Thank you for sharing such a wonderful code and idea, my main question is how you ensure you do not use future data during testing, since your transformer has no postional encoding. Of course, this is the concern of your CNN architecture. Looking forward to your reply!!!

[hnczj]

Thank you for sharing such a wonderful code and idea!!!

[langgege-cqu]

Firstly, the FI-2010 dataset is publicly available and valuable for LOB data. If you have research on mid-price trend prediction, you can run down the models I have published (deeplob, deepfilio, ocet), and you should know that I have not used future data. Secondly, I believe that the CNN part of OCET is equivalent to performing some positional encoding, as can be seen in the ablation experiment of OCET. I think the feature of the LOB mid-price dataset are too small, only a price quantity of 10 levels from the past. I think the important role of Transformer in it is feature aggregation rather than feature extraction. I hope it will be helpful to you.

[hnczj]

Thank you for the clarification and additional insights on using transformers for LOB data. I really appreciate you taking the time to explain the nuances of your architecture decisions. The points about CNN providing an equivalence to positional encoding and transformers mainly serving aggregation with small LOB features make complete sense.

I'm looking forward to digging deeper into the implementations you've generously open sourced for deeplob, deepfilio, and ocet.

Thank you again for taking the time to clarify and teach - it's extremely valuable and sets a positive example I aspire to follow.

[langgege-cqu]

Nice going!