When Structure Matters: Do Network-Based Portfolio Strategies Outperform Conventional Models?

Abstract

This thesis investigates whether network-based portfolio allocation models outperform conventional strategies in terms of risk-adjusted performance and robustness in structurally diverse return environments. Motivation arises from the limitations of traditional allocation methods in capturing interdependencies among assets or factors. We implement five strategies, including Equal Weighting (EW), Inverse-Variance Weighted (IVW), Minimum Variance Portfolio (MVP), and two network-based approaches, namely the Centrality-based Correlation Network (CCN) and the Centrality-based Directed Network (CDN). In the empirical analysis, CDN outperforms traditional benchmarks within asset-based universes by achieving higher Sharpe ratios, lower portfolio concentration, and greater allocation stability. In the factor-based setting, the CCN delivers the highest Sharpe ratio, although it results in higher turnover and concentration. In contrast, when asset and factor return structures overlap, both network-based models underperform simpler heuristic benchmarks such as EW and IVW. Bootstrap analysis shows that, despite favorable average Sharpe ratios in some regimes, none of the strategies achieves statistically significant outperformance. This holds both in absolute terms and when compared directly to the Equal Weighting benchmark. We construct a simulation study that spans three stylized regimes. CDN performs well in the directed regime by taking advantage of the directional influence to reduce systemic exposure. CCN, in contrast, does not capture directional dynamics and, therefore, offers no advantage in this setting. In the dense symmetric and hybrid regimes, structural clarity fades, and all models, including the network-based ones, tend to converge toward naive allocations. We conclude that network-based portfolio strategies can provide superior performance and allocation stability compared to traditional methods, but only when return dependencies are clearly defined and persist over time. In settings where structural patterns are weak, noisy, or ambiguous, conventional heuristics such as EW and IVW remain more reliable and robust choices.