An Artificial Neural Network Approach to Algorithmic Trading
(2023) In Master's Theses in Mathematical Sciences FMSM01 20222Mathematical Statistics
- Abstract
- The field of machine learning has advanced significantly in recent decades, and, at the same time, computational power has improved to the point where training large machine learning models, such as artificial neural networks, is now accessible. Consequently, there has been a rise in the use of these models within the financial sector, with some firms leveraging them to assist with investment decisions. Using neural networks, or machine learning models in general, for investing refers to investment strategies that are constructed, at least partially, by training algorithms on historical data to identify patterns that may recur in the future. The rationale behind this approach is that historical data contains structure that will be repeated... (More)
- The field of machine learning has advanced significantly in recent decades, and, at the same time, computational power has improved to the point where training large machine learning models, such as artificial neural networks, is now accessible. Consequently, there has been a rise in the use of these models within the financial sector, with some firms leveraging them to assist with investment decisions. Using neural networks, or machine learning models in general, for investing refers to investment strategies that are constructed, at least partially, by training algorithms on historical data to identify patterns that may recur in the future. The rationale behind this approach is that historical data contains structure that will be repeated in the future, meaning that past price developments of an asset hold valuable information for predicting future price developments. This approach challenges traditional market theories, such as the efficient market hypothesis. Despite this, price patterns have proved reliable enough to allow multiple investors to reap substantial financial gains and development into organizations with headcounts in the thousands, solely focused on identifying and trading these pricing patterns.
This thesis aims to construct and assess artificial neural network models intended for use in trading algorithms. Given historical returns, the models are trained to forecast the direction of asset price returns the following day. The predicted return directions are then input into a trading algorithm that computes daily portfolio values. The performance of these models is then benchmarked against naive trading strategies and the underlying asset itself. The Sharpe ratio, Sortino ratio, gross return, and maximal drawdown are benchmark values used to assess the models. The selected underlying assets are government bond contracts, as requested by Handelsbanken Fonder.
The results do not demonstrate significant improvement in return estimation over the simple benchmark models. However, among the tested models, the Trend model outperformed the others with an average Sharpe ratio of 1.28, Sortino ratio of 2.21, and gross returns of 15.04%. This suggests that incorporating trend analysis may provide some value in predicting returns, although further research is needed to confirm its effectiveness in different market conditions.
This thesis was written in collaboration with Handelsbanken Fonder as the concluding part of a master’s degree in engineering at Lund University. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9113439
- author
- Bengtsson, Timmie LU
- supervisor
-
- Erik Lindström LU
- Carl Hvarfner LU
- organization
- course
- FMSM01 20222
- year
- 2023
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Financial Markets, Machine Learning, Long Short-Term Memory, Gated Recurrent Unit, Recurrent Neural Networks, Time Series Analysis, Algorithmic Trading
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMS-3467-2023
- ISSN
- 1404-6342
- other publication id
- 2023:E9
- language
- English
- id
- 9113439
- alternative location
- https://github.com/TimmieBengtsson/masters_project
- date added to LUP
- 2023-04-18 09:41:22
- date last changed
- 2023-04-27 15:57:15
@misc{9113439,
abstract = {{The field of machine learning has advanced significantly in recent decades, and, at the same time, computational power has improved to the point where training large machine learning models, such as artificial neural networks, is now accessible. Consequently, there has been a rise in the use of these models within the financial sector, with some firms leveraging them to assist with investment decisions. Using neural networks, or machine learning models in general, for investing refers to investment strategies that are constructed, at least partially, by training algorithms on historical data to identify patterns that may recur in the future. The rationale behind this approach is that historical data contains structure that will be repeated in the future, meaning that past price developments of an asset hold valuable information for predicting future price developments. This approach challenges traditional market theories, such as the efficient market hypothesis. Despite this, price patterns have proved reliable enough to allow multiple investors to reap substantial financial gains and development into organizations with headcounts in the thousands, solely focused on identifying and trading these pricing patterns.
This thesis aims to construct and assess artificial neural network models intended for use in trading algorithms. Given historical returns, the models are trained to forecast the direction of asset price returns the following day. The predicted return directions are then input into a trading algorithm that computes daily portfolio values. The performance of these models is then benchmarked against naive trading strategies and the underlying asset itself. The Sharpe ratio, Sortino ratio, gross return, and maximal drawdown are benchmark values used to assess the models. The selected underlying assets are government bond contracts, as requested by Handelsbanken Fonder.
The results do not demonstrate significant improvement in return estimation over the simple benchmark models. However, among the tested models, the Trend model outperformed the others with an average Sharpe ratio of 1.28, Sortino ratio of 2.21, and gross returns of 15.04%. This suggests that incorporating trend analysis may provide some value in predicting returns, although further research is needed to confirm its effectiveness in different market conditions.
This thesis was written in collaboration with Handelsbanken Fonder as the concluding part of a master’s degree in engineering at Lund University.}},
author = {{Bengtsson, Timmie}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{An Artificial Neural Network Approach to Algorithmic Trading}},
year = {{2023}},
}