A.I. Generated Approach
End-to-End Deep Reinforcement Learning
Applying Deep RL to control tasks such as Atari Games and Go has achieved groundbreaking success in the recent years. The difference between Deep RL agents and traditional expert agents with handcrafted features is that the former produces outputs from inputs data completely on its own, free of programming. We have changed from "if a do b" to optimization a total reward
regardless of the action. Our agent is freely exploring different trading strategies to find the best trades that human may neglect.
Machine Vision and Technical Analysis
Deep Convolution Neural Network (CNN) model has surpassed human performance in ImageNet visual recognition challenge. We are applying the most up-to-date CNN model to Forex price data. With millions pf parameters in the model, the network is able to pick up features that human traders have ignored.
Systematic Approach Deprived of Human Emotion
It's a well known fact that human can be trained to be great traders in a simulation setting, but fail terribly when trading with real money. Emotions and doubts are evolution's gift to human kind, but only those who are deprived of it can execute risk management calmly and reasonably in a consistent manner. Such traits are by design the most capable for computers. Now coupled with non-linearity provided by deep neural network, the machine can do better what we cannot do.
GPU Powered Trading Execution
Our model is trained, tested and run on the most cutting edge GPU hardware. Each agent is catered with one GPU dedicating to trading one security at one time frame. Specialization in computation is what we strive for in our company. We are seeking partnership with NVIDIA to build our in-house GPU cloud computing server, enabling the most efficient and stable live A.I. application.
Scalability and Flexibility in Model Design
Our models are designed from the ground up with open source Machine Learning libraries such as Keras, Tensorflow and OpenAI gym. They are easily scalable to trading stocks, futures and other financial securities. We have a active model training platform that can plug in different format of data, assign multiple trading time scales, and optimize customized reward function in regard to risk measurement metrics.