Volatility Forecasting and Return Prediction under Market Regimes: Evidence from High-Frequency Chinese Equity Data

Abstract: This study investigates whether regime-dependent volatility forecasting and machine-learning-based return prediction can be jointly integrated to improve both statistical forecasting performance and economic strategy outcomes in equity markets. Using high-frequency CSI 300 Index data from 2005 to 2023, a sequential twostage framework is developed. In the first stage, realized volatility is modeled using regime-augmented HARQ specifications combined with Markov-switching GJR-GARCH filtering to capture long-memory dynamics, asymmetry, and structural market regimes. In the second stage, volatility forecasts, regime indicators, and return-related predictors are incorporated into an XGBoost return-prediction model estimated through a strictly walk-forward out-of-sample procedure. The empirical results demonstrate that regime-aware volatility forecasting consistently outperforms baseline HARQ models across forecast evaluation metrics and is generally supported by formal forecast comparison tests. In contrast, return predictability remains weak, state-dependent, and concentrated primarily in low-volatility regimes. Although naive predictive trading strategies generally fail after accounting for realistic transaction costs, carefully designed implementations incorporating volatility scaling, low-volatility gating, threshold calibration, and turnover controls can improve defensive economic performance. The findings suggest that the practical value of predictive systems in financial markets may depend less on generating strong unconditional return forecasts and more on transforming weak state-dependent signals into economically robust portfolio allocation rules. Overall, the study contributes by integrating econometric volatility modeling, regime classification, machine-learning return prediction, and implementation realism within a unified framework.

• The paper introduces an integrated regime-sensitive HARQ and XGBoost framework for volatility forecasting and return prediction in high-frequency Chinese equities.
• The methodology leverages Markov-switching GJR-GARCH and walk-forward validation to extract regime-dependent signals and enhance predictive accuracy.
• Implementation realism through transaction cost controls and exposure gating leads to robust portfolio outcomes and improved drawdown management.

Regime-Aware Volatility Forecasting and Return Prediction for High-Frequency Chinese Equities

Introduction

This paper introduces an integrated regime-sensitive volatility and return-prediction architecture for high-frequency Chinese equity markets, specifically the CSI 300 index. The core methodological contribution is the unification of a regime-augmented realized volatility forecast using HARQ and Markov-switching GJR-GARCH, with a downstream XGBoost-based return-prediction module. The empirical design leverages strictly walk-forward out-of-sample validation and incorporates comprehensive controls for implementation realism, including explicit transaction costs, turnover constraints, and volatility-conditioned signal gating. The central question interrogated is whether regime-aware hybrid frameworks yield not only statistical improvements in volatility forecasts but also economically meaningful, robust portfolio outcomes in a market with well-known regime structure and low unconditional return predictability.

Realized Volatility Modeling with Regime Dependence

The volatility modeling approach centers on exploiting rich high-frequency data characteristics through realized variance and quarticity measures. Extensive distributional diagnostics reveal canonical stylized facts, including pronounced clustering, persistence, and extreme heavy tails in both returns and volatility.

Figure 1: Distribution of daily returns for the CSI 300 index is leptokurtic with negative skew, highlighting extreme market movements.

Figure 2: QQ plot showing severe tail risk and deviation from Gaussianity in daily returns.

Motivated by these diagnostics, the first stage utilizes a HARQ backbone for log-realized variance, incorporating lagged realized variance components (daily, weekly, monthly) and realized quarticity, with all features constructed from 5-minute returns.

Crucially, HARQ residuals are subjected to a two-state Markov-switching GJR-GARCH model with Student-$t$ shocks. This hybridization enables identification of persistent latent volatility regimes with regime probabilities filtered using only contemporaneous data—critical for out-of-sample integrity. Periods of high $p_t$ correspond to empirical crisis episodes, confirming the macroeconomic interpretability of the extracted regime signal.

Figure 4: Time series plot of log-realized variance shows volatility clustering and regime shifts over nearly two decades.

Figure 6: ACF of log-realized variance demonstrates long-memory and persistent autocorrelation up to monthly lags.

Figure 3: Regime identification from the MS-GJR-GARCH overlaying realized volatility; states of persistently elevated $p_t > 0.5$ align with market stress.

The regime probability $p_t$ is then fed back as an exogenous input into an augmented HARQ, leading to substantial outperformance over HARQ-only baselines in all forecast error metrics (MSE$_{log}$, QLIKE, LMAE) and in Mincer–Zarnowitz regressions.

Machine Learning Return Prediction under Regime Constraints

Return forecastability on daily Chinese equity returns is challenged by an extremely low signal-to-noise ratio. XGBoost is selected for nonparametric learning power, trained recursively under a strictly walk-forward regime. The predictor set fuses volatility estimates (HARQ + $p_t$-based forecasts and lagged volatility), regime features (filtered regime memberships, interactions), and recent return dynamics.

Figure 5: Feature importance for the XGBoost predictor set is dominated by realized and predicted volatility, regime features, and lagged returns, underscoring volatility's relevance for risk premia.

A pronounced regime-dependence emerges: statistically significant positive correlations and direction-correct forecasting (Hit Ratio $\sim$54%, HAC $p<0.001$) exist predominantly in low-volatility states, but degrade or invert during detected high-volatility regimes. The time-varying rolling information coefficient analysis further corroborates this pattern.

Figure 10: Rolling IC for return predictions reveals transient, nonstationary pockets of predictability that cluster in low-volatility intervals.

Figure 7: The rolling Signal$\times$Risk IC and high-volatility regime probability $p_t$ over time demonstrate positive IC in calm regimes and unreliable or negative IC during stressed regimes.

Implementation Realism: Robust Trading Strategy Design

Economic evaluation is conducted through a suite of signal engineering and allocation rules. The main proposal—Low-vol Gated Weekly Signal$p_t$0Risk—applies return forecasts scaled by anticipated volatility (forecasted $p_t$1), natural gating to low-volatility periods, and conservative thresholding and turnover controls. Parameters (signal threshold, rebalancing frequency, no-trade bands, transaction costs) are selected ex ante and further stress-tested for sensitivity.

Figure 8: Schematic of expanding-window walk-forward design for recursive model re-estimation, enforcing strict out-of-sample forecast discipline.

Out-of-sample performance (2014–2023) evidences strong drawdown mitigation and volatility targeting relative to buy-and-hold and momentum-type benchmarks.

Figure 13: Out-of-sample equity curves: the proposed strategy generates stable growth and substantially mitigates major drawdowns observed in the benchmark.

Notably, Sharpe and Sortino ratios, maximum drawdown, and return volatility are all superior to benchmarks, despite only marginal improvement in raw annualized returns. This outcome holds across multiple regime, feature-ablation, and hyperparameter robustness checks; multi-testing corrections confirm that positive active information ratios arise from implementation discipline rather than ex post parameter tuning.

Figure 9: Annual return comparison shows the strategy’s defensive bias—outperforming in crash years, lagging in strong bull years.

Figure 11: Annual maximum drawdown consistently favors the regime-aware strategy, with capitalization of risk-off positioning during market stress.

Empirical and Practical Implications

• Regime-aware HARQ models, whether via hard or soft regime variables, yield robust improvements in volatility forecasts and outperform simple long-memory benchmarks, especially in turbulent periods.
• Return predictability is fragile and primarily conditional, being statistically and economically meaningful only in low-volatility (low $p_t$2) states. In high volatility, return forecasts not only degrade but may become anti-predictive.
• Naïve application of predictive signals fails under realistic costs; only architectures with volatility-aware gating, exposure scaling, and turnover suppression maintain positive economic performance.
• Practical allocation rules must focus less on maximizing average returns and more on downside protection and adaptive exposure suppression, in line with observed regime-state conditionality.

Limitations and Future Directions

Despite the design’s empirical rigor, important constraints remain. The framework is single-market and does not model cross-asset or multi-country dependencies. Realistic frictions, such as Chinese short-sale constraints, are only partially incorporated. XGBoost, while powerful, does not resolve model selection or instability issues endemic to low-flow financial prediction. Transferability to other asset classes with different microstructure and regime dynamics is untested. Future research should aim for joint structural estimation of regime-state and returns, richer cost modeling, and application to high-frequency futures or ETF vehicles.

Conclusion

This study demonstrates that realized volatility in Chinese equities can be robustly predicted via regime-augmented models, and that such forecasts provide critical state information for risk-aware allocation. Despite limited unconditional return predictability, regime-aware frameworks enable conditional transformation of weak, low-volatility-aligned signals into strategies with improved defensive properties—specifically, volatility targeting and drawdown reduction—contingent upon disciplined implementation and avoidance of overfitting. The main implication is that practical economic value in these markets is derived not from raw statistical outperformance, but from regime-dependent signal reliability and engineering of exposure control.

Figure 12: Sensitivity of strategy Sharpe ratio to signal thresholding $p_t$3 and no-trade band $p_t$4 demonstrates the importance of ex ante conservative parameterization for robustness.