QAML Asia stock-selection framework Report Interpretation

Overview

Morgan Stanley extends its Quant + Analyst + Machine Learning stock-selection framework to Asia Pacific ex Japan. The central conclusion is that ML is most useful as a differentiated source of information and diversification, rather than as a sleeve to maximize; the institution therefore prefers a 75/25 Quantamental/ML allocation and retains it as the strategic benchmark.

Core views

QAML combines three stock-selection pillars in the MSCI AC Asia Pacific ex Japan IMI universe: Core Metrics, an equal-weighted set of Value, Quality, Momentum, Growth and Low Risk signals; the analyst-informed Dynamic Metric Pool; and an AdaBoost ML sleeve using fundamental and more than 100 technical indicators. The first two form the Quantamental sub-composite. Securities are scored sector-relatively, the top quintile is held long and the bottom quintile short, with long-book leverage adjusted to make the portfolio beta-neutral. Signals are measured at month-end, portfolios rebalance monthly and reported USD results are pre-cost. The report’s main investment case is diversification across changing sources of alpha rather than reliance on a dominant signal. Over the full sample since 2002 through 31 July 2026, QAML delivered a 23.6% long-term CAGR, ahead of the Dynamic Metric Pool at 21.0%, Core Metrics at 20.1% and ML at 17.9%. QAML also led absolute returns across every reported horizon. Risk-adjusted leadership among standalone sleeves rotated: the Dynamic Metric Pool led over five and three years, Core Metrics over 10 years, and ML over one year and YTD. Morgan Stanley therefore views the composite’s benefit as stronger compounding from complementary inputs rather than proof that every sleeve consistently outperforms on its own. ML supplies the most differentiated component. Core Metrics and the Dynamic Metric Pool had a 0.93 return correlation, while ML’s correlation with each was only 0.33. The QAML Composite showed its strongest style correlations with Quality at 0.35 and Momentum at 0.33, followed by Value at 0.26, but the report says no single style factor dominates. A Barra Momentum neutralization exercise reduced returns and CAGR/Vol, showing Momentum was a meaningful historical contributor, yet returns remained positive after neutralization. Morgan Stanley treats this as a sensitivity diagnostic rather than causal attribution, because factor exposures overlap and cannot recreate a counterfactual zero-Momentum portfolio. The ML model uses AdaBoost to identify signals that distinguish future outperformers from underperformers across rolling 12- and 60-month samples, macro regimes and calendar seasonality. It uses 50 boosting rounds and a learning rate of 1.0. Fundamental signals remain repeatedly important, but more recent ML selection increasingly emphasized trading activity and market structure: Share Turnover 252d rose from outside the top 20 factors in 2002-2010 to a 20.0% selection frequency in 2011-2020 and 28.8% in 2021-2026. The report cautions that selection frequency is descriptive of early boosting iterations, not a measure of final factor contribution. Drawdown analysis motivates treating ML exposure as a portfolio-calibration decision. In June-July 2022, losses were broad-based across QAML pillars, although ML declined most; reducing ML alone would thus have offered only partial protection. In September-October 2024, losses were much more concentrated in ML, making the composite more sensitive to ML weight. Relative to historical 50/50 QAML, a 75% Quantamental/25% ML allocation raised long-term CAGR/Vol from 3.27 to 3.33, reduced full-sample maximum drawdown from -21.8% to -19.6%, and reduced the 2024 maximum drawdown from -16.5% to -11.4%. Eliminating ML cut the 2024 drawdown further to -5.9%, but reduced CAGR from 23.6% to 21.0% and CAGR/Vol from 3.27 to 3.12. Morgan Stanley consequently selects 75/25 as the preferred APxJ strategic allocation, not as a universally optimal weight. Adaptive QAML keeps the stock-selection architecture unchanged but dynamically varies ML risk around the 75/25 baseline. A monthly ML Risk Score equally weights relative ML volatility, ML drawdown risk, ML-to-Quantamental correlation, and Momentum/Low Vol factor stress. Based solely on prior month-end information and its trailing 10-year distribution, the allocation moves to 50/50 when risk is below the 10th percentile, remains at 75/25 through the middle 80%, and shifts to 100/0 when risk is at or above the 90th percentile. From January 2012 to July 2026, it spent 123 months, or 70.3% of the sample, at the 75/25 baseline; 21 months at 50/50; and 31 months at 100/0. Adaptive QAML produced broadly similar returns but lower volatility and drawdowns than static 75/25, with its protection more useful in ML-concentrated stress than in broad-based stress. Morgan Stanley nevertheless retains static 75/25 as its preferred benchmark pending further validation. Implementation remains unresolved. Average monthly whole-book turnover over 2012-2026 was 44.7% for static 50/50, 36.5% for static 75/25, 33.8% with no ML, and 37.5% for Adaptive QAML. The report says lower turnover is not evidence of investability because results exclude transaction costs, taxes, financing, borrow, liquidity and market impact. Small-cap inclusion expands the IMI opportunity set but adds liquidity, capacity and borrow constraints. A standard APxJ universe excluding small caps still showed QAML efficacy, but its long-term CAGR and CAGR/Vol fell to 16.8% and 2.0, respectively, from 23.6% and 3.3 in the APxJ IMI universe.

Analysis framework

Morgan Stanley constructs sector-relative, beta-neutral long/short backtests, compares each QAML pillar and their composite across long and shorter horizons, then tests correlation, style exposure, drawdowns and alternative ML weights. It separately examines ML parameter robustness, uses a historical risk score to test adaptive allocation, and flags the need to validate pre-cost results against APxJ trading frictions.

Methodology notes

• Quantitative, Factor, and Portfolio TheoryMulti-factor model

  Multi-factor Quantamental stock selection

  Core Metrics averages Value, Quality, Momentum, Growth and Low Risk signals, while the Dynamic Metric Pool adds analyst-suggested metrics subject to correlation and dispersion controls.

• Quantitative, Factor, and Portfolio TheoryStyle factor analysis

  Correlation and Barra style-exposure diagnostics

  The report compares component correlations and sensitivity to Quality, Value and Momentum proxies to identify how differentiated QAML’s return sources are.

• Other

  AdaBoost machine-learning classification

  The ML sleeve sequentially selects signals that distinguish future outperformers from underperformers, emphasizing observations that were difficult to classify in earlier rounds.

• Other

  Adaptive ML risk budgeting

  A standardized score from ML volatility, drawdown, correlation and factor stress determines whether ML exposure is increased, maintained or reduced at the next monthly rebalance.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• QAML Asia strategy

  APxJ beta-neutral long/short stock-selection framework combining Quantamental and ML sleeves.

  Strengths

  Strongest historical absolute return profile across reported horizons; ML adds differentiated information.

  Weaknesses

  Performance is pre-cost and risk-adjusted leadership rotates across components.

  Comparison

  Preferred 75/25 allocation improved historical return-risk and drawdown metrics versus original 50/50; no-ML reduces drawdown further but gives up return and diversification.

  Risks

  ML-specific drawdowns, broad factor stress, turnover, liquidity, borrow, financing, market impact and capacity constraints.

Key data

• QAML Composite long-term CAGR23.6%Pre-cost APxJ IMI backtest since 2002 through 31 July 2026; above all standalone pillars.
• QAML Composite long-term CAGR/Vol3.27Original 50% Quantamental / 50% ML specification.
• Preferred 75/25 CAGR/Vol3.33Versus 3.27 for original 50/50 QAML.
• Full-sample maximum drawdown-19.6%Preferred 75/25 allocation versus -21.8% for 50/50.
• 2024 maximum drawdown-11.4%Preferred 75/25 allocation versus -16.5% for 50/50.
• ML correlation with each Quantamental pillar0.33Versus 0.93 correlation between Core Metrics and the Dynamic Metric Pool.
• Average monthly turnover36.5%Static 75/25 over 2012-2026, versus 44.7% for 50/50 and 37.5% for Adaptive QAML.

Impact & implications

The report argues that ML should be assessed for its incremental diversification and information contribution, not only standalone return. For APxJ, Morgan Stanley favors a lower strategic ML risk budget than the original 50/50 design; Adaptive QAML may further improve risk efficiency when ML-specific stress is elevated, but requires post-cost and robustness validation.

Risks

• All historical results are pre-cost and may not survive APxJ transaction costs, taxes, financing, borrow and market-impact assumptions.
• Small-cap inclusion can increase liquidity, capacity, market-impact and borrow constraints.
• Adaptive QAML is not a universal drawdown hedge; it provides more limited protection when stress is broad-based across all pillars.
• Country, currency, sector and factor exposures remain relevant diagnostics in the regional long/short implementation.
• Capacity cannot be inferred from the current backtest and requires security-level liquidity and market-impact analysis.

What to watch

• Post-cost testing of the 50/50, preferred 75/25 and Adaptive QAML specifications using APxJ transaction-cost, tax, financing, borrow and market-impact assumptions.
• Walk-forward robustness of Adaptive QAML across lookback windows, state thresholds and switching frequency.
• Whether APxJ-specific yield and yield-curve regime inputs improve on the current US rate proxies.
• The economic trade-off between Adaptive QAML’s risk reduction and its incremental turnover.
• Extension of QAML and ML-risk calibration from APxJ to Emerging Markets.