Prediction of company bankruptcy. Learn about how Machine Learning finds insights of the Czech Business Landscape, presented by Lucie Beranová, Ph.D. Student at Prague University of Economics and Business (VSE) and Data Scientist at Vodafone.
*Machine Learning School for Business Schools 2021: Virtual Conference.
2. #BigMLSchool
Introduction
• Financial condition of a company is important for the
company, investors, suppliers, customers and many
others.
• Easy to find if a company is in an insolvency
proceeding, but not if it will enter insolvency in the
future.
• Goal of this work: Predict with a machine learning
model the probability of a company getting into
bankruptcy in 3 years and find out how accurate this
model can be. -> Classification problem.
Prediction of company bankruptcy 2
3. #BigMLSchool
Data source
• Albertina database
– Collects information from public registers about Czech
companies.
– Simple filter for active versus bankrupt companies
– Financial indicators, list of documents from the registers,
data about owners, registered office etc.
• Used random sample of companies with complete data
for 1 year:
– active companies – year 2019
– bankrupt companies – year signifying 3 years before
bankruptcy
Prediction of company bankruptcy 3
5. #BigMLSchool
Prediction of company bankruptcy 5
• Uploaded data
from csv to BIGML
• Marked not
prefered variables
like id
Data in BIGML
• District
• Region
• Capital
• Market segment
• Market segment 2
• Market segment 3
• Equity
Plant capacity
• Total revenues
• Total liabilities
• Education
• Proportion of women
• TARGET
9. #BigMLSchool
• Can help find the
best supervised
learning model for
our data
• Optimization metric:
ROC AUC = metrics
for checking any
classification
model’s
performance –
higher = better.
Prediction of company bankruptcy 9
OPTIML
10. #BigMLSchool
Results OPTIML:
• Training time: 2h 45m
• Evaluated 56 models:
– 1 decision tree
– 52 ensembles
models
– 1 logistic
regression
– 2 neural networks
• Model with the
highest AUC ROC:
bootstrap decision
forest.
Prediction of company bankruptcy
10
13. #BigMLSchool
• TEXT
• QUICKTEST
• MIRAZADLU
• CIZIZDROJE
• OKAMLIKVID
• ZAVTOTAL
• OBRAT
• RENTTRZEB
• CASHFLOW
• OATOTAL
Feature importance
Prediction of company bankruptcy 13
Debt ratio
Current liabilities
Cash position ratio
Liabilities total
Annual sales
Profitability of sales
Current assets
14. #BigMLSchool
Decision tree
• Interpretable results, hierarchical rules, but is not robust
Prediction of company bankruptcy 14
Annual sales
TEXT – contains „notarial deed“
TEXT – doesn‘t contain „water supply“
15. #BigMLSchool
Examples of „notarial deed“
Prediction of company bankruptcy
proposal for dismissal
of the managing director
proposal for dismissal of the
managing director
decision to amend the contract
Very often a connection with execution can be seen.
16. #BigMLSchool
Conclusion
• Original model was programming in Python and R -> used for
my thesis :
• BERANOVÁ, Lucie. Prediction of company insolvency using
data science methods. Praha, 2020. SupervisorTomáš Kliegr.
• Results of original approach were very similar.
Prediction of company bankruptcy 16