Twenty years of daily data on 500 stocks is about 2.5 million rows, which sounds large until you remember the rows are highly correlated cross-sectionally and serially, and that the true independent sample is closer to the number of distinct market episodes, perhaps a few hundred.
Where neural networks genuinely earn their place in finance is on data that is actually large and information-rich: limit order book sequences, transcripts and filings, satellite or transaction data. For predicting tomorrow's return from twelve technical indicators, they add flexibility no one needed.
If you do use one, treat the architecture, width, depth, and training schedule as hyperparameter choices subject to the same multiple-testing discipline as everything else, and be sceptical of any result you cannot reproduce on a different period.
Related: bias-variance-tradeoff, hyperparameter, overfitting, sample-size