Launching QAML: A Transparent, Multi-Layer Stock Selection Framework Delivering Robust Risk-Adjusted Returns Across US, Europe, and Japan

Core Conclusion

QAML (Quant + Analyst + Machine Learning) is a fully transparent, cascading stock selection framework that combines five core factor families (Value, Quality, Momentum, Growth, Low Risk), a dynamic metric pool derived from sector analyst insights, and an AdaBoost machine learning layer incorporating over 100 technical indicators. Since 1996, the Enhanced QAML Composite – which adds Quality-Value and Momentum-Revisions interaction filters – has produced post-cost Sharpe ratios of 0.70 (US), 1.49 (Europe), and 1.20 (Japan). Over the last five years, the framework maintained or improved these figures: US 1.25, Europe 1.31, Japan 0.89. The key investment implication: QAML provides a scalable, transparent alternative to black-box quant strategies, but its efficacy is contingent on maintaining beta-neutral implementation, controlling high turnover costs, and adapting regional factor weights.

Market Mispricing – Three Key Underappreciated Elements

1. Analyst-driven dynamic metric pool enhances recent performance – In the past five years, the dynamic pool’s Sharpe ratios consistently surpassed those of the static core metrics across all three regions. The additional analyst-suggested metrics (e.g., 26 growth indicators in the US) delivered positive Sharpe ratios across all samples, with the last five years showing significantly stronger results than the full period. This suggests the market undervalues the timeliness of analyst-informed inputs for capturing evolving factor leadership.

2. Quality-Value and Momentum-Revision filters provide material risk-adjusted gains – Over the last five years, applying the two interaction filters boosted the Composite’s Sharpe ratio by +0.17 in the US and +0.23 in Japan. The improvement is predominantly driven by the long side: removing stocks that rank in the bottom 10% on both Quality and Value (or that have strong residual momentum but weak earnings revisions) eliminates high-risk, low-return positions. Europe shows only +0.02 improvement, indicating regional specificity.

3. Low overfitting risk due to full transparency and regime adaptability – Unlike many machine learning quant models, QAML’s cascading architecture (core → dynamic pool → ML layer) is fully traceable. The ML sub-composite exhibits low correlation with six traditional factor returns (monthly correlations typically <0.3), providing genuinely orthogonal signals. Sensitivity tests show performance is stable across reasonable parameter ranges (learning rates, training rounds), and regime-based training (using OECD CLI, inflation, yield curve states) further reduces overfitting. The market may underprice the framework’s ability to adjust to different market regimes, especially VIX regimes: trend-following indicators perform best in falling VIX, while mean-reversion signals excel in rising VIX.

Evidence Chain

Core Metrics Performance (pre-cost, since 1996):

• US: Quality metrics strongest (low accruals Sharpe ~0.5, ROIC ~0.4); Value (FCF yield ~0.35) and Momentum (residual momentum ~0.4) also contribute; Growth and Low Risk have near-zero or negative Sharpe.
• Europe: Value dominates (FCF yield Sharpe ~0.6); Quality and Momentum are positive but weaker; Growth ~0.3; Low Risk near zero.
• Japan: Value is the primary driver (all three value factors Sharpe >0.5); Quality is weak (ROE dispersion historically low at 3.3% vs US 8.4%); Momentum is negative in most sectors; Growth ~0.1.

Dynamic Metric Pool:

• Over the full sample (1996–2025), the dynamic pool does not consistently outperform core metrics, but its recent five-year advantage is decisive across all regions, validating the need for periodic refreshes.

Machine Learning Layer:

• AdaBoost uses seven training universes (12-month, 60-month, all history, four regime-based). Predictions are sector- and beta-neutral.
• Technical indicators exhibit nonlinear signal profiles across quintiles, justifying the ML approach over simple ranking.
• The ML sub-composite’s correlation with core factor returns is low (e.g., US: correlations with Value, Quality, Momentum, Growth, Low Risk, Size all below 0.15).
• Simplifying the ML layer (reducing indicators from ~200 to 26 key ones, lowering training rounds) cuts US 10-year Sharpe by more than 30%, highlighting the value of the full indicator set.

Enhanced Filters:

• Quality-Value interaction is concave: quality deteriorates fastest in the cheapest decile. Excluding the bottom decile of both (or bottom 5% in Europe) improves Sharpe and reduces maximum drawdown per volatility by ~5% (US full sample).
• Momentum-Revisions filter: residual momentum and earnings revisions have concurrent correlation of 0.5–0.7; removing stocks in the top quintile of residual momentum but bottom decile of earnings revisions eliminates conflicting signals. The improvement is primarily from the long side.

Regional Heterogeneity:

• US: Healthcare ROIC Sharpe (pre-cost, 10-year) = 1.38; Communication Services 6-month momentum = 0.44.
• Europe: Healthcare Up/Down EPS revisions Sharpe = 1.02; Materials FCF yield = 0.86.
• Japan: Banks Fwd P/E Sharpe = 1.54; Pharmaceuticals 9-month momentum = 1.18; but most sectors show negative momentum Sharpe (e.g., Banks -0.32).
• Size factor is universally negative across regions and sectors.

Cost and Implementation:

• Assumes 2 bps transaction costs, beta-neutral and sector-neutral implementation. Beta-neutral yields materially higher Sharpe than cash-neutral (US long-term cash-neutral Sharpe is 40% lower).
• Average monthly turnover exceeds 200% (US 223%, Europe 210%, Japan 234%).

Key Risks

1. Factor leadership shifts – If Value or Momentum underperforms for prolonged periods (e.g., US value post-COVID drawdown), QAML’s exposure to these factors could cause relative losses. The dynamic pool may help but cannot eliminate regime risk.
2. Dynamic pool staleness – The metric pool relies on periodic analyst reviews. Failure to refresh inputs as market conditions evolve will likely erode the recent performance advantage.
3. ML overfitting – Although sensitivity tests suggest robustness, the framework’s complexity (200+ indicators, 50 training rounds) raises the probability of sample-specific overfitting. The simplification penalty (Sharpe drop >30%) shows that vulnerable points exist.
4. Regional applicability limits – Europe’s Quality-Value filter adds only +0.02 Sharpe; Japan’s momentum filter is effective but the overall Sharpe for Enhanced Composite over the last five years (0.89) is still below the US (1.25). The model is not uniform.
5. Turnover and cost risk – >200% monthly turnover implies substantial trading costs. Although the backtest includes 2 bps costs (with short funding costs and dividend withholding), real-world slippage could be higher. Cash-neutral implementation cuts Sharpe by ~30%+ (US >40%), indicating high sensitivity to hedging methodology.
6. Specific market regimes – Technical indicators show extreme sensitivity to VIX: e.g., US 12-day EMA returns 17.8% in falling VIX but -6.5% in rising VIX. A prolonged high-VIX period would hurt the ML layer’s trend-following signals.

Valuation & Trading Implications

• Recommended implementation: Beta-neutral, sector-neutral long/short portfolio with monthly rebalancing. Use the Enhanced QAML Composite as the core construction tool.
• Cost management: Given turnover >200%, mandate low-commission execution and consider limiting to highest-liquidity stocks (top 90% of market cap within each region) to reduce impact. The backtest already excludes bottom 10% by size, and including smaller stocks boosts Sharpe for Europe (twofold improvement) but may increase costs.
• Regional customization: For Japan, overweight Value and underweight Quality; for US, emphasize Quality and momentum; for Europe, maintain balanced exposure but note the weak Quality-Value filter.
• Risk overlay: Monitor VIX regimes and consider dynamically adjusting the weight of trend vs. mean-reversion indicators in the ML layer. The framework already does this through regime-based training, but explicit risk budgeting can avoid disproportionate drawdowns.
• Monitoring frequency: Re-assess the dynamic metric pool at least semi-annually; test correlation thresholds (0.2–0.5 range has similar Sharpe, but US/Japan optimal at 0.5, Europe at 0.4).
• Alternative use cases: The QAML framework can be applied to single-sector portfolios by leveraging the sector-specific factor Sharpe tables (Exhibits 97–102) to tilt exposures. However, the full composite’s diversification benefit is superior.

Appendix Data Summary (Compressed)

Factor Definitions Summary: The framework uses 82 fundamental equity factors (Value, Growth, Momentum, Revisions, Quality, Risk, Size) and 50+ technical indicators (moving averages, oscillators, volume, volatility, short interest). Full definitions are provided in Exhibits 105–106. Critical technical indicators: 12-day EMA (trend), Bollinger Bandwidth (mean-reversion), market beta (risk). Their performance varies sharply across VIX regimes (see Exhibit 103).