By In this study, we construct financial networks in which nodes are represented by assets and where edges are based on long-run correlations. We construct four networks (complete graph, a minimum spanning tree, a planar maximally filtered graph, and a threshold significance graph) and use three centrality measures (betweenness, eigenvalue centrality, and the expected force). To improve riskreturn characteristics of well-known return maximization and risk minimization benchmark portfolios, we propose simple adjustments to portfolio selection strategies that utilize centralization measures from financial networks. From a sample of 45 assets (stock market indices, bond and money market instruments, commodities, and foreign exchange rates) and from data for 1999 to 2015, we show that irrespective of the network and centrality employed, the proposed network-based asset allocation strategies improve key portfolio return characteristics in an out-of-sample framework, most notably, risk and left-tail risk-adjusted returns. Resolving portfolio model selection uncertainties further improves risk-return characteristics. Improvements made to portfolio strategies based on risk minimization are also robust to transaction costs.