Interpretable Machine Learning dengan Model Agnostik pada Prediksi Financial Distress menggunakan Metode Backpropagation Neural Network

Nurfadhila, Aisyah Azka (2024) Interpretable Machine Learning dengan Model Agnostik pada Prediksi Financial Distress menggunakan Metode Backpropagation Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5006201066-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (4MB) | Request a copy

Abstract

Model machine learning sering disebut sebagai model Black-Box karena sulit diinterpretasikan. Namun, interpretasi model sangat penting untuk keyakinan dan penerimaan hasil prediksi, termasuk di sektor industri manufaktur. Salah satunya adalah dalam memprediksi kondisi financial distress yaitu kondisi dimana perusahaan tidak mampu melunasi kewajibannya dan merupakan tahapan sebelum kebangkrutan. Pihak Bursa Efek Indonesia (BEI) pada laman Market Bisnis melaporkan bahwa terdapat 9 perusahaan yang berisiko mengalami delisting pada tahun 2023, dengan 3 perusahaan bergerak di sektor industri manufaktur, sehingga klasifikasi financial distress dengan rasio keuangan menjadi penting. Rasio keuangan yang digunakan adalah rasio keuangan pada model Ohlson. Penelitian ini menggunakan data laporan keuangan perusahaan sektor industri manufaktur yang terdaftar di BEI mulai kuartal I 2017 hingga kuartal III 2023, dengan status awal kondisi keuangan perusahaan berdasarkan Earnings Per Share (EPS). Penelitian ini juga dilakukan penanganan outlier dan multikolinieritas karena model agnostik mengasumsikan bahwa adanya hubungan antar variabel prediktor dapat menyebabkan bias saat interpretasi. Hasil penelitian menunjukkan tidak terdapat perbedaan antara model BPNN tanpa penanganan dan dengan penanganan multikolinieritas. Penerapan winsorization 1% dan 2,5% juga tidak terdapat perbedaaan terhadap nilai evaluasi yang dihasilkan. Nilai evaluasi yang dihasilkan adalah sama, yaitu AUC 0,9476, Akurasi 0,9592, dan F1-Score 0,9722. Arsitektur model BPNN terbaik tanpa penanganan multikolinieritas adalah BPNN (9-1-1), sedangkan model dengan penanganan multikolinieritas adalah BPNN (8-1-1). Interpretasi global dengan PDP, Feature Interaction, dan Permutation Feature Importance menghasilkan bahwa variabel TLTA, OENEG, NITA, dan INTWO membentuk hubungan positif terhadap financial distress, sedangkan variabel SIZE, CLCA, dan CFOTL membentuk hubungan negatif. Variabel CHIN memiliki interaksi tertinggi dan variabel INTWO adalah variabel terpenting dalam memprediksi financial distress. Intrepretasi lokal dengan LIME dan Shapley Value menghasilkan nilai kontribusi negatif dan positif terhadap financial distress di setiap variabel
======================================================================================================================================
Machine learning models are often referred to as Black-Box models because they are difficult to interpret. However, model interpretation is critical for confidence and acceptance of predictive results, including in the manufacturing industry sector. One of them is in predicting financial distress conditions, which is a condition where the company is unable to pay off its obligations and is a stage before bankruptcy. The Indonesia Stock Exchange (IDX) on the Business Market page reports that there are 9 companies at risk of delisting in 2023, with 3 companies engaged in the manufacturing industry sector, so the classification of financial distress with financial ratios is important. The financial ratios used are the financial ratios in the Ohlson model. This study uses data on the financial statements of manufacturing industry sector companies listed on the IDX from the first quarter of 2017 to the third quarter of 2023, with the initial status of the company's financial condition based on Earnings Per Share (EPS). This research also handles outliers and multicollinearity because the agnostic model assumes that the existence of a relationship between predictor variables can cause bias during interpretation. The results showed no difference between the BPNN model without handling and with multicollinearity handling. The application of 1% and 2,5% winsorization also made no difference to the resulting evaluation value. The resulting evaluation values are the same, namely AUC 0,9476, Accuracy 0,9592, and F1-Score 0,9722. The best BPNN model architecture without multicollinearity handling is BPNN (9-1-1), while the model with multicollinearity handling is BPNN (8-1-1). Global interpretation with PDP, Feature Interaction, and Permutation Feature Importance results in that TLTA, OENEG, NITA, and INTWO variables form a positive relationship to financial distress, while SIZE, CLCA, and CFOTL variables form a negative relationship. The CHIN variable has the highest interaction and the INTWO variable is the most important variable in predicting financial distress. Local interpretation with LIME and Shapley Value produces negative and positive contribution values to financial distress in each variable.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Backpropagation Neural Network, Financial Distress, Interpretable Machine Learning, Model Agnostik, Ohlson, Agnostic Model, Backpropagation Neural Network, Financial Distress, Interpretable Machine Learning, Ohlson.
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Actuaria > 94203-(S1) Undergraduate Thesis
Depositing User:	Aisyah Azka Nurfadhila
Date Deposited:	11 Jul 2024 05:33
Last Modified:	26 Aug 2024 06:24
URI:	http://repository.its.ac.id/id/eprint/108247

Actions (login required)

View Item