Complexity Analysis of Reinforcement Learning Models Applied to Stock Trading
Details
- Authors: Erich Schwarzrock, Jason Davis, Hezi Owuor
- Title: Complexity Analysis of Reinforcement Learning Models Applied to Stock Trading
- Supervisior: Prof. Dr. Jörg Osterrieder
- Degree: Bachelor of Science
- University: Zurich University of Applied Sciences
- Year: 2022
- Status: Working Paper
Summary
Empirical data analysis of the performance of Double-Deep Q-Learning models of varying complexity for stock trading amongst Forex and Equity Index markets.
Abstract
In this project, we analyze the effect of data complexity on the performance of financial reinforcement learning models. We created six models which were identical except for the complexity of their learning data. The goal for each was to make as much money as possible by investing in only a single stock. We trained these models on daily index fund data, intraday index fund data, and daily foreign exchange data. We then analyzed the effect that the different data complexities had on both the training and testing returns. Simpler models cannot learn anything and will perform poorly, while if a model is too complex, the agents will overfit the training data and perform poorly on testing data. State spaces with moderate complexity tend to perform the best.
Important links
Data
- Daily Closing Price (SPY)
- Daily Closing Price (NDAQ.O)
- Daily Closing Price (DIA)
- Daily Closing Price (USO)
- Daily Closing Price (GLD)
- 5 Minute Price (SPY)
- 5 Minute Price (NDAQ.O)
- 5 Minute Price (DIA)
- 5 Minute Price (USO)
- 5 Minute Price (GLD)
- Daily Closing Price (GBPUSD)
- Daily Closing Price (EURUSD)
- Daily Closing Price (USDCHF)
- Daily Closing Price (USDJPY)
- Daily Closing Price (NZDCAD)
Data source: Refinitiv