Investor-Tuned ML Portfolios with Turnover, Risk, and Arbitrage Difficulty Controls

PDF of paper available on www.linkedin.com/in/reggie-hyde

We develop a practical, multi-objective portfolio engine that converts machine-learning return forecasts into tradable positions under real-world frictions. Following the NN3 architecture of Gu, Kelly, and Xiu (2020), we produce monthly stock-level expected risk premia from a high-dimensional set of firm–macro interactions and industry dummies. These forecasts feed a convex optimizer that maximizes expected return while applying smooth penalties for three implementation levers: turnover, ex ante variance, and exposure to hard-to-arbitrage names. The problem is convex, and with valid inputs and a feasible constraint set standard solvers reliably return the global optimum. To capture implementation frictions, we introduce an arbitrage-difficulty score built from firm-level liquidity, trading-cost, and idiosyncratic-risk measures combined with market volatility and short-rate conditions. Penalty intensities are investor-tuned, allowing the same engine to align with the objectives of retail investors, pension funds, mutual funds, and hedge funds. We evaluate 29 penalty configurations across both calm and crisis periods, score portfolios on five practitioner metrics (realized return, Sharpe ratio, ex ante risk, turnover, arbitrage exposure), and select mandate-specific settings. Results show that penalty choices materially shape portfolio behavior, and they indicate a numerical “sweet spot” for risk-aware mandates. The framework offers a transparent bridge from ML signals to implementable portfolios and creates
a clean path to decision-aware learning, such as Smart Predict-then-Optimize, that aligns model training with downstream objectives