Difference between revisions of "AI for Trading Strategies"
(Created page with "== Details == * Authors: Romain Délèze & Danijel Jevtic * Title: AI for Trading Strategies * Supervisior: Prof. Dr. Jörg Osterrieder * Degree: Bachelor of Science * Univer...") |
|||
| (4 intermediate revisions by the same user not shown) | |||
| Line 14: | Line 14: | ||
== Abstract == | == Abstract == | ||
In recent years, much research has been done in predicting stock price trends using machine learning algorithms. However, machine learning methods are often criticized by financial practitioners. They argue that neural networks are a "black box." In this bachelor thesis, we show how four different machine learning methods (Long Short-Term Memory, Random Forest, Support Vector Machine Regression, and k-Nearest Neighbor) perform compared to already successfully applied trading strategies such as Cross Signal Trading and a conventional statistical time series model ARMA-GARCH. The aim is to show that machine learning methods perform better than conventional methods in the crude oil market when used correctly. A more detailed performance analysis was made, showing the performance of the different models in different market phases so that the robustness of individual models in high and low volatility phases could be examined more closely. The focus was on the robustness and performance of the machine learning methods. It was essential to keep complexity in proportion to performance. All analysis work was realized with the software Python. It was shown that the machine learning methods such as Random Forest, on average, made better decisions in highly volatile market periods than ordinary statistical models such as ARMA-GARCH. The Random Forest was characterized by a very high true-positive rate (recall), which set it apart from the other models. Likewise, the final performance result provided conclusions about the robustness of the individual models. It could be shown in which market phases the models made good or bad trading decisions. In the case of the k-Nearest Neighbor model, which failed to detect the Corona crisis, it was shown that not every machine learning model responded well to different market phases. Comparing this with the Random Forest model, it can be seen that the most significant monthly gain was achieved there across all models. | In recent years, much research has been done in predicting stock price trends using machine learning algorithms. However, machine learning methods are often criticized by financial practitioners. They argue that neural networks are a "black box." In this bachelor thesis, we show how four different machine learning methods (Long Short-Term Memory, Random Forest, Support Vector Machine Regression, and k-Nearest Neighbor) perform compared to already successfully applied trading strategies such as Cross Signal Trading and a conventional statistical time series model ARMA-GARCH. The aim is to show that machine learning methods perform better than conventional methods in the crude oil market when used correctly. A more detailed performance analysis was made, showing the performance of the different models in different market phases so that the robustness of individual models in high and low volatility phases could be examined more closely. The focus was on the robustness and performance of the machine learning methods. It was essential to keep complexity in proportion to performance. All analysis work was realized with the software Python. It was shown that the machine learning methods such as Random Forest, on average, made better decisions in highly volatile market periods than ordinary statistical models such as ARMA-GARCH. The Random Forest was characterized by a very high true-positive rate (recall), which set it apart from the other models. Likewise, the final performance result provided conclusions about the robustness of the individual models. It could be shown in which market phases the models made good or bad trading decisions. In the case of the k-Nearest Neighbor model, which failed to detect the Corona crisis, it was shown that not every machine learning model responded well to different market phases. Comparing this with the Random Forest model, it can be seen that the most significant monthly gain was achieved there across all models. | ||
| − | The conventional statistical models - ARMA-GARCH - performed surprisingly well and were on par with the machine learning models. Nevertheless, the extreme values were analyzed in more detail. It was found that the Random Forest model made the best decisions at the 95 | + | The conventional statistical models - ARMA-GARCH - performed surprisingly well and were on par with the machine learning models. Nevertheless, the extreme values were analyzed in more detail. It was found that the Random Forest model made the best decisions at the 95% interval level. This shows that the Random Forest model always made better decisions than all other models during periods when a lot is going on in markets, which is reflected in the total return. Surprisingly, the ARMA-GARCH model made the best decisions at the 99% level but performed worse on average when the market was "calmer." It was shown that better results could be achieved with proper machine learning methods than with conventional statistical methods. For further investigation, these models would also have to be analyzed in other markets. |
== Important links == | == Important links == | ||
* [https://www.overleaf.com/read/tggxxqfjfdsb Overleaf] | * [https://www.overleaf.com/read/tggxxqfjfdsb Overleaf] | ||
| − | + | * [https://github.com/delezrom/BA_oest_03/upload/main GitHub] | |
| + | * [https://drive.google.com/drive/folders/1zlmBS7o1R80MCjh5M4KU73Gl73UiLyiI Google drive] | ||
| + | * [Arxiv] | ||
== Data == | == Data == | ||
Latest revision as of 14:33, 26 June 2022
Details
- Authors: Romain Délèze & Danijel Jevtic
- Title: AI for Trading Strategies
- Supervisior: Prof. Dr. Jörg Osterrieder
- Degree: Bachelor of Science
- University: Zurich University of Applied Sciences
- Year: 2022
- Status: Working Paper
Summary
Investigation of different machine learning algorithms on the crude oil market.
Abstract
In recent years, much research has been done in predicting stock price trends using machine learning algorithms. However, machine learning methods are often criticized by financial practitioners. They argue that neural networks are a "black box." In this bachelor thesis, we show how four different machine learning methods (Long Short-Term Memory, Random Forest, Support Vector Machine Regression, and k-Nearest Neighbor) perform compared to already successfully applied trading strategies such as Cross Signal Trading and a conventional statistical time series model ARMA-GARCH. The aim is to show that machine learning methods perform better than conventional methods in the crude oil market when used correctly. A more detailed performance analysis was made, showing the performance of the different models in different market phases so that the robustness of individual models in high and low volatility phases could be examined more closely. The focus was on the robustness and performance of the machine learning methods. It was essential to keep complexity in proportion to performance. All analysis work was realized with the software Python. It was shown that the machine learning methods such as Random Forest, on average, made better decisions in highly volatile market periods than ordinary statistical models such as ARMA-GARCH. The Random Forest was characterized by a very high true-positive rate (recall), which set it apart from the other models. Likewise, the final performance result provided conclusions about the robustness of the individual models. It could be shown in which market phases the models made good or bad trading decisions. In the case of the k-Nearest Neighbor model, which failed to detect the Corona crisis, it was shown that not every machine learning model responded well to different market phases. Comparing this with the Random Forest model, it can be seen that the most significant monthly gain was achieved there across all models. The conventional statistical models - ARMA-GARCH - performed surprisingly well and were on par with the machine learning models. Nevertheless, the extreme values were analyzed in more detail. It was found that the Random Forest model made the best decisions at the 95% interval level. This shows that the Random Forest model always made better decisions than all other models during periods when a lot is going on in markets, which is reflected in the total return. Surprisingly, the ARMA-GARCH model made the best decisions at the 99% level but performed worse on average when the market was "calmer." It was shown that better results could be achieved with proper machine learning methods than with conventional statistical methods. For further investigation, these models would also have to be analyzed in other markets.
Important links
- Overleaf
- GitHub
- Google drive
- [Arxiv]
Data
- Brent Crude Oil futurs
Data source: Yahoo Finance