Tried your approach, it took me 4 days to re-create the experiment from scratch in JavaScript using FlashScore and Fifaindex and TensorflowJS.
The main challenge was to match players with 100% accuracy between the two data sources.

I’ve trained my model differently, I didn’t like your loss and optimizer.
The avarage ROI for 1000 runs of simulation on 100 randomly picked events from the Premier League with fixed bet size of 2% of bankroll is rather promising: 7.5%.

But it’s still inferior to a model I made from arbitrary rules for picking bets.