A-Share ETF Rotation Backtest — Full Breakdown & Transcript

【论文复现】行业ETF切点轮动夏普1.3

0h 07m video Published Aug 5, 2026 Transcribed Sep 17, 2026 量化Quantgirl

21 views Recent velocity 0.0 views/hour View full performance history →

Advanced 4 min read For: Quantitative finance professionals and advanced investors interested in ETF rotation strategies and backtesting methodologies.

AI Trust Score 65/100

⚠️ Average / Some Fluff

"The title promises a research-based study and delivers a detailed backtest, though the use of proxies instead of full RL training slightly oversells the 'deep reinforcement learning' angle."

AI Summary

This video presents a backtest study of A-share industry ETF rotation using a smart switching-point combination framework. The research applies deep reinforcement learning concepts, but this backtest uses heuristic proxies instead of full RL training. The default 'Blend' strategy achieves an annualized return of about 16.52% with a maximum drawdown of -15%, demonstrating the framework's potential on Chinese thematic ETFs.

Chapters

1 Introduction and Framework 00:00 2 Backtest Setup and Data 01:02 3 Performance Comparison 02:36 4 Return Distribution and Composition 04:20 5 Cautions and Limitations 05:29

[00:00]

Framework Introduction

The study proposes a switching-point method that uses deep reinforcement learning to simultaneously decide risk aversion and rebalancing cadence. It achieved significant excess returns on US industry ETFs, and this backtest applies it to 12 Chinese thematic ETFs.

[00:30]

Backtest Setup

The backtest period is from June 2021 to July 2026. The default Blend scheme achieves an annualized return of about 16.52%, a Sharpe ratio of 1.30, and a maximum drawdown of about -15%. After deducting 20 basis points of costs, the Sharpe ratio remains at 1.28.

[01:02]

Strategy Core Metrics

The strategy always holds full long positions with non-negative weights summing to one. The core metrics are the risk aversion parameter LAMBDA (1 to 100) and the holding period CH (5 to 60 trading days). After the holding period, weights are rebalanced using a mean-variance framework based on rolling 252-day daily returns, supporting MV, MSD, and CVaR utility functions.

[01:32]

Trading and Rebalancing

The dynamic scheme trades daily, while the static baseline rebalances every 30 days. The dynamic scheme follows the H value from the switching rule. Initial net value is set to 1000, and the risk-free rate is taken as 0. Data sources differ: the original study used Yahoo daily data for US ETFs, while this backtest uses locally stored A-share ETF data.

[02:08]

Data and Benchmark

The backtest uses 12 A-share ETFs covering themes like gold, soybeans, Nasdaq, ChiNext, and new energy vehicles. The benchmark is the CSI 300 plus a defensive ETF. The control group uses heuristic MSV with volatility sleeves or static grids, with no reinforcement learning training. The estimation window is 252 trading days, static rebalancing every 30 days, and dynamic holding periods between 5 and 60 days.

[02:36]

Performance Comparison

The chart shows net values and drawdowns for each scheme. The default Blend is the most stable, with maximum drawdown around -15%. In contrast, the full portfolio and CSI 300 benchmarks have drawdowns of -33% and -40%, respectively, indicating that thematic diversification alone is insufficient.

[03:01]

Alternative Schemes

The 'Other' scheme has the highest cumulative return but the deepest drawdown at -30%, with volatile net value swings, making it unsuitable as a default. The static switching-point scheme achieves about 17% return with a Sharpe of 1.12 but suffers a -17% drawdown, which is risk-adjusted worse than Blend. Blend's longest underwater period is about 89 trading days.

[03:42]

Risk-Adjusted Performance

Blend has the highest Sharpe ratio at 1.30. The MSV-GT scheme has a slightly higher annualized return of 16.97% but a deeper drawdown of -17.53%. The full portfolio only achieves 6.23% annualized with a Sharpe of 0.32, showing that switching-point grids significantly improve risk-adjusted returns.

[04:06]

Cost Sensitivity

Blend's annualized turnover is 1.2 times. After deducting 20 basis points, the annualized return drops to 16.2% and the Sharpe ratio to 1.282, a decline of only about 0.02. This indicates the strategy is not a thin-edge strategy that collapses with costs, showing strong real-world tolerance.

[04:20]

Return Distribution

Blend's returns are evenly distributed from 2022 to 2025, with interval returns of 10.4%, 11.4%, 1.4%, 17.4%, and 40.6% respectively. In 2026 up to July, it recorded 7.6%. The Sharpe ratio in 2025 was as high as 2.38%, but this strength is not extrapolatable as normal; in 2026 it fell to 0.64%, showing market dependence.

[05:00]

Blend Composition

Blend is a half-and-half mix: one half uses the improved dynamic scheme with MSV utility and volatility sleeves, and the other half is the static switching-point improved dynamic with volatility sleeves. The target 18% volatility sleeve and MSV suppress concentration, avoiding the issue of maximum weight averaging 0.73.

[05:13]

Robustness Check

The rolling window comparison confirms that similar schemes maintain Blend's drawdown at -15.1%, still leading. The actionable part is clear: use Blend as the default configuration, applying switching-point logic to A-share industry ETFs to obtain risk-adjusted returns with low cost burden.

[05:29]

Cautions and Limitations

If real-world impact costs are significantly higher than the 20 basis point assumption, one should prioritize reducing holding concentration or lowering rebalancing frequency rather than directly applying the old heuristic. Two things are absolutely not to be done: treating the old heuristic as a dynamic scheme for live trading (its high returns come with extreme volatility and drawdowns), and equating this proxy's conclusions with the original deep reinforcement learning results without actually running PPO.

[06:08]

Future Directions

Exploring other utility functions like CARA within the volatility sleeve framework is a possible extension, but requires independent validation of robustness. The summary states that the current best executable scheme is Blend with about 16.5% annualized return, Sharpe 1.30, and max drawdown about 15%, even after costs.

[06:44]

Key Takeaways

The switching-point scheme, when fully applied to the whole portfolio and compared with CSI 300, shows that combining switching points with volatility sleeves or mixing dynamic approaches is more reasonable than using the old heuristic. The old heuristic's high returns cannot be used as evidence of intelligent switching. The improved scheme effectively reduces concentration risk through MSV utility and volatility sleeves, which is the characteristic the dynamic part should have.

[07:09]

Final Disclaimer

This backtest is entirely based on A-share ETF proxies, with no PPO training. Conclusions apply only to this universe and window. Liquidity impact in live trading still needs separate evaluation.

The Blend strategy, combining switching-point logic with volatility sleeves and MSV utility, offers a robust risk-adjusted performance on A-share industry ETFs. However, its results are proxy-based and require careful consideration of costs and market dependence before live deployment.

Mentioned in this Video

PPO

tool

A2C

tool

Yahoo Finance

tool

💡 Key Takeaways

📊

Blend's Strong Performance

Demonstrates the framework's potential with a high Sharpe ratio and controlled drawdown.