series, is bootstrap aggregation or cross-validation more appropriate?

$\endgroup$
0
$\begingroup$

Raw regression is usually not as good performance-wise but there are variations of said algorithms to accommodate time series more effectively like Long Short Term Memory (LSTM) neural networks, Gated Recurrent Units Neural nets among others (GRU) as well as other non neural net approaches and closed to graph theory like Dynamic Bayesian Networks.

$\endgroup$

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