This study proposes an asset allocation strategy recommendation system that
reflects the structural inefficiency and nonlinear dynamics of financial markets.
Specifically, a Markov Switching Model (MSM) is employed to classify market
conditions into t...
This study proposes an asset allocation strategy recommendation system that
reflects the structural inefficiency and nonlinear dynamics of financial markets.
Specifically, a Markov Switching Model (MSM) is employed to classify market
conditions into two distinct regimes—bull and bear—while a machine learning
technique, the Light Gradient Boosting Machine (LightGBM), is applied to predict the
optimal asset allocation strategy for each regime. This framework overcomes the
limitations of traditional static asset allocation approaches by providing a dynamic
and adaptive investment methodology, thereby offering both academic and practical
implications.
The traditional Mean–Variance Optimization (MVO) model (Markowitz, 1952)
derives the optimal portfolio based on expected returns and the covariance matrix of
assets. However, in practice, its application is constrained due to the difficulty of
estimating expected returns and the model’s high sensitivity to input parameters.
Furthermore, the time-varying and structurally changing nature of asset returns
often diminishes predictive accuracy under fixed model assumptions. To address
these challenges, various alternative approaches such as the Risk Parity model—
which allocates assets based on risk contributions rather than expected returns—
and simpler strategies like the Minimum Variance Portfolio (MVP), Equal Weight
(EW), Dynamic Momentum (DM), and Traditional Asset Allocation (TAA) have been
widely employed in practice.
Building on this background, this study quantitatively identifies market regimes
through the MSM and evaluates the daily performance of five asset allocation
strategies: Dynamic Momentum, Risk Parity, Minimum Variance, Traditional Asset
Allocation, and Equal Weight. Subsequently, a LightGBM-based classification model
is trained on regime-dependent patterns of strategy performance, enabling the
system to recommend the strategy expected to yield the highest return under similar
market conditions.
The empirical analysis utilizes daily return data from major U.S. ETFs between
August 2003 and 2024, with Winsorization applied to mitigate the influence of
outliers. The backtesting results demonstrate that the proposed machine learning–
based recommendation system achieved an annualized return of 17.45%, a Sharpe
ratio of 1.655, and a maximum drawdown of 14.22%. Compared with the unadjusted
model (18.82%, 1.412, and –19.69%, respectively), the proposed system exhibited
slightly lower returns but achieved a 17% improvement in risk-adjusted
performance and a 5.47% reduction in drawdown.
Moreover, this study empirically confirms that the LightGBM classifier effectively
captures complex nonlinear interactions among variables, thereby improving both
regime identification accuracy and strategy recommendation precision compared
with traditional statistical models. Nevertheless, this research does not consider
transaction costs, taxes, or slippage, and it simplifies the regime structure into two
states (bull and bear). Future studies could extend this framework by incorporating
multi-regime models, macroeconomic indicators, and time-series forecasting
methods to further enhance its practical applicability and predictive performance.