We present a hierarchical three-phase optimization framework designed to automate the strategic intelligence of portfolio rebalancing across fragmented multi-account structures (e.g., Spousal RRSP, LIRA, and individual RRIF). This framework is specifically engineered to address the challenges of financial contagion and regime instability by optimizing for Regime Detection Intelligence on a 4-week “Strategic Pulse,” providing a decision-support path that filters local volatility noise.

The proposed architecture utilizes a nested “Protocol Stack” of evolutionary and deterministic layers:

Phase 1 (Objective-Specific Asset Selection): A Multi-Objective Genetic Algorithm ($NSGA$-$II$ via $pymoo$) solves the combinatorial problem of asset selection across multiple time horizons. Rather than a singular Pareto-front, this phase harvests elite sets for each objective separately—targeting safety-oriented vs. return-oriented goals across training sets with known historical behaviors. To ensure cross-asset comparability, a novel “Proxy EPS” normalization metric is used for Short and Trend-Following ETFs, while Actual EPS is utilized for traditional equities. An external 1D macro-signal modulates the relative weighting between the 4 quality and 4 momentum parameters during this selection process, ensuring the candidate sets are pre-conditioned for the observed macro-regime.

Phase 2 (Regime Intelligence Layer): A second $pymoo$-driven layer tunes 7 parameters governing a Markov chain that specifies transitions between discrete points in an 8-dimensional decision space. This phase operates exclusively on the elite specialist sets generated in Phase 1 as “fodder” for the optimization. Each point in the resulting trajectory defines the 8 parameter weights required for the downstream allocation, effectively automating the "Intelligence" of the regime pivot based on macro-congruence.

Phase 3 (High-Autonomy Execution Layer): The optimized 8D weights are passed to a Mixed-Integer Linear Program ($MILP$) which performs a simultaneous cross-account allocation. The $MILP$ solves the combinatorial challenge of distributing the global strategy across multiple legal entities and tax-sheltered buckets. It employs a two-step refinement to maximize the global look-back objective while minimizing transaction costs across all accounts simultaneously, constrained to not compromise the primary objective value by more than $x\%$.

Validation and Preliminary Results:
The framework was evaluated across multiple high-volatility training regimes (2007–2012, 2020, 2022, and 2020–2024). Preliminary results yield a terminal annualized return of 5\% for 2007–2012 and 24\% for 2020–2024. While currently in-sample, the consistency across disparate regimes—despite small degrees of freedom—validates the model’s structural adaptability and serves as a proof-of-concept for the final out-of-sample test on 2025–2026 data.