I did another experiment. According to your suggestion, I used 3 years of data for training, 10% OSS and the trading period is one month. To avoid focusing on one type of market condition, I spread the trading periods between 2019 to 2023. The trading periods are separated by 4 months. In addition, I added Monte Carlo simulation for the market variations to increase the robustness. The result is shown above.
It seems that the result has no improvement. Do you know what is going wrong?