Final Year Project (FYP) : Application of Fuzzy Neural Networks (FNNs) for bank failure prediction and missing data reconstruction
I have learnt how a Fuzzy Neural Network (FNN) works as well as how aritifical neural networks and fuzzy logic systems complement each other in FNNs. In this FYP, I have applied a Mamdani-type FNN, namely "SaFIN(FRI/E)++", to bank failure prediction. To further improve its prediction accuracy, I have extended the original model with a novel hierarchical feature selection. I have also invented a missing data reconstruction algorithm in order to rebuild missing values in the given data set. The extended prediction model combined with the proposed missing data reconstruction method, can serve as a powerful hybrid tool for bank failure prediction.
[1] Extended a deep five-layer fuzzy neural, SAFIN(FRIE)++, with hierarchical feature selection to improve its prediction accuracy
[2] Analyzed the data set that includes financial statements of banks in the United States over the course of 20 years
[3] Performed online learning to train the model and used 5-fold Cross Validation (CV) for training and testing in bank failure prediction
[4] Achieved over 93% accuracy with the new extended model on an imbalanced data set with very low type I and type II error
[5] Implemented a missing data reconstruction algorithm based on a combination of longitudinal and lateral
reconstruction by using DENFIS, i.e. known for its regression accuracy
[6] Successfully reconstructed missing data with relatively low mean absolute error (MAE) and decent correlation accuracy (R)
[7] Tested the quality of the reconstructed data by predicting bank failure using the extended model
[8] Achieved about 1-2% better accuracy with the reconstructed data set than with the original data set for one-year ahead prediction
To find out more details about:
SAFIN(FRIE)++: https://repository.ntu.edu.sg/handle/10356/72908
HCL: https://ieeexplore.ieee.org/abstract/document/4359859?section=abstract
DENFIS: https://ieeexplore.ieee.org/document/995117