Juan Jose Suarez

1

A NEURAL NETWORK MODEL TO PREDICT BUSINESS
FAILURE IN CONSTRUCTION COMPANIES IN THE UNITED
STATES OF AMERICA

• Juan Jose Suarez
• Construction Engineering and Management
• Department of Civil and Coastal Engineering
• November 16, 2004

2
Agenda

• Problem Statement
• Literature Review
• Objectives
• Neural Network Models
• Conclusions
• Recommendations
• Questions

3
Problem Statement

• The construction industry has one of the highest
  rates of bankruptcy
• 2001 40,099 cases filed
• 2002 38,540 businesses declared bankruptcy

4
Problem Statement (cont.)

• (2003) 8 of the 30 companies with the highest
  percentages of bankruptcy assets were related to
  the construction industry (Public companies)
• Construction companies are normally private
• In 2001, the construction industry reported 1.20
  of the total amount of bankruptcies (481 cases)

5
Problem Statement (cont.)
6
Problem Statement (cont.)
7
Literature Review
Bankruptcy Models
Mathematical Models
Neural Networks Models
Statistical Prediction Models
8
Statistical Prediction Models
Statistical Prediction Models
Beaver 1967
Altman 1968
Serrano and Molinero 2000
9
Univariate Analysis

• Beaver (1967)
• Financial ratios to discriminate between failed
  and non-failed firms
• Predict bankruptcy five years before
• Most important factor Cash flow/total debt ratio

10
Altmans Z-Score Model

• Altman (1968)
• Z0.012 X1 0.014 X2 0.033 X3 0.006 X4
  0.999 X5
• X1 working capital/total assets
• X2 retained earnings/total assets
• X3 earnings before interest and taxes/total
  assets
• X4 market value of equity/book value of total
  liabilities
• X5 sales/total assets
• Sixty-six companies (50 healthy companies - 50
  unhealthy companies)
• Companies manufacturing and retail industry

11
Linear Discriminant Analysis (LDA)

• Serrano and Molinero (2000)
• Same ratios that Altman
• Conclusion using the combination of a linear
  discriminant analysis and neural networks they
  could predict the two areas

12
Mathematical Prediction Models

• Feller (1968). Mathematical models the
  gamblers ruin approach
• Company will fall into bankruptcy when its net
  liquidation value becomes negative
• Net Liquidation Value (Total Asset Liquidation
  Total Liabilities)
• There are no mathematical models completely
  accepted to predict bankruptcy

13
Neural Networks Models
14
Neural Networks Prediction Models
Neural Network Models
Feed-Forward Back-Propagation Model
Back Propagation Algorithm
Probabilistic Neural Networks without Patterns
Normalized
Probabilistic Neural Networks
Three-Perceptron Network
Genetic Algorithm
15
Back Propagation Algorithm

• Dorota Witkowskah(1999)
• Neural network models to help banks to identify
  business failures
• 75 companies
• 13 testing sample
• 62 training data

16
Back Propagation Algorithm (cont)

• Less accurate when less than 10,000 iterations
  are performed

17
Probabilistic Neural Networks

• Zheng Rong Yang at the University of Exeter, UK
• 2,408 companies (1989 to 1995), 33 financial
  ratios
• Accuracy 95.5 percent predicting company
  survival, and 92.37 percent predicting company
  failure

18
Genetic Algorithm

• Xiaotong Li and Jatinder N. D. Gupta at the
  University of Alabama in Huntsville
• Two data sets gathered by Altman, Frydman and Kao
  in 1985
• First data set 200 companies - four financial
  ratios
• Second data set 200 companies - six financial
  ratios
• Companies Manufacturing and retail industry

19
Feed-Forward Back-Propagation Model

• Gregory Golinski at New York University (1998)
• (5-5-1) five ratios suggested by Altman (1981 to
  1997)
• 104 companies (50 failed companies 50
  healthy companies)
• Results 96 percent prediction rate (Predicting
  bankrupt companies in the first year)
• Companies Manufacturing and retail industry

20
Probabilistic Neural Networks without Patterns
Normalized

• Z. R. Yang at University of Portsmouth, and
  Marjorie B. Platt and Harlan D. Platt at
  Northeastern University (1998)
• 38 companies
• 30 successful companies
• 8 companies that failed
• Results 100 percent accuracy non-bankruptcy
  firms
• Companies United States gas and oil industry

21
Three-Perceptron Network

• Marcus D. Odom, and Ramesh Sharda (1998)
• 129 companies (1975 to 1982)
• 64 healthy companies
• 65 bankrupt companies
• Input variables those recommended by Altman in
  1968
• (5-5-1) 81.48 percent accuracy (bankrupt
  companies)

22
Three-Perceptron Network (cont)
23
Objectives

• Gather financial information from construction
  companies (public) that either had fallen into
  bankruptcy in the United States or are currently
  in business
• Identify financial variables
• Quantify the impact of those variables
• Identify and train a neural network algorithm
• Create a neural network model

24
Type of Data Collected

• Financial information from bankrupt construction
  companies (Chapter Seven or Chapter Eleven), and
  healthy companies
• 50 came from healthy companies, and 50 from
  bankrupt companies
• Name of the company
• Scope of work
• Yearly financial statements
• Information about the owner(s) and employee(s)
• Geographic dispersion
• Amount of work performed by the company
• Growth rate

25
How Data Was Collected
26
Software Used

• Statistical Data analysis R
• This software met the three following
  characteristics
• Provide graphic results
• Relatively easy to use
• Have the necessary statistical tools
• Neural Networks - NeuroSolutions 4.2

27
Data Analysis

• Twenty-six financial ratios from 67companies (34
  Healthy and 33 Unhealthy)
• Boxplots 26 financial ratios (After data were
  standardized)

28
Data Analysis (cont.)

• Cross correlation matrix graphics (26 financial
  ratios)

29
Data Analysis (cont.)

• Seven financial ratios as inputs in the neural
  network model
• Current Ratio
• Gross Profit Margin Ratio
• Debt-to-Assets Ratio
• Debt-to-Equity Ratio
• Account Receivables-to-Turnover Ratio
• Total Assets-to-Turnover Ratio
• Equity-to-Debt Ratio

30
Data Analysis (cont.)
31
Mistakes In Bankruptcy Prediction

• Type I error occurs when the prediction model
  classifies a company that filed bankruptcy as a
  healthy company
• Type I errors are usually more costly for model
  users

• Type II error occurs when the model classifies a
  healthy company as a failed company

32
Bankruptcy Prediction Model (1 yr in advance)

• Healthy construction companies 88.89 percent of
  the time
• Unhealthy companies 100 percent of the time

33
Bankruptcy Prediction Model (2 yrs in advance)

• Healthy construction companies 90 percent of
  the time
• Unhealthy companies 77.78 percent of the time

34
Bankruptcy Prediction Model (3 yrs in advance)

• Healthy construction companies 77.78 percent of
  the time
• Unhealthy companies 69.23 percent of the time

35
Bankruptcy Prediction Model (General Model)

• Healthy construction companies 32.14 percent of
  the time
• Unhealthy companies 38.71 percent of the time

36
Altmans Z-Score Model

• Z0.012 X1 0.014 X2 0.033 X3 0.006 X4
  0.999 X5
• X1 working capital/total assets
• X2 retained earnings/total assets
• X3 earnings before interest and taxes/total
  assets
• X4 market value of equity/book value of total
  liabilities
• X5 sales/total assets
• If output value lt 1.81, company classified as
  unhealthy
• If output value gt 2.67, company classified as
  healthy
• (Gray Area) for values between 1.81 and 2.67

37
Altmans Z-Score Model (cont.)
38
Weights of the Financial Ratios

• Disadvantage with using neural networks the user
  cannot know the importance of each one of the
  inputs in the model (weight)
• Numerical Analysis
• Data used mean of the data for each one of the
  years
• A weight of one was given to each one of the
  seven variables
• The variable means are multiplied by the weight
  and added together to find a base number
• The weight of one of the variables was changed
  from 1 to 5 meanwhile the other weights were
  kept as one
• The variables with higher increments were chosen
  as the most important variables

39
Weights of the Financial Ratios (cont.)
40
Gray Point

• Results from the neural network model two
  independent probabilities
• U (Pu) probability for that company to be
  unhealthy
• H (Ph) probability for that company to be
  healthy
• Odds ratios can be defined as a ratio of the
  probability that an event will occur versus the
  probability that the event will not occur.
• If PHgtPU, Healthy Odds ratio HORPH/PU
• If PHltPU, Unhealthy Odds ratio UORPU/PH
• HOR (Healthy Company Odds Ratio) odds that the
  company is healthy relative to it being unhealthy
• UOR (Unhealthy Company Odds Ratio) odds that the
  company is unhealthy relative to it being healthy
• The gray point happens when the Odds ratio is
  equal to one

41
Conclusions

• The results obtained by using the three neural
  network models were
• One year before business failure (7-2-2-1)
• Healthy construction companies 88.89 percent of
  the time
• Unhealthy companies 100 percent of the time
• Two years before business failure (7-7-2-1)
• Healthy construction companies 90 percent of
  the time
• Unhealthy companies 77.78 percent of the time
• Three years before business failure (7-8-2-1)
• Healthy construction companies 77.78 percent of
  the time
• Unhealthy companies 69.23 percent of the time

42
Conclusions (cont.)

• Three significant financial ratios in bankruptcy
  prediction debt-to-equity ratio, debt-to-assets
  ratio and gross profit margin ratio
• These three variables are the main financial
  ratios in bankruptcy prediction for the type of
  construction companies used in this research
  (heavy construction, utility construction, and
  commercial construction.)
• The neural network model will provide the user
  with two numeric indicators (the probability for
  that company to be healthy and the probability to
  be unhealthy). The higher number will indicate
  whether the company is either healthy or
  unhealthy

43
Conclusions (cont.)

• The results provided by a neural network do not
  have a Gray Area. Instead, those results have
  a possible Gray Point. (when the probability
  that the company is healthy is the same as if the
  company is unhealthy
• The accuracy of the model can be improved if more
  data from public construction companies and from
  private construction companies are used to train
  the model

44
Research Limitations

• Lack of data from private construction companies
  concerning financial conditions
• Data from construction companies (heavy
  construction, utility construction, and
  commercial construction) were used to train and
  test the model
• The software used to train and test the neural
  networks was a student version. This version
  limited the development of a software
  application. (Although it proved adequate for
  this initial study)

45
Recommendations

• An upgraded version of NeuroSolutions should be
  used in order to develop software
• More data should be collected from public
  construction companies to improve the accuracy of
  the neural network model
• Data from private companies should be collected
  and added to the existing data
• Financial ratios should be calculated and
  analyzed using data from private companies. The
  same should be done for smaller companies

46
Recommendations (cont.)

• Data from other fields of the construction
  industry such as building construction should be
  gathered to improve the accuracy of the model
  (create general model)
• Surety companies should be surveyed to determine
  the financial ratios they use to measure the
  fiscal health status of a construction company
• A computer with a faster processor should be
  used in order to reduce the training time

47
Questions?