What are the advantages (and disadvantages) of using machine learning techniques, e.g.
- Neural networks
- SVM
over the traditional times series analysis based techniques e.g.
- ARIMA models?
What are the advantages (and disadvantages) of using machine learning techniques, e.g.
over the traditional times series analysis based techniques e.g.
The advantage is that, with proper specification, you can fit other ranges of problems. If your true data-generating is really an ARIMA model or close to it, then ARIMA is the best model. If it's not, other approaches may work better.
If you choose to use other supervised learning methods for fitting a timeseries model (don't forget a simple baseline), the modelling part, where you decide on the features and crossvalidation strategy you choose, is still a lot of work and it will require careful consideration. With neural networks, there are a lot of architectures to choose from.
In my own subjective practice, I have found that ARIMA is actually not a very effective model. Your mileage may vary, especially if you are an expert in that field. In any case, with modelling as an ARIMA model or as a supervised learning approach, there is a lot of choices involved that can greatly improve or destroy your performance.
In general, I would say the ARIMA models are already suited for a timeseries model, so you get results quicker, while the general supervised learning methods are more general, and will require more work. They are also more flexible though. Timeseries are a very broad topic, and there is no one model that will give the best or even good predictions in any case. Do you have covariates, seasonality, stationarity? Here, as in any problem, the question is, what causes the variation in the numbers, and how can I model it?
A popular and apparently effective framework that you don't mention is that of the Bayesian (structural) models, see for example Prophet.