Is it not possible to backtest these sorts of things? If it were, I would assume results would be the first bullet point on articles like this, but I also don't understand why it wouldn't be possible.
Basically, the backtesting failed to match reality for the recent past, they tweaked parameters to get it to fit, and that produced surprisingly high predictions for the future. I don't think that would leave them anything to backtest on, since they'd used effectively all their historical data to fit the model to.