Actually, the “likely voter” model has proven wrong over the last few cycles.
Ignore at your peril.
A likely voter model without certain balances and you’d be correct. If you can produce a registered voters poll that is remotely near the mark of accuracy, I’d love to see it.