Is this an idealized condition?

[Dreamlikec]

I see your code in 'RL/agent/low_level/ddqn_pes_risk_aware.py', the default transaction_cost is set to 0.00015, As we known most of exchanges have a commission fee rate of 0.001, So is the result in your paper as such low commission fee rate ?

[qinmoelei]

The commission fee rate is from Binance. If you have an account of VIP 9 and use BNB to settle the commission fee, the rate is 0.00015.

It is quite common in distributed exchanges to have such a low commission fee since those exchanges, unlike traditional markets like the stock markets, need to attract customers all by themselves, and lowering the commission fee is the most common way (Binance will even cancel a trading pair's commission from time to time). Since we are doing high-frequency trading, the commission fee rate has an extremely high impact on our results, and therefore, we choose the commission fee with which most of the orders are conducted in this market.

The commission fee rate for different trading pairs differs in my paper. For more details, see Appendix D.3 in the paper.

[Dreamlikec]

Thank you for your reply, so another question : could you please recommend a crypto RL-based implementation profitable as you known for an individual trader ??

[qinmoelei]

In response to your concern, I would like to clarify a few points:

No ready-to-use solution: Open-source GitHub repositories typically offer trading frameworks, but there’s no solution where you can simply input an API key and expect profitability in real trading without substantial tuning. Zero-sum nature of trading: Trading is often a near zero-sum game. Public technical indicators lose effectiveness quickly once widely used, making profitability hard to sustain with open strategies. Framework vs. Strategy: Most repositories provide general tools for trading but not plug-and-play strategies. Effective trading requires adapting the framework to the market’s specific conditions. Importance of customization: For real profitability, it’s crucial to focus on building unique technical indicators rather than relying on widely available machine learning models or existing strategies. Other key factors: Profitability depends not just on the asset but also on trading frequency, risk tolerance, and one’s role in the market (e.g., market making, algorithmic trading).

[Dreamlikec]

Thank you for your response. I understand that building a profitable algorithm is a systematic and complex task. In my experience with reinforcement learning in real trading environments, I found that even when the model performs well during training and backtesting, its performance in actual application is relatively average. I've noticed overfitting in reinforcement learning and the more complex phenomena in actual trading.

Originally, I wanted to ask if you or your team have used reinforcement learning to trade cryptocurrencies in real scenarios and achieved relatively good returns. Based on your experience, what thoughts or directions do you think I should explore? I look forward to your insights, thx.