Dynamic portfolios via triangulated maximally filtered graph networks: a centrality-driven multi-strategy approach

Abstract

This paper develops an empirical approach to evaluate network-based portfolio strategies relative to traditional benchmarks in the Casablanca Stock Exchange. Using the Moroccan All Shares Index, daily closing prices are used to compute historical log-returns. A rolling-window backtesting procedure over 2013–2022 ensures robustness across market regimes. Time-varying dependencies are estimated with the dynamic conditional correlation generalized autoregressive conditional heteroskedasticity (GARCH) model, capturing evolving correlations. These correlation matrices are filtered through the triangulated maximally filtered graph (TMFG) and minimum spanning tree (MST) to construct financial networks. Centrality measures identify systemically important and peripheral stocks, forming nine portfolio strategies, including TMFG- and MST-based selections, ESG-integrated portfolios and classical benchmarks such as minimum-variance (MVP) and equal-weighted portfolios. Performance is evaluated using Sharpe ratio, Sortino ratio, maximum drawdown and cumulative returns. Results indicate that the MVP provides downside protection but underperforms in bullish markets, whereas network-based strategies, particularly the TMFG selection portfolio, show consistent outperformance. ESG portfolios yield solid returns but with higher volatility, highlighting the need for risk control. Integrating network topology with dynamic correlation estimation enhances portfolio resilience in emerging markets, offering practical insights for investors and regulators.