Please consider myself as newbie to data science, specifically to Time Series. As I can see in time series we are trying to predict the future based on past values. Isn't this problem can be formulated as any supervised machine learning problem with time also as one of the feature? Aren't they more accurate than timeseries models considering they can capture more than one predictors, apart from time? In real time data the time series data is always affected more than time itself isn't it? Is it still relevant?
$\begingroup$
$\endgroup$
1
-
$\begingroup$ Not if you are interested in causality or in inference. Recurrent neural networks is type of model which people appear to find "hot" $\endgroup$– RepmatJun 11, 2017 at 19:42
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Yes, you are right. Time series analysis for prediction is nothing else than supervised machine learning. I argue here why time series analysis is not commonly considered machine learning, but in my opinion, it really comes down to terminology.
Note that there is a lively research community looking at ML algorithms for time series forecasting, e.g., neural networks. However, just because a method "is machine learning" or sounds sexy doesn't mean it will necessarily outperform simple time series methods like exponential smoothing, ARIMA - or even a simple historical average: Is it unusual for the MEAN to outperform ARIMA?
answered Jun 11, 2017 at 20:51
-
$\begingroup$ Agree Stephan Sir. After working for couple of years I do realize sometimes Holt Winter itself can generalize pattern better! $\endgroup$ Aug 24, 2020 at 6:24