Financial advisors focus on asset class as a way to help investors diversify their portfolios to maximize returns. Investing in several different asset classes ensures a certain amount of diversity in investment selections. Each asset class is expected to reflect different risk and return investment characteristics and perform differently in any given market environment. However, does the traditional classification really reflects the relationship between return and risks? In this report, we focus on The (Mis)Behavior of Hedge Fund Strategies: A NetworkBased Analysis provided by Eduard Baitinger and Thomas Maier (2020), which discuss a network-based methodology that models hedge fund strategies across the superordinate subordinate dimension to gain new insights into their interrelation. In addition, we make some extensions on their result.
There are three basic properties conveyed by the resulting financial network. First, the “backbone” of hierarchical clustering unveiling connections between assets across the superordinate–subordinate dimension. Second, the centrality properties of underlying assets that closely relate to the hierarchical “backbone” of the system. Central or strongly interconnected assets can be more strongly affected by shocks as shocks usually propagate faster to central assets and vice versa. Third, a financial network shows the clustering properties of underlying assets, that is, their proximity in the metric space. Information on the clustering behavior of assets can be useful for analyzing and improving portfolio diversification. Hence, we can use these properties to redefine the classification of underlying asset and try various strategies.
In this report, we will replicate the financial network of the paper first, then continue to extend two general research areas appear to be of particular interest mentioned in this paper, namely, network-based forecasting and network-based asset allocation.
For strategies to buy only peripheral nodes, Figure 1 shows that the long peripheral nodes strategy is obviously not better than equally weight, but not too bad either. On the other hand, we adopt the alpha strategy and found that alpha exists, while keeping only half of the nodes.
We try to use rolling window to backtest. From Figure 2 and Figure 3 we can see that regardless of the length of the rolling window, we can get positive alpha and sharpe.
Divide the nodes into 10 groups
In this part, will show the experiments of dividing the nodes into 10 groups base on network. Among them, the performance of group wei7 is particularly outstanding, which can be seen in Figure 4. We are also curious about the influence of grouping under different window lengths, so we decided to plot all the combinations and different universe. We find that the performance is still well.
In conclusion, we successfully verified the concept proposed by the author and found relevant supporting papers. The author leaves the topic of network application to the readers, so that we conducted a series of experiments based on the concept of network. Although we do not have complete SPY components for each period, we try to verify and assist from various angles, such as using equally weight and trying our best to restore the historical SPY components. The two main contributions of this study are to prune targets in the portfolio whilemaintaining similar or even better performance and to show the alpha strategies.