In South Africa, Regulation 28 (Reg 28) caps how much retirement funds can invest in certain asset classes, ensuring the diversification and protection of members' savings.

Reg 28 first came into effect in 1962 under South Africa's Pension Funds Act of 1956. It was designed to move the retirement industry away from the old prescribed assets system, where funds were mostly stuck in safe but low-return options like cash and government bonds, and into a smarter, more balanced approach to investing. It set firm limits on volatile investments, like capping stocks at 75% and property at 25%, pushing funds toward a practical, middle-of-the-road strategy. Over the years, it has been adapted to today's world: the big 2011 update emphasised trustees' responsibilities, and 2022 changes which increased offshore investments to 45% and introduced infrastructure.

But when every fund operates under the same rulebook, how do we know whether a fund is truly behaving as advertised? This article introduces a graph-theory-based diagnostic framework for testing fund consistency and uses the Reg 28 fund universe as a deliberate proof-of-concept to validate the method itself. The test consisted of a sample of 53 Reg 28-compliant funds, with 60-month return series ending October 2025. Fund names were anonymised. The network was constructed using Spearman ranked return correlations. Cluster stability assessed via rolling 3-, 6-, and 12-month windows.

Why test Reg 28 funds?

Reg 28 funds make an ideal test case precisely because the expected answer is already known. By regulation, all these funds must stay within the same asset class limits: equity, bonds, property, and offshore allocation, which means they are structurally constrained to behave similarly. Therefore, applying a reliable clustering method to this universe should produce broadly homogenous groupings, stable over time, with limited dramatic separation between clusters. If the method surfaced wildly distinct clusters or erratic groupings, that would be a red flag about the method and not the funds.

Validating this approach on a well-understood, regulation-constrained universe is important, as it can then be applied with greater confidence to less constrained fund sets where findings are less predictable and the diagnostic value is even higher.

Mapping funds as a network using Graph Theory

The graph theory approach treats each fund as a node in a network. Where two funds behave similarly through correlated returns or comparable asset allocations, an edge (connecting line) is drawn between them. Applied across an entire fund universe, this produces a web of relationships that groups funds with kindred characteristics, without imposing any pre-defined categories.

Two graph structures are used to progressively refine this network.

Minimum Spanning Tree (MST)

The MST connects all fund nodes using the minimum number of links necessary, no loops, no redundancy. It acts as a noise filter, retaining only the most significant pairwise relationships and surfacing natural clusters. Its limitation: by focussing solely on the strongest link between each fund, it discards secondary relationships that may still be meaningful.

Planar Maximally Filtered Graph (PMFG)

The PMFG starts where the MST ends. It restores important connections the tree had dropped, while keeping the graph drawable without crossing edges. In our study of 53 funds, the MST produced 52 edges; the PMFG produced 151, recovering triangular, multi-fund relationships that better reflect shared strategies and overlapping exposures.

Clustering is then performed on the PMFG using the Directed Bubble Hierarchical Tree algorithm. This is a method that identifies fund communities objectively from network topology, without requiring the analyst to pre-specify the number of clusters.

What the analysis found

Two dominant fund clusters emerged from the data, which broadly reflect mild strategic tilts (growth-oriented versus income-focussed) within the shared Reg 28 balanced mandate. However, several findings stand out:

• Low separation, by design. Low separation confirms the method works. Modularity was 0.087 and conductance 0.490, which indicates significant cluster overlap. This is the expected result for a regulation-constrained universe. Rather than a limitation, it is the validation: the method correctly detects the homogeneity that Reg 28 is designed to impose.
• Stable membership over time. Most funds maintained consistent cluster affiliation 60-70% of the time across rolling windows, rising further over 12-month rolling periods. This suggests the clustering is capturing genuine strategic orientation, not short-term market noise.
• Hybrid funds as bridges. A small number of funds sat between both clusters, exhibiting high centrality and betweenness. All were funds of funds: an unsurprising result, given their multi-strategy mandates and keeping in mind that child portfolios were also in the same sample. Bridge funds warrant close monitoring: a position between clusters can reflect deliberate diversification, an early sign of style drift or a source of systematic risk, possibly amplifying effects of shocks.
• Longer windows, clearer signals. Three-month cluster assignments were relatively volatile; 12-month windows revealed stable, persistent groupings. Fund "personalities" emerge over longer horizons: consistent with what one would expect in a constraint-bound, well-regulated environment of active funds.

Practical applications of this framework as a diagnostic tool

Beyond these findings, the framework has broad applicability as a governance, due diligence, and risk monitoring tool:

• Peer group discovery: Surfaces natural fund communities that may differ from marketing categories, enabling more meaningful benchmarking.
• Style drift detection: A fund shifting position within the network over successive windows is an early warning of mandate creep – visible long before it appears in headline statistics.
• Intent vs outcome validation: Checks whether a fund occupies the cluster its mandate implies, confirming that implementation aligns with stated objectives.
• Broad applicability: While focussed on Reg 28 funds, the framework transfers readily to quantitative funds, sector strategies, or any pool of comparable investment vehicles.
• Hybrid entities: Can act as a warning sign that exposes false diversification.

Conclusion

Are Reg 28 funds doing what they say they are? Based on the evidence here, largely yes. The fact that results align precisely with what Reg 28's design predicts confirms the framework is objective, unbiased, and fit for purpose. Applied to a fund universe with known, regulation-imposed characteristics, the framework returned exactly the results one would predict. That alignment between expected and observed outcomes is precisely what validates a diagnostic tool.

Reg 28 is the proving ground. The real opportunity lies in applying this validated framework to entity sets where the answers are not already known, and where the insights could be genuinely unexpected. For analysts, trustees, regulators and all stakeholders, network clustering offers a richer diagnostic lens, one that complements traditional performance metrics and catches what conventional reports miss. Having established its reliability on familiar terrain, it is now ready to be deployed where it matters most.