Relevance
9/10
Directly addresses portfolio construction as a decision problem rather than a prediction problem. Bridges the gap between gradient-boosted trees (the workhorse of tabular ML in quantitative practice) and end-to-end decision-focused learning. Provides exact Kelly criterion optimization within a tree-boosting framework. Demonstrates that the choice of training objective (growth vs. surrogate) matters more than the choice of learner class. Includes practical deployment recipe (leave-one-out ensembling), transaction cost analysis, and honest boundary conditions. Highly relevant for practitioners wanting to use boosted trees for allocation without routing through surrogates.
Implementation Complexity
7/10
Requires understanding of XGBoost internals (multi-output mode, vector-leaf trees, custom objectives), second-order optimization theory, Kelly criterion, and careful evaluation protocol design. The core objective and derivatives are elegant (closed-form gradient and diagonal Hessian), but the full pipeline includes: feature engineering (7,871 candidates), stochastic local search over (hyperparameter, feature list) states, purged walk-forward selection, deployment walk-forward with expanding windows, leave-one-out ensembling, and transaction cost modeling. The dependency-free numpy reference implementation is ~200 lines. Requires xgboost >= 3.4 for correct min_child_weight semantics under vector leaves.
Reproducibility
5/5
Exceptional reproducibility: code, data, and scripts regenerate every number from one committed CSV without an API key. Includes a dependency-free pure-numpy reference implementation (~200 lines) tested to 10^-10 against brute-force evaluations. All derivatives verified by finite differences. Fixed seeds for every fit. Full search logs committed. One command (uv sync && uv run pytest && bash run_all.sh) regenerates all tables and figures. Raw data snapshot (38 series, 1985-2026) committed to repository.