A recurrent mistake in quantitative research is to choose a sophisticated model before proving that the problem requires one. LSTM, TCN and Transformer architectures are powerful, but they are not automatically better than a strong tabular baseline.
The reason to use a sequence model is temporal structure. Imagine a market approaching an important level. Aggressive flow increases, price response weakens, liquidity changes, the level is tested, then either accepted or rejected. The order and interaction of these events may contain information that a set of static features cannot represent well.
A useful research ladder therefore starts with a simple baseline. Next, add lagged features to a tabular model. Then compare that with a true sequence model. Finally, destroy the temporal order while preserving the same values. If the sequence model does not lose performance after shuffling the order, it probably did not learn meaningful temporal structure.
This is also why I prefer multi-stage market-state problems to direct next-return prediction. Questions such as approach → pressure → test → break/hold → continuation/recovery give the model a structure that corresponds to market behaviour.
Complexity should be earned. A Transformer is valuable when it extracts information that simpler models cannot — not when it merely reproduces the same signal with a larger training bill.