Proceedings of the 2015 International Conference on Industrial Engineering and Operations Management Dubai, United Arab Emirates (UAE), March 3 – 5, 2015
An Application of "Neuro-Logit" New Modeling Tool in Corporate Financial Distress Diagnostic Waleed E. Almonayirie Research Scholar, DBA Program Swiss Business School (SBS-UAE)
Abu Dhabi, UAE
[email protected] Abstract— During the last decades and recession of 2007-2009 witnessed many global financial crises. Consequently, this research represents a proactive study via introducing new modeling tool; in order to diagnose the financial distress and assess its probability of occurrence. The Neuro-Logit is a new approach for diagnosis, prediction and forecasting corporate financial distress. This tool acts as Logit (Logistic Regression Analysis), but the equations are built based on the basic algorithm of ANN (Artificial Neural Network). ANN and Logit are widely used as modeling tools in many business applications; Neuro-Logit model reduces most of ANN and Logit limitations. The sample in this research has been drawn from the available financial statements (Financial Ratio-Based Model) that are belonged to most active non-financial firms in Egyptian stock markets. The observations are quarterly basis observations, covering six-year time period (2004-2009). The overall results show that Neuro-Logit model has superior outcome comparing to legacy Logit model, where the overall classification accuracy rate almost 86% with Type I Error 10.13%, 85.33% harmonic mean between Recall and Precision values and also good Kappa coefficient (0.7169) and Matthew Correlation Coefficient (0.7217). The paper revolves the diagnosing the financial health of the firms, and is an extension for the latest Egyptian model in 2007 which concerns with six-year span 2000-2005. The time span of the paper for model building is threeyear (2005-2007) which is covering prior-recession time. The paper can be considered as second trial of supervised financial distress prediction model and the fourth Egyptian model with superior outcome supporting to be recommended in corporate financial failure assessment and diagnosis. Also the research is presenting empirically an innovative modeling approach, where the ANN is used as statistical tool. Keywords— Corporate Financial Distress Diagnosis, Financial Analysis, Modeling Approach, Artificial Neural Networks and Logistic Regression Analysis
I.
INTRODUCTION
Corporate Failure is very important problem that affects the economy at both levels macro level and firm level. Prediction models were created along 45 years attracted the intention of both, firms and government organizations; to enhance auditing, investment and critical decisions in the studied markets. These models could be useful to take preventive and/or corrective actions that providing stability through its higher accuracy. The present research study is motivated by that need to predict the corporate failure in Egyptian setting with new Egyptian model to facilitate the decision making in the right time.
A. Research Limitations As most of researches, the study has some limitations which may affect the accuracy of the model, as following: •
The analyzed data in this study was obtained from public financial statements which may subject to creative accounting. Companies facing failure may distort their published accounts and this will skew the model results.
•
Some corporate financial statements did not disclose figures on cash flow or operating expenses. II.
LITERATURE REVIEW
Reference [1] had been represented in 1968 and is considered the pioneered model of corporate failure/bankruptcy prediction, using Multivariate Discriminant Analysis (MDA). The main advantage of the MDA approach to predict corporate failure is its ability to reduce a multidimensional problem to a single score with a high level of accuracy; where MDA combines information from multivariate independent variables (e.g. ratios) into a single score that is used to classify an observation into either of two a-priori and mutually exclusive group. Many models were created after Altman Z-score, and according to [2] the most modeling tools are (MDA, Logit and ANN, respectively). These models may not specifically tell what is wrong, but it should encourage identifying problems and taking effective action; to minimize failure incidence [3]. A. Logistic Regression Analysis (LRA/Logit) Logit is a statistical model calculated based on natural logarithm of the odds ratio [4]. The Logit generic equation as following: Pj = 1/ (1+ EXP (-Yj))
(1)
Yj = A0+A1*X1+A2*X2+…+An*Xn
(2)
Where: Pj the Logit output for company j (the probability of failure) Yj the equivalent to MDA score X1 to Xn set of independent variables A1 to An regression coefficients and A0 is the intercept of Yj Reference [4] introduced the conditional Logit in the area of corporate financial distress prediction. Consequently, Ohlson’s Logit model combines several companies’ attributes
978-1-4799-6065-1/15/$31.00 ©2015 IEEE
into a Logit output that indicates the probability of failing. A company is classified as failed or non-failed if its Logit output is below or above a priori chosen cut-off probability respectively. In addition, the coefficients in a Logit model indicate the relative importance of the independent variables. Logit models can also include qualitative variables expressed as nominal data (e.g. 1 = male). Finally, Logit models enjoy a degree of non-linearity because of the model’s logistic function. Logistic regression describes the relationship between a dichotomous response variable (success/failure) and a set of independent variables. The independent variables may be continuous or discrete with dummy variables. Logit does not require the restrictive assumptions regarding normality distribution of independent variables or equal dispersion matrices nor concerning the prior probabilities of failure as required in MDA. B. Logistic Regression Analysis Limitations Logistic regression analysis (LRA) is based on two assumptions, [2]: •
The dependent variable is required to be dichotomous, with the groups being discrete, non-overlapping and identifiable.
•
The cost of Type I and Type II error rates are considered in the selection of the optimal cut-off probability
However, Logit models have been criticized on the following grounds, [5]: •
Sensitivity to the problem of multi-co-linearity
•
Assuming a logistic probability distribution
•
Sensitivity to outliers and missing values.
C. Artificial Neural Network Reference [5] briefly shows what is the ANN? The ancestry of the current Multi-Layer Perceptron (MLP) can be traced back to the “neurode” proposed by W. McCulloch and W. Pitts in 1943. The suggested “neurode” addresses the problem of separating patterns into two categories. Fig. 1 shows an artificial neuron j which has n inputs x1,…,xi,…,xn and each input is weighted before reaching the neuron j by the connection weight w1j,…,wij,…,wnj. In addition, the activation function determines the output of the neuron Yj. The purpose of the activation function is to ensure that the output of the neuron is bounded. There are different kinds of activation functions but the sigmoid function which takes values between 0 and 1 is widely used and has the form:
Fig. 2. Feed Forword ANN
Yj = 1/ (1+ EXP (-Σ Wij*Xij))
(3)
Neural networks, as Fig. 2, consist of neurons or neurodes and connections that are organized in layers. These layers are organized basically as input layer, hidden layer and output layer. Each neuron receives a number of inputs and produces an output which will be the input to another neuron in the next layer. The connections between different neurons are associated with weights that reflect the strength of the relationships between the connected neurons. Since 1990, when the first study had appeared by Bell et al. which proposed a comparative study between ANN and statistical models in banking sector, the studies that concerning corporate failure models based on ANN, were generated sequentially with different objectives and hypotheses, in 1991 two papers were represented [6] and [7]. After 15 years a state of art paper had introduced in 2006 to summarize almost 25 studies. To be able to improve the approach by neuronal networks, efforts have to focus on: the construction of the system, the clustering of entry variables and the adjustment of learning parameters that depend greatly on human intervention [8]. Most of the researches were systematic comparison between neural networks and statistics provided a good survey of the appropriate methodologies for the failure prediction, showing that neural networks are better than statistics in modeling company failure prediction. As far as the overall correct classification was concerned, ANNs were proved to be superior as it provided the highest prediction results in insolvency [3]. The advantages of Artificial Neural Networks are they do not require the pre-specification of a functional form, or the adoption of restrictive assumptions about the characteristics of statistical distributions of the variables and errors in the model. By their nature, ANN systems are able to work with imprecise variables and with model changes over time. They are also able to adapt to the appearance of new cases which represent changes in the situation [9]. D. Neural Network Limitations As with any system, ANN has its limitations. These include [2] and [5]: The learning stage could be very long. •
Fig. 1. The Basic Neuron
The system might not achieve a stable absolute minimum configuration but might lock on local minimums without being able to move to the optimum.
•
The system might give rise to oscillating behavior in the learning phase.
•
There is no theory allowing determining the optimal structure of the system.
•
ANN often assimilated to "black boxes" in which it is difficult to extract relevant relationships among variables.
E. The Egyptian Models More specifically to the Egyptian setting, the first introduced model (Egyptian Z-score) was utilized by [10], in 1988. After ten years from the first Egyptian Z-score, the second study appeared and it was the first one concerning in using ANN and the study is a comparative one between the three tools [11]. By 2007, the latest Egyptian model had been introduced with main objective of strengthening the model by applying three independent variables classes (accounting information plus macroeconomic variables and corporate governance measures); in order to predict corporate failure [12]. Logit had been used to test the predictive accuracy of four models: •
Model I: Financial ratios only
•
Model II: Economic variables with financial ratios
•
Model III Corporate governance measures plus financial ratios
•
Model IV (which has the highest accuracy rates): The three classes' variables had been gathering. III. NEURO-LOGIT MODEL DESCRIPTION
As shown above in Eq. (1) and Eq. (3), the basic element of the ANN (Neuron) is equivalent to Logit [13].
Fig. 3. Neuro-Logit Element
Also [14] inspire the idea of hybrid model which using two modeling tools, one of it is ANN and use one of it as preselector and the other is the main model. Based on that, the researcher created the Neuro-Logit approach. Basically NeuroLogit model is ANN that acts as statistical tool (Logit), in other meaning Neuro-Logit is logistic regression analysis using back propagation learning algorithm in calculating its coefficients. Neuro-Logit is one layer ANN with two neurons or two Logit equations (one for distress and the other for health). The model training or building phase is split into two stages: Selecting Stage, the significant highly weighted coefficients (financial ratios) are selected; in order to reduce the number of input variables. Building stage: the final weights or coefficients are calculating.
IV. RESEARCH METHODOLOGY A. The Sample As [12] the sample is drawn from the most active listed firm in the Egyptian stock markets (14 listed firm and one unlisted), in quarterly basis for two reasons: firstly, due to lack of failure recording. Secondly, to have the seasonality effect as recommended in [15]; where most of the studies are annually basis. Total observations are initially 185 observations covering time span six-year. Due to lack of recording the distressed or failed firms financial statements in the Middle East generally, Egypt particularly, a supervised approach is applied which means corporate financial distress will be assigned via applying some criteria. In [12] the failing is defines according to aggressive criteria; so based on [12] and [16], the target output is basically based on the following criterion: A company is to be considered among healthy companies if and only if the following conditions are fully satisfied: •
Positive value of net operating cash flow
•
Net operating profit
•
Positive value of working capital
TABLE I. is showing the final sample observations where each failed firm must supported with the previous two years information, therefore the final sample is 157 observations that will be divided into two groups (118 observations in estimation group that built the model and covering just three-year time period (2005-2007), 39 observations in indication group which test the validity of the model). B. The Independent Variables The following table (TABLE II.) shows the chosen financial ratios, representing the independent variables or the model inputs. According to [12], the left hand side of has been assigned and the right hand side has been chosen based on [2]. These ratios are covering the financial analysis groups (liquidity, profitability, leverage, cash flow and turn over ratios). Also the ratios are extracted from the main three financial statements (Income, Balance Sheet and Cash Flow). TABLE I. Total Sample
THE SAMPLE STRUCTURE Quarterly Basis Observations Distressed
Healthy
Total
2004
0
2
2
2005
19
14
33
2006
22
25
47
2007
26
23
49
2008
8
8
16
2009
3
7
10
Total
78
79
157
TABLE II. Ratio
THE INDEPENDENT VARIABLES
Description
CA/CL TL/TA RE/TA QA/TA SE/TL OPM
Ratio
Current Assets / Current Liabilities Total Liabilities / Total Assets Retained Earnings / Total Assets Quick Assets / Total Assets Stockholder’ Equity / Total Liabilities Operating Profit Margin
NI/SE NI/S
Net Income / Sales
WC/TA
Working Capital / Total Assets
WC/S
Working Capital / Sales
Sales/Current Assets
EBIT/I
S/TA
Sales / Total Assets
LTL/SE
CA/TA
Earnings Before Taxes / Total Assets Earnings Before Interest and Taxes/ Interest Long-Term Liabilities / Stockholder’ Equity
EBT/TA
S/ CA
QA/CL
Description Net Income / Total Assets Net Income / Stockholder’ Equity
NI/TA
Quick Assets / Current Liabilities Current Assets / Total Assets
V. DATA ANALYSIS AND FINDINGS
CF/S
Cash Flow / Sales Cash Flow / Total Liabilities Cash Flow / Current Liabilities Cash Flow / Total Assets
CF/TL CF/CL CF/TA
C. Research Hypothesis and Results Testing Criteria: This empirical study has main null hypothesis (H0): “Neuro-Logit would not be superior financial ratios based model”. In order to verify that after building the model the following tests (TABLEIII & TABLE IV) are applied then compared to results from [12] which is the latest Egyptian Model represented in 2007 (EM07). TABLE III. Actual
THE PERFORMANCE MATRIX
TABLE IV.
Distressed (1)
Healthy (0)
Distressed (1)
TP
FN
Healthy (0)
FP
TN
Description
Calculation
α
Type I Error
FP / (FP+TN)
CCR
Correct Classification Rate
(TP+TN) / (TP+TN+FP+FN)
Se
Sensitivity
TP / (TP+FN)
F1(Se, PPV)
MCC Kappa
TABLE V.
OPM
TP is True Positive: refers to number of correctly classified distressed firms.
•
TN is True Negative: refers to number of correctly classified healthy firms.
•
FP is False Positive: refers to number of incorrectly classified distressed firms.
•
FN is False Negative: refers to number of incorrectly classified healthy firms.
Negative Predictive Value Harmonic Mean between Recall and Precision (Positive Predictive Value) Matthew's Correlation Coefficient Cohen's Kappa Coefficient
NPV
RE/TA
N
ACCURACY RATES MEASURES
Measure
CL/CA
Where: • N is referring to the total sample (observations). •
Also ten ratios from 22 are considered significant to build the final model. After re-calculating the test measures of [12] the TABLE VII had been obtained:
Ratio
Tested
Total Sample
Using VBA in MS-Excel 2007, Neuro-Logit final two equations and its coefficients are shown in below TABLE V., by comparing the driving ratios, there is overlapping between the two models in three financial ratios (RETA, OPM and CACL, which underlined in TABLE V. and TABLE VI.).
NI/SE WC/TA S/WC I/EBIT
LTL/SE CF/TL CF/CL Constant
TN / (TN+FN)
(2*TP) / ((2*TP)+FP+FN)
((TP*TN)-(FP*FN)) / √(TP+FP)*(TP+FN)*(TN+FP)*(TN+FN) 2*((TP*TN)-(FP*FN)) / (N*(FP+FN))+( 2*((TP*TN)-(FP*FN)))
NEURO-LOGIT MODEL FINAL COEFFECIENTS
Description Current Liabilities / Current Assets Retained Earnings / Total Assets Operating Profit Margin Net Income / Stockholder’ Equity Working Capital / Total Assets Sales / Working Capital Interest / Earnings Before Interest and Taxes Long-Term Liabilities / Stockholder’ Equity Cash Flow / Total Liabilities Cash Flow / Current Liabilities Bias Item
Equation #1 Coefficients
Equation #2 Coefficients
‐0.992576051
9.997274256
‐11.43231463
‐8.936543185
0.883740856
3.649164523
‐4.242092746
2.69689032
0.143529306
‐0.946011613
‐0.507892565
1.224513244
‐1.164817387
3.341684043
5.185642521
‐24.55377315
‐2.989296777
8.278476706
0.846554663
‐3.047607656
1.06605206
‐2.950096033
TABLE VI.
THE LATEST EGYPTIAN (EM07) MODEL EQUATION AND COEFFECIENTS
Ratio
Description
Equation Coefficients
Constant
Intercept
0.457
RETA TLTA SETL CACL QATA OPM
Retained Earnings / Total Assets Total Liabilities / Total Assets Stockholder’ Equity / Total Liabilities Current Assets / Current Liabilities Quick Assets / Total Assets Operating Profit Margin
TABLE VII. EM07 Measuresa N TP TN FP FN CCR α Se NPV F1(Se, PPV) MCC Kappa
a.
Indication Group 78 14 42 12 10 71.79% 22.22% 58.33% 80.77% 56.00% 35.36% 35.29%
The null hypothesis has been rejected; where the Neuro-Logit has greater accuracy rates (18 out of 21 measures) than latest Egyptian model.
•
The main objective of research has been verified; where the new model has to possess enhanced capabilities: higher accuracy rates, more practical generalization with powerful integration and introduces the seasonality effect. Beside that it relies on the accounting information only (financial ratios).
•
Building new Egyptian mode for corporate financial distress assessment (the fourth model).
•
Reintroducing Artificial Neural Network (ANN) based corporate failure prediction model through a new approach (Neuro-Logit).
•
Neuro-Logit eliminates the pre-assumptions and the limitations of the traditional Logistic regression analysis.
•
Neuro-Logit reduces the ANN limitations (optimal structure and training time).
•
Introducing a new innovative statistical tool.
•
The fourth Egyptian model can be considered as Egyptian corporate financial health diagnostic tool and recommended to generalize its usage by who concerns (Managers, Auditors, Investors, Traders, bank credit scoring and companies rating).
-8.24 4.131 -0.053 0.442 -4.609 -3.354
EM07 ACCURACY RATES
Estimation Group 243 59 135 18 31 79.84% 11.76% 65.56% 81.33% 70.66% 55.83% 55.44%
•
Entire Sample 321 73 177 30 41 77.88% 14.49% 64.04% 81.19% 67.28% 50.79% 60.64%
VII. FUTURE WORK
The Researcher Recalculated based on [12] information.
The driving ratios of Neuro-Logit are covering most of the financial analysis groups (liquidity, profitability, leverage, cash flow and turn over ratios), also the ratios will be extracted from the main three financial statements (Income, Balance Sheet and Cash Flow), representing short-term and long-term as well.
•
Developing the model; in order to reduce its limitations.
•
Applying it on financial institutions financial distress diagnosis.
The output from Neuro-Logit is dual probability (financial health and distress probabilities), but when both of them represent one case (distress/distress or healthy/healthy), the correct one will be considered true; due to the failing uncertainty and lack of its databases. If both are incorrect the observation will be classified incorrectly (false case). According to that the TABLE VIII had been established.
•
Corporate governance and micro economic variables usefulness can be investigated.
•
The output can be multinomial or categorical to represent the financial health levels (from two case (success and fail) to eight cases (from strong healthy,..., fair healthy and distress)
•
Using the Neuro-Logit model in other applications just as (Biostatistics Applications...)
Comparing the TABLE VII and TABLE VIII, Neuro-Logit outperforms the EM07; as from 21 measures, Neuro-Logit has 18 accuracy rates better than EM07. The three accuracy rates that EM07 has, are indication group Sensitivity, estimation group Type I Error and indication group Negative Predictive Value (58.33%, 11.76% and 67.86% respectively). Some accuracy rates exceed 90 % (estimation group Sensitivity and NPV 91.38% and 91.23% respectively). The overall Kappa coefficient and MCC indicate that the model reliability is very good (71.96 and 72.17 respectively). NeuroLogit has zero prediction group Type I Error (the more costly error) and overall 10.13%. VI. CONCLUSION The study has achieved its goals as following:
TABLE VIII. NL10 Measures N TP TN FP FN CCR α Se NPV F1(Se, PPV) MCC Kappa
NEURO-LOGIT ACCURACY RATES
Estimation Group 118 53 52 8 5 88.98% 13.33% 91.38% 91.23% 89.08% 78.08% 77.98%
Indication Group 39 11 19 0 9 76.92% 00.00% 55.00% 67.86% 70.97% 61.09% 54.36%
Entire Sample 157 64 71 8 14 85.99% 10.13% 82.05% 83.53% 85.33% 72.17% 71.96%
REFERENCES [1]
E. Altman, “Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy,” Journal of Finance, September 1968. [2] G. Hossari, “A Dynamic Ratio-Based Model for Signaling Corporate Collapse,” JAMAR, Vol. 4 · No. 1; 2006. [3] KC. Chung, S. S. Tan, D. K. Holdsworth, “Insolvency Prediction Model Using Multivariate Discriminant Analysis and Artificial Neural Network for the Finance Industry in New Zealand,” International Journal of Business and Management, January 2008. [4] J. Ohlson, "Financial Ratios and the Probabilistic Prediction of Bankruptcy," Journal of Accounting Research, 18(1), pp. 109–131, 1980. [5] A. Argyris, “Predicting Financial Distress Using Neural Networks: Another Episode to the Serial?,” HANKEN, Swedish School of Economics and Business Administration, Department of Accounting, 2006. [6] P. K. Coats and L. F. Fant, "A Neural Approach to Forecasting Financial Distress," Journal of Business Forecasting Methods and Systems, vol. 10, no. 4, pp. 9-12, 1991. [7] Raghupathi, Schkade and Raju, A neural network approach to bankruptcy prediction, NN in Finance and Investing: Using AI to improve real-world performance TRIPPI/TURBAN, Irwin Professional Publishing, revised 1996, pp. 227-241, first publication in Proceedings of the IEEE 24th Annual Hawaii International Conference on Systems Sciences, 1991. [8] M. Perez, Neural Networks Applications in Bankruptcy Forecasting: A State of The Art, Allocataire de Recherche Modeme, Centre de Recherche de l’IAE, ESA CNRS 5055, Université Jean Moulin Lyon 3, and published at Neural Computer & Application. 15, pp. 154-163. 2006. [9] E. Severin, B. Shs n°3, Neural Networks and Their Application in the Fields of Corporate Finance, hal-00325117, version 1 – 26, Sep 2008. [10] A. Youssef, "Enhancing Financial Analysis with Statistical Methods to Rationalize Investment Decisions in Egyptian Money Market", Ph.D. Thesis, Faculty of Commerce, Cairo University, 1988, Unpublished.
[11] T. Hassanein, "Measuring the Ability of Cash Flows to Predict the Going Concern Using Artificial Neural Networks," The Scientific Journal for Research and Commercial Studies, Helwan University, (3/4), pp. 289–333, 1998. In Arabic. [12] H. S. Abou El Sood, “The Usefulness of A Composite Model to Failure Prediction,” Boston College; 2008 ABR & TLC Conference Proceedings; Orlando, Florida, USA [13] D. Ventura, The Perceptron Algorithm, Computer Science Department, Brigham Young University, January 8, 2009. [14] J. Yim, H. Mitchell, “A comparison of corporate distress prediction models in Brazil: Hybrid Neural Networks, Logit models and Discriminant Analysis,” Nova Economia, Belo Horizonte, 15 (1), pp.7393, 2005 [15] M. Nwogugu, Decision-Making, Capital Markets Risk and Corporate Governance: A Critique of Bankruptcy/Recovery Prediction Models – Part One, New York 2003 [16] T. Lee, Y. Yeh, and R. Liu, Can Corporate Governance Variables Enhance the Prediction Power of Accounting-Based Financial Distress Prediction Models, Working Paper No.2003-14, Institute of Economic Research, Hitotsubashi University, 2003
BIOGRAPHY Waleed E. Almonayirie has more than fourteen-year experience in satellite communication (Technical Support, Operations and Pre-Sales) and works as VSAT Expert in UAE. He earned B.Sc. in Electronics and Communication from Faculty of Engineering, Helwan University, Egypt. He has awarded by MIBA (Master of International Business Administration), Global Finance, ESLSCA Business School, Egypt. He is now enrolling the DBA program in Swiss Business School (SBSUAE). His interests in Artificial Neural Networks, Pattern Recognition, Optimization, Modeling, Financial Analysis, Cryptosystems, Quantitative Finance and Management Science. He is publishing his work of the MIBA research project and the DBA dissertation. This paper is based on the MIBA graduation research project.
