Kang Gao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kang Gao has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Ce Guo
Imperial College London, London, United Kingdom
1 shared publication - Changling Zhou
Research Center of Cyberspace Security, Peking University Changsha Institute for Computing and Digital Economy, Changsha, China
1 shared publication - Min Shi
School of Control Science and Engineering, Shandong University,Jinan,China
1 shared publication - Namid Stillman
Simudyne, London, United Kingdom
1 shared publication - Perukrishnen Vytelingum
Simudyne Limited, United Kingdom
1 shared publication Qingnan Lai
Computer Center, Peking University,Beijing,China
1 shared publication
Latest Publications
- Open Access
- 25 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 270-278
Deeper Hedging: A New Agent-based Model for Effective Deep Hedging
We propose the Chiarella-Heston model, a new agent-based model for improving the effectiveness of deep hedging strategies. This model includes momentum traders, fundamental traders, and volatility traders. The volatility traders participate in the market by innovatively following a Heston-style volatility signal. The proposed model generalises both the extended Chiarella model and the Heston stochastic volatility model, and is calibrated to reproduce as many empirical stylized facts as possible. According to the stylised facts distance metric, the proposed model is able to reproduce more realistic financial time series than three baseline models: the extended Chiarella model, the Heston model, and the Geometric Brownian Motion. The proposed model is further validated by the Generalized Subtracted L-divergence metric. With the proposed Chiarella-Heston model, we generate a training dataset to train a deep hedging agent for optimal hedging strategies under various transaction cost levels. The deep hedging agent employs the Deep Deterministic Policy Gradient algorithm and is trained to maximize profits and minimize risks. Our testing results reveal that the deep hedging agent, trained with data generated by our proposed model, outperforms the baseline in most transaction cost levels. Furthermore, the testing process, which is conducted using empirical data, demonstrates the effective performance of the trained deep hedging agent in a realistic trading environment.
- 25 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 270-278
Open AccessDeeper Hedging: A New Agent-based Model for Effective Deep Hedging
We propose the Chiarella-Heston model, a new agent-based model for improving the effectiveness of deep hedging strategies. This model includes momentum traders, fundamental traders, and volatility traders. The volatility traders participate in the market by innovatively following a Heston-style volatility signal. The proposed model generalises both the extended Chiarella model and the Heston stochastic volatility model, and is calibrated to reproduce as many empirical stylized facts as possible. According to the stylised facts distance metric, the proposed model is able to reproduce more realistic financial time series than three baseline models: the extended Chiarella model, the Heston model, and the Geometric Brownian Motion. The proposed model is further validated by the Generalized Subtracted L-divergence metric. With the proposed Chiarella-Heston model, we generate a training dataset to train a deep hedging agent for optimal hedging strategies under various transaction cost levels. The deep hedging agent employs the Deep Deterministic Policy Gradient algorithm and is trained to maximize profits and minimize risks. Our testing results reveal that the deep hedging agent, trained with data generated by our proposed model, outperforms the baseline in most transaction cost levels. Furthermore, the testing process, which is conducted using empirical data, demonstrates the effective performance of the trained deep hedging agent in a realistic trading environment.
- 23 September 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 57, 69-73
Identification and Analysis of IoT Devices Based on Multilayer Perceptron
loT devices have provided great convenience for people's life and work, but hacking risks often tag along with this. Once attacked, a large amount of privacy information will be leaked and security problems will be triggered. Therefore, it is necessary to analyze and strengthen the security of loT devices. This paper proposed an loT devices identification method based on multilayer perceptron, which uses active detection to quickly find the loT devices in the network. Firstly, we use Nmap to obtain the opening ports and operating system information of IP addresses. Secondly, for devices which provide web service, we use WebDriver APls to obtain the web page title and URL, and take these data as the identification features. During the experiment, 2545 active IP addresses information in Peking University campus network are collected and tagged, 2000 as the training set and 545 as the test set. The F1 score is 0.94 from multilayer perceptron model, and the accuracy is 95.41%, precision is 95.31%. Our work will be the foundation for further identification of loT device types.
- 23 September 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 57, 69-73
Identification and Analysis of IoT Devices Based on Multilayer Perceptron
loT devices have provided great convenience for people's life and work, but hacking risks often tag along with this. Once attacked, a large amount of privacy information will be leaked and security problems will be triggered. Therefore, it is necessary to analyze and strengthen the security of loT devices. This paper proposed an loT devices identification method based on multilayer perceptron, which uses active detection to quickly find the loT devices in the network. Firstly, we use Nmap to obtain the opening ports and operating system information of IP addresses. Secondly, for devices which provide web service, we use WebDriver APls to obtain the web page title and URL, and take these data as the identification features. During the experiment, 2545 active IP addresses information in Peking University campus network are collected and tagged, 2000 as the training set and 545 as the test set. The F1 score is 0.94 from multilayer perceptron model, and the accuracy is 95.41%, precision is 95.31%. Our work will be the foundation for further identification of loT device types.