US ML-Enhanced Portfolio Optimization

This report implements the AI-Driven Quantitative Portfolio Optimization Engine (MLPO) on US large-cap equities: a 30-name diversified panel across technology, healthcare, finance, consumer, energy, industrial, and utilities sectors, benchmarked against SPY.

Out-of-sample evaluation uses quarterly walk-forward rebalancing with a rolling 36-month estimation window, training on history from late 2013 through November 2020, then simulating through the latest available prices. A 4% annualized risk-free rate proxies US cash/T-bill returns. Four constructions are tracked out-of-sample: mean-variance (MV) optimized, ML-enhanced MV (XGBoost views via Black-Litterman), a cap-weight reference, and SPY.

Interactive exhibits quantify return and risk (CAGR/total return, Sharpe/Sortino, max drawdown, beta/alpha, information ratio, and VaR/CVaR) plus SPY-relative behaviour (tracking error, correlation, up/down capture, and turnover), along with allocation and model diagnostics (sector concentration and ranked forecast signals).

Results are research illustrations of method behaviour—not investment advice or live trading signals.

Stage 1 — Universe and prices. Thirty liquid US tickers with aligned adjusted closes from Yahoo Finance; names with insufficient history are dropped before estimation.

Stage 2 — Feature engineering. Technical indicators (moving averages, RSI, MACD, momentum, volatility) feed the gradient-boosted forecasting model.

Stage 3 — Training window. At each quarterly rebalance, the prior 36 months form the estimation sample for means, covariances, and ML features.

Stage 4 — Optimization. Pure MV maximizes Sharpe under long-only bounds and a volatility ceiling; ML + MV replaces expected returns with Black-Litterman posteriors from ML views.

Stage 5 — Walk-forward simulation. Weights are held for one quarter; 0.1% slippage applies to turnover at each rebalance.

Stage 6 — Risk analytics. Cumulative paths, drawdowns, Sharpe/Sortino, beta/alpha versus SPY, VaR/CVaR, and sector concentration are reported below.

Mean-variance optimization. Weights maximize Sharpe subject to ${\sum }_i{w}_i=1$, box constraints, and $\sqrt{252w^{\top }\Sigma w}\le {\sigma }_{\max }$ with ${\sigma }_{\max }=22.5\%$ annualized.

Machine-learning views. XGBoost predicts forward returns from technical features; out-of-sample $R^2$ maps to view confidence, with forecasts shrunk toward historical means before Black-Litterman blending ($\tau =0.025$).

Black-Litterman. Cap-weight notionals imply equilibrium returns; investor views $Q$ with diagonal $\Omega$ produce posterior expected returns for the second MV solve.

Performance metrics. CAGR, Sharpe, Sortino, max drawdown, beta, alpha, VaR, and CVaR follow standard definitions on daily portfolio returns versus SPY.

Compare ML + MV to MV only to see whether ML views improve the risk-return trade-off; compare both to SPY for passive benchmark context.

Sector allocation bars show industry concentration in the latest optimized book. Signal tables rank cross-sectional forecasts from the most recent training window (annualized, model-implied)—illustrative, not orders. Use the risk and relative analytics above to judge whether any outperformance (if observed) comes with better downside control (drawdowns + VaR/CVaR) and how the portfolio participates in market up- and down-moves (capture metrics).

Yahoo Finance prices and simplified transaction costs (slippage on turnover only) may differ from live execution.

The fixed 30-stock panel is not the full S&P 500; results may not generalize to broader universes.

ML forecasts are noisy; views are shrunk toward historical means by design.

Past backtest performance does not guarantee future results.

Interactive charts below cover the full scorecard: strategy performance (CAGR/total return, Sharpe/Sortino, max drawdown, volatility), relative analytics vs SPY (tracking error, correlation, up/down capture, and turnover), and risk profile (beta/alpha, information ratio, VaR/CVaR).

You also get behaviour over time (calendar-year returns, drawdown and cumulative wealth paths, monthly returns), plus allocation and model outputs (sector concentration and ranked cross-sectional signals). Regenerate with npm run data:us-mlpo after refreshing market data.