Fuzzy-embedded long short-term memory (FE-LSTM) with application in stock trading
Deep learning has been increasing in popularity in recent years due to its high accuracy and effectiveness in many applications. However, a major drawback of deep learning systems is the lack of interpretability as it functions like a black box where the prediction results are often unexplainable ev...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2022
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/156487 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | Deep learning has been increasing in popularity in recent years due to its high accuracy and effectiveness in many applications. However, a major drawback of deep learning systems is the lack of interpretability as it functions like a black box where the prediction results are often unexplainable even to experts. On the other hand, traditional modelling techniques such as the fuzzy inference system have interpretable results but suffer from low accuracy due to its limited learning capabilities.
This paper proposes the Fuzzy Embedded Long Short-Term Memory (FE-LSTM) architecture which integrates a fuzzy inference system with a deep neural network to leverage on the strengths of both systems. The proposed hybrid architecture aims to achieve high prediction accuracies with interpretable results. This is done by first applying a discrete incremental clustering (DIC) algorithm to fuzzify the data inputs and then simultaneously feeding the fuzzified inputs into the parallel long short-term memory network and the fuzzy inference system. Here, the LSTM network uses back-propagation to learn from the data while the fuzzy inference system uses a pseudo-outer product rule generation to interpret the LSTM network which consists of human explainable IF-THEN rules. Finally, the output is defuzzified to obtain a crisp value.
The implemented FE-LSTM is benchmarked against common neural networks and fuzzy inference systems through several time-series benchmark experiments and the prediction of stock prices. The prediction results obtained are highly encouraging with the FE-LSTM outperforming other systems in terms of accuracy. Next, the FE-LSTM is used as a predictor in a technical trading system that uses a look forward MACD indicator to maximize trading profits. This trading system shows promising results when benchmarked against both the buy-hold strategy and the vanilla MACD strategy. |
|---|