Quantamental Portfolio Allocators: Deriving Alpha from Fundamental Metrics with ML

In recent work, Neural Portfolio Allocators: Cross-Asset Optimization Strategies with Attention-Based Transformers and Multi-Agents [13], we examined the limitations of classical portfolio optimization frameworks—such as Markowitz mean–variance optimization and risk parity—which rely on linearity, stationarity, gaussianity and stable correlation assumptions that rarely hold in real financial markets. That study established the advantages of incorporating neural networks and machine learning into multi-horizon portfolio management, extending their applicability beyond short-term algorithmic trading and into long-horizon allocation and risk control. A broad comparison across linear econometric models (ARIMA/VAR), stochastic processes (MGARCH/GBM), and neural architectures (RNNs/LSTMs) illustrated when and why model augmentation is optimal, even demonstrating that stochastic transformer-based architectures consistently outperform traditional approaches in long-horizon allocation tasks. These results spanned the full investment lifecycle, from data construction and universe selection to allocation, leverage control, and tail-risk management using extreme value theory. Building on those findings, this research project formalizes quantamental portfolio construction as a complete, end-to-end architecture that integrates classical financial metrics and fundamental analysis with academically robust, institutional-grade strategies employed in modern quantitative finance [see Visual Overview]. ‎(Quantamental Pipeline Visual Overview) Alongside our GitHub repository, the accompanying paper is intended as a pedagogical yet technically rigorous example of how fundamental analysis, machine learning, and modern risk management can be augmented into a single coherent investment pipeline via ML. It is written as an instructional reference encompassing equity research through allocation and optimization; Emphasizing transparency, reproducibility, and practitioner-centric investment strategies, rather than a journal submission. Content Outline & Paper Structure: Data, Screener, & Metrics Sleeve. Our report begins with defining our asset universe and data screening across large-cap U.S. equities, combining market data and standardized financial statements. A dedicated fundamental metrics sleeve is constructed using valuation, profitability, balance-sheet quality, and shareholder return measures. These features are winsorized, normalized, and aggregated into a composite cross-sectional Z-score that serves as the initial ranking mechanism. This stage produces an interpretable ordering of equities based on relative fundamental strength and accounting quality. Machine Learning & Probabilistic Rank. Building on the static screener, the paper introduces a machine learning refinement stage designed to estimate directional probabilities rather than point forecasts. Logistic regression is used as a baseline probabilistic classifier, with extensions including K-fold cross-validation, probability calibration, shallow neural networks, and LightGBM ensembles. Models are trained in a strictly rolling, out-of-sample cross-sectional framework. Performance is evaluated using ranking metrics such as Area under the curve, calibration loss, and realized hit rates. SHAP-based attribution is used to analyze feature importance and stability across time. The resulting output is a probabilistic rank that captures conditional likelihoods of future outperformance given evolving fundamentals. Utility-Based Allocation. These probabilistic signals are then translated into portfolio weights using a utility-based allocation framework grounded in Arrow-Pratt risk aversion and constant relative risk aversion preferences. Expected returns and covariance matrices are estimated from market data, and constrained mean-variance optimization is performed across a sweep of risk-aversion parameters. This produces an efficient frontier that illustrates how allocations evolve as risk tolerance changes, including transitions between concentrated growth exposure and diversified equity and credit allocations. Tail Risk Hedging & EVT. The final sections extend the framework to active long and short portfolios and a leveraged implementation. Short candidates are analyzed using empirical return distributions, ruin probabilities, and tail diagnostics. Extreme value theory is applied to volatility-normalized returns to model downside risk more realistically than Gaussian assumptions allow. The leveraged portfolio analysis incorporates total return swaps, gross and net exposure constraints, and drawdown behavior under stress scenarios. Pareto-type tail behavior and ruin probabilities are explicitly quantified to illustrate the trade-offs between expected return amplification and tail risk.