Thread: Emacsen: Studying, studying and studying

Prediction of time series and neural networks https://uhra.herts.ac.uk https://stats.stackexchange.com https://cs231n.stanford.edu https://staff.ustc.edu.cn Some derivative projects in GitHub: https://github.com/emacsenli Forex data download: https://www.histdata.com LSTM: https://www.quora.com RNN Implementation of RNN using Theano: http://christianherta.de Keras, Theano: https://news.ycombinator.com Convolutional Neuronal Networks (LeNet

Very good contribution. I downloaded almost all the papers and I�m reviewing them little by little. Stanford�s seemed especially clear to me to understand how LSTM models are trained in the financial context. I don�t know if anyone has already implemented anything with Theano or if they�re all using Keras. Personally, I lean towards Keras for its simplicity, although I feel Theano gives you more control.

It gives me the impression that many get excited about neural networks without understanding the basics of statistics. Not everything is solved with LSTM and RNN, gentlemen. If you don�t know what autocorrelation or seasonality is, you shouldn�t be training a model. You�re playing with fire, but you don�t know where the fire extinguisher is.

Thanks for the links. Histdata�s already knew him, but I hadn�t thought of using it with neural models. It may be a gold mine with the right preprocessing. I�m thinking of putting together a balanced dataset with different market sessions to see how models behave with volatility changes.

People underestimate what it takes to train a neural network well. It�s not to lower the GitHub code, give it run and ready. You have to understand what you�re training and why it fails when it fails. I�ve seen too many posts celebrating a 55% correct as if they invented the Holy Grail. Anyway, that�s the level.

Has anyone tried combining LSTM with technical indicators such as input? For example, using RSI, MACD and volume as inputs instead of raw prices. In theory, that should help the model capture more interpretable patterns, but I don�t know if it ends up oversetting.

I'm tired of seeing everyone repeating that LSTM is the best thing for time series without even comparing to well-configured ARIMA models. Try both with the same data and you'll see that sometimes the classic beats the "modern." But of course, that doesn't look so sexy on Twitter.

I have trained an RNN model with Theano and the truth, it does give you more flexibility, but the cost in time is brutal. Keras gives you decent results with much less configuration. I think the key is how thin you want to adjust everything. If you settle for an average model, Keras is enough for you. If you go for the surgical, you better go with Theano.

If you're using histdata data, make sure you clean them well. They have holes and duplicates if you don't filter them. It happened to me once and my model learned anything. Those little details kill the prediction. And then one wonders why the model "doesn't learn".

I�ve been stuck with evolutionary networks applied to candlestick patterns for months. I turned the graphics to images and passed them through LeNet. The results are interesting, although they�re not yet consistent. I share it in case anyone else wants to explore that approach. I think there might be something useful there, even if it�s not mainstream.