The writing does not exclude proper backtesting. A more favorable reading would have him/her use a holdout or test set that is in the future of the historical data used for training. Then there would be no problem with that approach in regards to overfit/poor generalization.
There is still the problem of the market not reacting to his trades. It would be very easy for me to show a trade of $1m making money, but very hard to execute such a trade.
