The aim of this project is to train Deep Learning (DL) models enabling them to forecast next day's closing price using past information/data and evaluate their performance. The following models were explored: Multi-Layer Perceptron Feed-Forward Neural Network (MLP FFNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) RNN, Gated Recurrent Unit (GRU) RNN. This work can be further extended by implementing an automated mechanism that will build and evaluate different DL architectures. This mechanism will then select the optimal model and incorporate it as part of our trading strategy.
The aim of this project is to develop trading Bots, using Reinforcement Learning (RL) algorithms, that will provide "Buy" and "Sell" signals for any given time series of stock data and evaluate their performance in reaching the predefined objective. The Agent's action space covers the "Hold, Buy, Sell" options and the algorithms utilized in this work are: Deep Q-Learning (with experience replay), Double Deep Q-Learning (with experience replay), Monte-Carlo Policy Gradient (REINFORCE). The results outline the importance of exploring additional RL algorithms, since their efficiency can improve existing strategies. One major limitation of this project is the required computational power.
-
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: an introduction. Cambridge: The MIT Press.
-
Atienza, R. (2018). Advanced deep learning with Keras: apply deep learning techniques, autoencoders, GANs, variational autoencoders, deep reinforcement learning, policy gradients,and more. Birmingham: Packt Publishing.
-
Jason Brownlee - Machine Learning Mastery
-
Github repositories/gists:
- Husein Zolkepli's (huseinzol05) Stock-Prediction-Models repository (huseinzol05/Stock-Prediction-Models)
- Edward Lu's (edwardhdlu) Q-Trader repository (edwardhdlu/q-trader)
- Mo Kweon's (kkweon) policy_gradient.py (kkweon/policy_gradient.py)