Hello beautiful community !
I’m starting a small project on time series and various elements of prediction for financial market.
I’m looking for help in some of the projects aspects and models development.
Moderated by Moderator
Hi @Igor_Lessio, The time series model can be made using different model architectures like using Dense layer, Recurrent Neural Networks(RNN), Convolution Neural Networks(CNN) etc. The choice of layers depends upon the time series data and the goals of the modeling task. RNN has the vanishing gradient problem due to this it is difficult for RNN to learn from distant past information in a sequence. To overcome this problem LSTMs can be used. Please refer to this time series tutorial where the time series was implemented using different model architecture to get some info. Thank You.
1 Like
Hi Igor.
Nice to meet you.
It would it be interesting if you could share details about the project you’re working on, and/or on the kind of project and expertise you are looking to help with.
In the meantime and building on Kiran answer, there are informally three popular families of techniques used for forecasting:
-
The first family of techniques are neural networks with native timeseries consumption such as RNN, CNN, or transformers. As with all neural net methods, those techniques require some significant expertise and amount of data to work, but they can have fantastic results. This is what Kiran mentioned. The Keras Time series lists some examples.
-
The second family of techniques are statistical models such as ARIMA or ETS. Those techniques are very easy to use, very popular, and work with very little data. However, their power of expression is limited making them possibly less accurate than other methods in case of large and complex data with rich patterns. You can learn more about them in the online book Forecasting: Principles and Practice and you can experiment with them in seconds using the Simple ML for Sheets addon. There are a lot of open source python libraries for that (e.g. prophet, tsfresh, sktime, darts, kats). R also has a rich eco system of such libraries. Some of them might benefit from your help.
-
The third family of techniques consist in transforming your temporal data (e.g. time series) into tabular data, and then to use a standard ML method (Random Forest, Gradient Boosted Trees, Neural Network, Nearest Neighbor, etc.). This approach is both easy to use and works both with small and large datasets, with both simple and complex patterns. From past experience, decision forest models work exceptionally well in this case–which is one of the reasons TF-DF was created :). The Temporian tool is specialized into converting temporal data to tabular data, and the tutorial section shows examples such as predicting frauds in sequence of transactions or predicting future sales with TF-DF.
1 Like
Tnks a lot Kiran !
Is the road i started in but as i just started i would like to have someone to help me in the beginning or in all the project. Of course with proper compensation.
Really tnks for the links Mathieu.
I am gonna check all the material. Temporian is really really cool.
I prepare the APIs then i try to convert the results and make a test.