Cross-asset market transmission and spillovers Report Interpretation

Overview

Morgan Stanley extends its QAML machine-learning framework from stock selection to cross-asset market transmission. It concludes that cross-asset market states can help distinguish stronger from weaker subsequent risk-adjusted outcomes, and that a combined multi-horizon signal provides the most robust portfolio implementation.

Core views

The report asks whether the current state of one market contains useful information about what happens next in another, rather than merely whether markets move together. It applies an AdaBoost framework separately to each target asset, allowing the relevant source market, threshold and direction to differ by target and forecast horizon. The universe covers 116 assets across equities, fixed income, short rates, FX forwards, commodities and credit, using 221 signals from 13 liquid source markets. Inputs include 1-, 3-, 6- and 12-month volatility-adjusted returns, momentum and trend indicators, price position relative to historical ranges, tail behavior and volatility-regime measures. For each target asset, Morgan Stanley trains separate models for 1-, 3-, 6- and 12-month forward realised Sharpe ratios. Outcomes are classified as stronger or weaker relative to each asset’s expanding median forward Sharpe. The model uses expanding time-series windows, begins only after 60 eligible monthly observations, and adds observations only once their forward outcomes have fully realised. It calibrates each out-of-sample model score into an expected forward Sharpe relative to that asset’s own long-term expected Sharpe, so that the final signal indicates whether the current cross-asset environment looks historically better or worse than normal for that particular asset. The backtest indicates persistent performance across all four horizons. From February 2013 to July 2026, the equal-weighted Average 1/3/6/12-month long/short signal generated a 3.7% CAGR, 5.2% annualised volatility, a 0.71 net-of-cost CAGR/volatility ratio, and an 8.6% maximum drawdown. Its information ratios were 0.81 against an equal-weighted global universe and 0.79 against an inverse-volatility universe. The 1-month signal was the strongest individual horizon, with 2.7% CAGR, 4.0% volatility and a 0.68 ratio, while the 3-, 6- and 12-month signals also remained positive. Morgan Stanley interprets the stronger combined result as evidence that averaging horizons captures transmission occurring at different speeds and reduces dependence on a single forecast window. The authors test whether the result depends on model specification. Reducing the feature set to 52 volatility-adjusted return signals lowered the combined strategy’s CAGR/volatility ratio to 0.62, versus 0.71 for the full 221-signal baseline, indicating that technical market-state measures add information beyond returns alone. A full-signal alternative calibration based directly on conditional expected forward Sharpe produced a 3.9% CAGR and 0.75 CAGR/volatility ratio. Morgan Stanley describes this direct calibration as a complementary, broader cross-asset allocation specification, whereas the baseline is more tactical because it measures expected Sharpe relative to each asset’s own long-run norm. Both long and short books contributed positively over the full sample, although the long book was the larger and more persistent return driver while the short book added value more episodically. Exposure and return contribution differed materially: short rates dominated long-book gross notional after volatility scaling, while FX forwards dominated the short book. Yet equities and FX forwards were the leading long-term contributors to strategy returns, and short rates made a more modest return contribution despite their large notional weight. All six target asset classes contributed positively over the full sample, with leadership changing over time. A constrained portfolio test imposed fixed asset-class allocations before applying the same volatility sizing. This reduced natural concentration in FX forwards and short rates and broadened representation of equities, fixed income, commodities and credit. Individual horizon results weakened substantially under the constraint, but the Average signal retained positive net-of-cost performance: 3.4% CAGR, 5.1% volatility, a 0.66 CAGR/volatility ratio, and information ratios of 0.72 and 0.68 versus the two global benchmarks. The authors view this resilience as support for the multi-horizon specification, because averaging reduces sensitivity both to transmission speed and to the asset-class mix through which signals are expressed. The transmission analysis identifies commodities and fixed income as the dominant information sources for most target asset classes. For equities, higher Brent 6- and 12-month volatility-adjusted returns were historically associated with weaker subsequent outcomes, while stronger copper three-month momentum was associated with better outcomes. For fixed income, higher Brent 12-month momentum was associated with stronger subsequent outcomes, while stronger JPY six-month momentum was associated with weaker outcomes. For short rates, the US two-year Treasury’s distance from its 252-day low was associated with stronger outcomes, while the S&P 500 being closer to its 252-day high was associated with weaker outcomes. For FX forwards, stronger Brent momentum was associated with weaker outcomes, while gold and Nasdaq price-position signals were associated with better outcomes. For commodities and credit, front-end rates, equities and commodity signals all add information beyond own-market momentum. In credit specifically, stronger Brent 12-month momentum was associated with weaker subsequent outcomes, while stronger Euro-Schatz price-position and range signals were associated with better outcomes. Morgan Stanley stresses that these are not fixed one-way rules: the same source signal can have different, including non-linear, implications for different targets. The report’s central conclusion is therefore that cross-asset spillovers form a dynamic, target-specific and state-dependent network, rather than a simple hierarchy in which one market consistently leads another.

Analysis framework

Morgan Stanley builds asset-specific AdaBoost classifiers using expanding historical samples and percentile-based market states. It converts model outputs into economically interpretable forward-Sharpe expectations, ranks assets monthly, takes the top and bottom 20%, applies inverse-volatility sizing and a one-business-day implementation lag, and evaluates results net of fixed transaction, roll and margin-cost assumptions. It then tests alternative feature sets and calibrations, constrained asset-class allocations, return attribution and source-market signal importance.

Methodology notes

• Other

  AdaBoost machine-learning classification with expanding time-series training windows

  The model sequentially combines simple classifiers to identify source-market states that historically distinguish stronger from weaker future risk-adjusted outcomes for each target asset.

• Quantitative, Factor, and Portfolio TheorySharpe and Information Ratios

  Forward realised Sharpe calibration and benchmark-relative information ratios

  The report labels outcomes using forward realised Sharpe, calibrates model scores to expected Sharpe, and evaluates strategy performance relative to equal-weighted and inverse-volatility global-universe benchmarks.

• Quantitative, Factor, and Portfolio TheoryMulti-factor model

  Multi-horizon cross-asset signal ensemble

  The strategy combines separate 1-, 3-, 6- and 12-month forecasts so that information transmitted at different speeds can contribute to the portfolio signal.

Key data

• Target universe116 assets across six asset classesEquities, fixed income, short rates, FX forwards, commodities and credit.
• Source markets and signals13 source markets; 221 candidate signalsSignals combine risk-adjusted returns and technical market-state measures.
• Average signal performance3.7% CAGR; 5.2% annualised volatility; 0.71 CAGR/volatility; (8.6%) maximum drawdownFebruary 2013–July 2026, net of implementation costs.
• Average signal information ratios0.81 vs global equal-weighted universe; 0.79 vs global inverse-volatility universeRegression-based benchmark-relative performance measure.
• Alternative direct calibration3.9% CAGR; 5.2% annualised volatility; 0.75 CAGR/volatilityUses the full signal set and ranks directly on conditional expected forward Sharpe.
• Constrained Average signal3.4% CAGR; 5.1% annualised volatility; 0.66 CAGR/volatilityFixed asset-class targets imposed before volatility sizing.

Impact & implications

Morgan Stanley’s findings support treating cross-asset information as target-specific and state-dependent rather than relying on broad correlations or fixed directional relationships. The report favors the combined multi-horizon implementation because it remained positive across alternative specifications and under a materially different asset-class allocation structure.

What to watch

• Whether signal importance varies materially by geography within each asset class.
• Whether signal rankings remain stable under alternative weighting schemes, including rank-based approaches.
• Whether VAR-based lead-lag analysis captures information complementary to the model’s non-linear spillover relationships.