5. Now, I think it is going to be my last shot. I am going to create a lot of different features, including market profile-related features, moving averages, and whatnot. So far, managed to create 26 features, with minimum correlation so they provide unique information. Let's see how it works.

4. Therefore, what remains is to concatenate all other symbols and create a large input data. But the issue with the stock market is that different scrips and instruments follow different patterns, and that's why... you guessed it right, the model won't merge.

3. You have to resample it to at least a one-minute dataframe, where you sum up all the order book data of depth 5 so it correctly represents the underlying order book imbalance. But now, you're back to point 1; even though it's not daily data, your sample size is extremely reduced.

2. I started with the tick data, and that's why I selected Transformers. But here is the issue, Tick data is extremely noisy, and transformers always and almost try to overfit on the noise; your model won't merge.

Okay, so after working on transformers using time-series data, here are my inputs.
1. Transformers are extremely data-hungry. If you are working on daily data, then forget about them. And if you are working on tick data, then read along else skip below posts