A Methodological Framework for Intrinsic Explainability in Portfolio Allocation: Constraint-Aware Portfolio Reasoning Network

Abstract

Black-box portfolio models can produce allocation weights without a reconstructable account of how market signals, constraints, and risk controls shaped the decision. This study introduces the Constraint-Aware Portfolio Reasoning Network (CAPRN), a neuro-symbolic-inspired framework for intrinsic, decision-level explainability in portfolio allocation. Its implemented configuration uses a one-layer long short-term memory encoder and exposes factor relevance, constraint pressure, temporal state, rule-bias effects, and asset-level preference scores before producing long-only, fully invested weights. Conditional value-at-risk, a shuffled mini-batch wealth-path surrogate, and an equal-weight-deviation regularizer connect these variables to risk controls, while chronological drawdown and realized turnover are evaluated separately out of sample. CAPRN is evaluated on a ten-asset universe using strict walk-forward testing, performance measures, ablations, decision narratives, deletion and insertion diagnostics, counterfactual constraint tests, and explanation-quality metrics. CAPRN remains economically viable out of sample but neither uniformly outperforms equal-weight and mean–variance benchmarks nor exhibits statistically significant return or Sharpe-ratio dominance. Its internal variables are inspectable and stress-testable, although its factor-indicator fidelity is weaker than the marginal attribution performance of SHAP and LIME. CAPRN should therefore be viewed as an auditable allocation and governance layer rather than a benchmark-dominant production optimizer. Its principal contribution is a reproducible reasoning pathway connecting market information, constraint responses, risk controls, and final portfolio weights for practitioner oversight and regulatory reporting.