Joseph Iyanda, PhD, is an academic, researcher, and quality assurance professional at Botho University with interdisciplinary expertise in information technology, artificial intelligence,
business, and education. His research interests include Artificial Intelligence, Machine Learning, Data Science, Explainable AI, Feature Engineering, Deep Learning, Financial
Engineering, intelligent decision systems, and data-driven enterprise applications. His doctoral research investigated the effects of feature engineering on structured and unstructured
enterprise data for predictive modelling and machine learning. He has been a Reviewer for reputable publishers. He has authored and co-authored scholarly publications and book
chapters in digital learning, assessment transformation, intelligent systems, artificial intelligence, and emerging technologies. Dr Iyanda also contributes to academic quality
assurance and programme evaluation and has served as a subject matter expert in qualifications and accreditation-related activities. He is a Certified Artificial Intelligence Practitioner (CAIP),
and also member to the Blockchain Council.

Reliable AI-Assisted Portfolio Decision Systems in Finance: Constraint-Aware Validation, Covariance
Regularisation and Hierarchical Allocation

Financial AI/ML validation is typically viewed as a single end metric for measuring the improvement of
financial AI/ML systems. However, financial AI/ML systems consist of multiple layers including numerical
model input layer, algorithmic predictive reliability layer, decision feasibility layer, and ultimately the realized
portfolio layer. Therefore, validation of financial AI/ML systems should be treated as a cross-layer problem
rather than a single end-metric validation problem. This study develops and empirically tests a constraint-aware
validation architecture for financial AI/ML-supported portfolio decision systems. We use daily value-weighted
Fama–French 10-, 30-, and 48-industry portfolios covering the period from 2000 to 2025. A strict walk-forward
validation approach is adopted using a one-trading-day information lag, a 252-day estimation window, a
monthly rebalancing frequency, and proportional transaction costs. We compare the sample covariance matrix to
a structured covariance estimate via Ledoit–Wolf shrinkage and parameter-free constant-residual-eigenvalue
Marčenko–Pastur (MP-CREM) regularisation. We then propagate the resulting covariance estimates through a
set of financial decision architectures including global minimum-variance (GMV) portfolios, quadratic-Kelly
portfolios, and Hierarchical Risk Parity (HRP) portfolios. Equal-weight and inverse-volatility portfolios are used
as benchmarks. Our empirical results for the 48-industry universe indicate that the mean condition number of
the sample covariance matrix is 971.1. The corresponding numbers for Ledoit–Wolf shrinkage and MP-CREM
regularisation are 471.4 and 265.2, respectively. While the two regularised covariance estimators result in
improved condition numbers, only the Ledoit–Wolf shrinkage estimator results in improved out-of-sample
quasi-likelihood forecast scores with a mean benefit of 0.0949 over the sample covariance matrix. The
corresponding 21-day stationary-bootstrap 95% interval for the difference is (0.0730, 0.1218). However,
downstream performance improvement is highly architecture- and constraint-specific. For instance, a 20%
maximum-weight constraint materializes in terms of reduced concentration and lower annualised volatility for
the quadratic-Kelly portfolio architecture but has no effect on the decision-making of the HRP portfolios.
Moreover, a simple GMV–MP Sharpe reversal under cap does not survive family-wise confidence control nor
key specification checks. The complete system execution pipeline also captures a range of additional metadata
including data hashes, solver traces, fully deterministic system configuration, and environmental metadata, as
well as automated feasibility checks. The results highlight an important insight for financial AI/ML researchers
and practitioners, namely that component quality, predictive accuracy, decision quality, and system
implementation robustness all constitute distinct validation layers that must be tested jointly before announcing
system-level improvement.