I decided to learn about such an area as the use of neural networks in econometrics, regardless of subsequent employment. One PhD explained to me that:
"In econometric research, the explainability of models is important; neural networks do not provide this. For time series, neural networks can be used, but only with a special architecture, for example, LSTM. For macroeconomic forecasting tasks, as a rule, neural networks are not used. ARIMA/SARIMA, VAR, ECM are used."
But on one forum they explained to me that
"A typical task in the field of time series analysis is to predict, from a sequence of previous values of a time series, the most likely next/future value. The Large Language Model (LLM), which underlies the same ChatGPT, predicts which word or phrase will be next in a sentence or phrase, i.e. in a sequence of words in natural language. The current ChatGPT is implemented using so-called transformers - neural networks, which after 2017 began to actively replace the older, but also neural network and also sequence-oriented LSTM (long short-term memory networks) architecture, and not only in text processing tasks, but also in other areas."
That is, the use of transformers in time series forecasting may seem promising? It seems that now this is a relatively young industry, still little studied?
Yes, Neural Networks can be used for almost anything. However, some types of neural networks are created for specific tasks. You might want to read these:
https://otexts.com/fpp2/nnetar.html
https://towardsdatascience.com/exploring-the-lstm-neural-network-model-for-time-series-8b7685aa8cf#:~:text=“The%20LSTM%20cell%20adds%20long,term%20trend%20in%20your%20data.
https://www.geeksforgeeks.org/time-series-forecasting-using-recurrent-neural-networks-rnn-in-tensorflow/
https://www.turing.com/kb/recurrent-neural-networks-and-lstm
So yes, for forecasting you can try to use DNN, RNN (LSTM is form of RNN also). Anyway, these articles should cover all you need to know about forcasting using ANN. But... ANN requires a lot of data to train, that is why ARIMA family models might work better often. But if you have a lot of data you can use almost any machine learning model for forecasting. Back in university days we tested almost all of them and LSTM and SARIMA gave very similar results with one beating another based on task, data and how far you want to predict. So yes, ANN might not always work better but in some cases it can outperform even best models.
Both perspectives are valid, in my opinion, and the choice of whether to use traditional econometric models or neural networks like LSTM or transformers in time series forecasting depends on the specific context and requirements of your project.
And according to the questions at the end of your text, yes, the use of transformers in time series forecasting appears promising. Transformers, known for their success in natural language processing and computer vision, offer advantages such as capturing long-range dependencies and parallel processing. These attributes make them powerful for understanding complex temporal patterns in time series data. This field is relatively young and evolving, leaving ample room for further research and innovation.
I recommend visiting his GitHub repositories, particularly the one with the following link:
https://github.com/jdb78/pytorch-forecasting
Yes!
In fact, Artificial Neural Network models have several points of contact with traditional econometric models.
When applying Artificial Neural Networks to the problem of forecasting economic time series, the choice of input neurons can be made using economic theory, to determine the relevant exogenous variables, and the procedures commonly used in time series analysis, such as autocorrelation and partial autocorrelation functions, tests for unit roots, graphical analysis, etc.
Furthermore, particular characteristics of a time series, such as seasonality, trend and cycle, can be learned by an Artificial Neural Network, thus enabling predictions to be made.
- Badges
- Science topic
- Similar topics
- Biological Science
- Histology
- Sectioning