[ML / Investment] Turnaround Company Prediction

Project Overview

This project aims to classify and predict whether a company classified as a "potential distressed company" (with an interest coverage ratio below 1 for two consecutive years) will either remain distressed or recover in the following year (t+1). To achieve this, nine optimal features were selected, and 11 classification and deep learning algorithms were tested. The XGB Classifier, which demonstrated the best performance, was chosen as the final model. Using the final model, 15 investment strategies were validated by combining three investment weightings and five trading strategies for companies predicted to recover. This approach allows for the early identification of distressed companies exacerbated by COVID-19, enabling proactive measures such as corporate loans. Additionally, the investment strategies can be applied to recovering companies during an upcoming economic recovery to generate returns.

0. Data Collection

• Source: KRX & TS200

• Collection Period: 2011–2019 (post-IFRS adoption, pre-COVID-19)

• Target Companies: Listed on KOSPI & KOSDAQ

• Excluded Companies: Financial and insurance firms, SPACs, delisted companies, foreign-owned companies

• Selected Features: 80 total (financial statement data, derived variables, and stock price data)

  
  

1. Preprocessing

• Removed duplicates of companies with two exchange codes (26 companies).

• Verified duplicates of companies with two names under the same exchange code.

• Deleted companies with only one data entry.

• Replaced consolidated financial statements with separate financial statements.

• Removed 999999, inf, and -inf values (except for interest coverage ratio).

• Winsorized inf values in the interest coverage ratio.

• Winsorized 2.5% of extreme values (excluding object features).

• Handled outliers in derived variables (growth rates):

  • Created growth rate columns if previous and current year values existed.

  • Set growth rate = 0 if both previous and current year values were 0.

  • Dropped rows if the previous year value was 0 and the current year value was positive.

  • Dropped columns generating excessive inf values.

• Filtered for potential distressed companies (interest coverage ratio < 1 for two consecutive years).

  • Only the earliest data point for each company was retained.

    
    

2. Labeling

• Label 1: Distressed companies (interest coverage ratio < 1 in year t+1, three consecutive years).

• Label 0: Recovering companies (interest coverage ratio ≥ 1 in year t+1, i.e., companies that normalized after being potential distressed).

  
  

3. Data Split & Scaling

• Train Set: 2012–2016

• Test Set: 2017–2018

• Scaling: Standard scaling applied separately to train and test sets to prevent data leakage.

  
  

4. Feature Selection

• Feature selection was performed only on the train set.

• Tests Conducted:

  • Normality tests (Shapiro, Anderson, KS, Jarque-Bera, Normal, Dist plot, QQ plot).

  • Independence test (Durbin-Watson).

  • Homoscedasticity test (Bartlett).

• Final features were selected based on overlap across at least three of the following models:

  • T-test (statistical analysis).

  • Lasso (Wrapper Method).

  • Stepwise (Embedded Method).

  • Kbest (Filter Method).

    
    

5. Modeling

Classification Models

• Logistic Regression

• KNN

• Decision Tree

• Naive Bayes

• SVM

Ensemble Models

• Random Forest

• XGBoost

• LightGBM

• Voting

• Stacking

Final Model: XGBClassifier

6. Evaluation (Test Set)

• Accuracy: 0.73

• Precision: 0.73

• Recall: 0.97

• F1-Score: 0.83

• AUC: 0.59

  
  

Key Insights

• The model exhibits relatively high recall, minimizing the risk of misclassifying distressed companies as recovering.

• From the perspective of recovering companies (Label 0), the model confidently identifies recovering companies, making it suitable for investment applications.