First, when using pure real-time macro information alone, the authors find that deep learning cannot help find any statistical evidence for forecasting both non-overlapping and overlapping excess bond returns. In contrast, some machine learning models can help find some statistical evidence for forecasting overlapping excess bond returns.
Second, when using both pure real-time macro information and yield curve information, they find that deep learning performs well for forecasting medium- and long-maturity overlapping excess bond returns, but such predictability is dominantly driven by yield curve information.
Third, all statistical evidence of predictability is much weaker than that found from using fully-revised macro data and generates minimal economic gains for a mean-variance investor, regardless of her level of risk aversion and whether she can take short positions.