What Bandy is saying is surprising because CTAs and especially Turtles made billions trading non stationary prices with trend-following systems and in fact it was the non stationary prices (trends) that provided the profits.
He points out that the Turtle method worked in an environment of higher inefficiencies in the market. Today's markets are are much more efficient. He actually says the Turtle method will not work today.
He is also mostly talking about machine learning algos. They DO need stationary data to work. So, as Val said, we need to break down the data into time splices that will be more stationary than long time slices (like years).