14% returns. Best Sharpe in the sample. And it doesn’t predict returns.

This new paper shows why forecast accuracy is the wrong objective in portfolio construction.

Instead of minimizing MSE, it trains models to optimize the final portfolio decision itself — using the Smart Predict–then–Optimize (SPO) framework.

What happens?

• Small forecast errors stop blowing up portfolio weights,

• Trading costs and turnover are handled inside training,

• Simple linear models outperform deep nets,

• And robustness shines when it matters most.

During COVID:

• Sharpe > 2.1,

• Drawdown < 10%,

• While standard Predict-then-Optimize lost over 30%.

The lesson is brutal and clear:

In markets, decision quality beats prediction accuracy.