Recurrent Neural Networks (especially LSTM), may work better than other, more traditional methods.

First of all, thanks to universal approximation theorem, a neural network can approximate any given algorithm, given sufficient computational power and a correct architecture.

Then, a neural network is fully modular. With ARIMA/Autoregressive Models, you're forecasting "only" on prior events. With a NN, you can integrate any other informations.

For example, I've worked on stock predictions, so instead of focusing only on the past, I added other relevant informations like stocks of related companies, Google Trends statistics on specific keywords, specific trading indicators, etc...

But the two models aren't mutually exclusive, I've seen works for TS forecasting with ARIMA as part of the input of a NN.

BUT, there are some drawbacks: LSTM may be computationally intensive, more than simpler models. Should you use DL if the training is taking 2 hours, just for 0.05 more accuracy? Also, NN are of course prone to overfitting. You may have to deal with data quality/quantity, and the usage of dropouts, L1 or L2 regularizers.

Each technique is having its own pros and cons,so try both and see which one fulfill your goals, and needs!

Clément Poiret Great description but I do not think the computational drawback is a real drawback for time series forecasting. Most time series problems do not deal with millions of data points. The computational drawback would arise if you work with thousands of images for instance.

So back to the question: ARIMA or neural nets ?

LSTM shows great performance but I basically would start with something simple like LR or ARIMA, get a first impression of the system and its features and go on with more complicated methods.

If you already have a clean dataset, it is not a big deal to test several methods.

I am agree with your statement., however for better accuracy estimates you may need to use Deep Learning via a Bayesian non parametric/or parametric modelling framework.

I am not sure. I do not know your system and aim ...so it is hard to say.

I really like Bayesian approaches but not for all problems.

...Basically a neural net does not give you details what happened and why you get a value of xy as forecast. A neural net learns trends and patterns in your time series data set and makes a forecast based on this experience. Another approaches gives you more insights. They uncover relations between features and output which is really important to understand your system.

Nonetheless, I know that neural nets perform really well ..especially on non-linear problems but it is a kind of unsatisfactory to get just a result without an explanation :-)

You need to test the algorithms first, then form the right conclusion.

HTH.

Dr. Samer Sarsam

Yes, machine learning approaches generally will give you better results compared to well-known parametric approaches (such as ARIMA or exponential smoothing). This happens because parametric approaches use very strict assumptions about the data generation process. Machine learning techniques are more flexible.

BUT: in order to use machine learning you need to have a sufficiently large set of data. In cases of hundreds of observations the traditional approaches usually work better.

But in order to find out which method is better for your data you need to perform empirical out-of-sample rolling-origin evaluation.

Let me recommend this chapter containing step-by-step instructions for rolling-origin out-of-sample evaluation of forecasting performance in practical settings:

Chapter Forecast Error Measures: Critical Review and Practical Recommendations

There are few studies in this area. A significant number of the NN research in forecasting and prediction suffer from validation and implementation problems. This render their results suspect (see ADYA and Collory, 1998).

Hai Alket Cecaj ,

Time series data sets are quiet interesting to analyse with time stamp across several period.

Recently i worked with time series data using PAST and MATLAB APP called time series NN forecasting. Results are impressive with various performance graphs as output.

According to my experience, when volatility increases ann perform is getting poorer. Look at my study Chapter THE ANALYSIS OF CRYPTO CURRENCY PRICE BEHAVIOR WITH ARTIFICI...