Contemporaneous correlations are important for portfolio optimization problems. We propose a newly developed machine learning tool, the OWL shrinkage method, which explicitly exploits stocks’ contemporaneous correlations by assigning similar positions to correlated stocks (the grouping property). We find strong evidence that OWL-based portfolio strategies outperform other benchmark strategies in the literature when stocks exhibit strong correlations. In particular, the OWL shrinkage method bridges the gap between the naive (but well performing) 1 / N portfolio strategy and the portfolio optimization framework: our OWL-based portfolio strategies yield very similar portfolio weights to (yet not the same as) the 1 / N portfolio strategy, but outperform the 1 / N portfolio strategy in terms of both the Sharpe ratio and turnovers. We also show that the superior performance in Sharpe ratio against the 1 / N portfolio is significant. • Exploit stock correlations using newly developed machine learning techniques. • Incorporate various portfolio weight-constraints in the optimization framework. • Our methods yield favorable out-of-sample results. • The superior performance against the equal weighted portfolio is significant. • Our methods produce minimal transaction cost compared to other benchmarks.