Evolving type-2 neural fuzzy inference system with embedded deep learning in dynamic portfolio rebalancing
This paper examines the benefits of integrating neuro-fuzzy system and deep learning architecture for making predictions in a noisy environment with dynamically changing data, and its feasibility in financial market applications. Previous research has been carried out to implement deep neural netw...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2021
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/148085 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | This paper examines the benefits of integrating neuro-fuzzy system and deep learning architecture for
making predictions in a noisy environment with dynamically changing data, and its feasibility in financial
market applications. Previous research has been carried out to implement deep neural networks to
extract features for the fuzzy system [1], but this combination does not make the learning model
entirely interpretable.
To benefit both from the interpretability property of the fuzzy system and the performance of deep
learning, the proposed Evolving Type-2 Neural Fuzzy Inference System with Embedded Deep Learning
model (eT2FIS-EDL) employs an embedded deep learning component parallel with the rule generation.
This proposed method resolves the black-box nature of deep learning architecture while maintaining
computation efficiency during the inference process. The eT2FIS-EDL is implemented in two variations,
one with convolutional neural network (CNN), and the other with long short-term memory (LSTM).
Type-2 fuzzy model is chosen in order to increase the model’s tolerance with noisy data. Learning
mechanism features including the modification, merger and deletion of rules are implemented to help
maintain the interpretability of the model, as well as to ensure its adaptation to shifts and drifts in data.
The eT2FIS-EDL demonstrated its effectiveness through the prediction tasks on a nonlinear system
dataset, as well as on real-life stock and exchange-traded fund (ETF) datasets. The model’s usefulness is
further illustrated through the implementation of dynamic portfolio rebalancing strategy. This strategy
combines reinforcement learning with eT2FIS-EDL to rebalance the portfolio with the considerations of
market trends, risks, and returns. Both CNN and LSTM variants are integrated into the learning model to
reduce the information lag. The experimental results are highly encouraging as the proposed
rebalancing strategy outperforms the other four existing portfolio management strategies, namely buy
and hold, equal weight (periodic) rebalancing, inverse variance rebalancing and max Sharpe ratio
rebalancing. |
---|