Relevance
9/10
Highly relevant to quantitative trading. The paper directly addresses portfolio optimization with practical constraints (transaction costs, liquidity, market neutrality). The uncertainty-aware factor selection provides a principled method for dimensionality reduction in factor models. The ensemble approach with low-correlated signals is directly applicable to multi-strategy portfolio construction. The 25-year out-of-sample backtest (2000-2024) with realistic implementation constraints (top 300 weights, 10 bps costs, monthly rebalancing) makes findings immediately actionable. The Sharpe ratio of 2.20 for Ensemble (A) and market-neutral properties are highly attractive for systematic trading.
Implementation Complexity
7/10
Moderate-to-high complexity. Requires: (1) Training CAE models with specific architecture (hidden layers [32,16], 50 experts, L1 regularization), (2) Feature engineering for Q-Boost (37 features per factor including lags, moving averages, rolling statistics, z-scores, momentum, peer lags), (3) GPU-accelerated XGBoost with RAPIDS, (4) Chronos-Bolt inference pipeline, (5) Bootstrap resampling for IID-BS, (6) Adaptive κ selection via log-sum-exp optimization, (7) Tangency portfolio construction and projection to asset weights, (8) Ensemble construction with expanding-window covariance estimation. The multi-model pipeline and adaptive selection add significant engineering overhead.
Reproducibility
3/5
The paper provides detailed hyperparameters (Table 3), feature engineering specifications (Table 2), and model configurations. Uses publicly available data (2000 largest US stocks, 153 long-short portfolios from Jensen et al. 2023). However, no explicit GitHub repository is mentioned. The CAE architecture, Q-Boost feature set, and Chronos-Bolt configuration are well-documented. The adaptive κ selection procedure with specific λ=1, η=2, H=12 parameters is described. Reproducibility is moderate given the complexity of the multi-model pipeline and reliance on specific data sources.