Thread: Emacsen: Studying, studying and studying

Many download GitHub repositories believing that they already have a money-making machine. Spoiler: it's not like that. The code serves as a base, but if you don't understand what you're doing, you're just running random lines. And so you don't win in this game.

I liked Quora�s article on the limitations of RNNs for price prediction. It reminds us that models have limits. Sometimes the best thing is to combine them with market logic or simpler strategies. Not everything is solved with deep learning.

I don't know if this happened to them, but I trained a model that worked perfectly in validation and then in real did not predict anything useful. The damn overoptimization... I'm starting to think that market data is too chaotic to predict them with realistic precision.

I think we should open a parallel thread to share network architectures that have worked well. Not only code, but concrete structures and configurations. There is a lot of smoke out there, but few people sharing things that really work.

I spent the last three months working with transformers applied to the market. Yes, like NLP. The structure works, but it costs a lot to train it. Is it worth it? It depends on your infrastructure. If you have a 3090 or something more powerful, give it. If not, do not try.

I ask a little humility from those who think they have the magic solution. If your predictions were so good, you would be on an island, not a forum. So we better share, learn, and leave the ego at the door.

I would like to see a serious benchmark comparing RNN, LSTM, GRU and traditional models like SARIMA and Prophet. But with real data, not with clean laboratory datasets. Does anyone know about such a publication? Otherwise, we should do it together.

I tried using RNN to predict directional changes, not exact prices. The model improved a lot because it didn�t have to be precise, just right direction. That little shift in focus improved the results as you don�t imagine. Sometimes the trick is what you ask the model, not the model itself.

If you are going to use histdata data, please respect the difference between GMT and server time. I have seen completely ruined models for not aligning timestamps well. The most important part of machine learning is data cleaning. Without that, everything else is castles in the air.

What if we stop playing data scientists and start earning money with a simple mobile media first? A lot of deep learning, but they don�t know what an EMA crossing is. So you can�t.