Global Journal of Management and Business Research, A: Administration and Management, Volume 22 Issue 1

We will now discuss a method used to extract the feature values from the trend during the 5-year period (Y t , Y t-1 , …, Y t-4 ). The following five patterns ( ① through ⑤ ) were considered using the current asset data included in Table 3, the results of which are shown in Table 4. ① Arithmetic mean of the negative value change rate: This takes as a feature value the mean value of the change rate over the 5-year period (year- equivalent mean value), as in “(Y t-4 + Y t-3 + Y t-2 + Y t-1 +Y t )/5.” ② Absolute minimum of negative value change rate: This takes as a feature value the absolute minimum of the negative change rates over the 5-year period, as in “min(|Y t |, |Y t-1 |, … , |Y t-4 |).” ③ Absolute maximum of negative value change rate: This takes as a feature value the absolute maximum of the negative change rates over the 5-year period, as in “max(|Y t |, |Y t-1 |, … , |Y t-4 |).” ④ Sum of negative value change rate: This takes as a feature value the sum of the negative change rates over the 5-year period, as in “sum(Y t , Y t-1 , … , Y t-4 ).” ⑤ Year-equivalent change rate between 4 years before and the final year: This takes as a feature value the change rate in years equivalent to between the first and last years of the 5-year period, as in “( (Y t - Y t- 4 )/Y t-4 )/4.” The procedure is as follows: Table 4 represents one company that had articulated differences in the rates of change. Here, “ ③ Absolute maximum of negative value change rates” had the largest negative value in the reduction of current assets, and was thus selected as the feature value { FV(Feature Value) }. The formula is expressed as follows, where Y t is obtained, as shown in (1). FV = -max (|Yt|, |Yt-1|, … , |Yt-4|) , {Yt, Yt-1, … , Yt-4} < 0 (2) Here, FV was obtained for each of the eight bankruptcy prediction indices (explanatory variables), which will serve as the input data for clustering in the next section. Table 4: Comparison of methods used to extract feature values from the 5-year trend (Y t , Y t-1 , … ,Y t-4 ) (unit: 1,000 yen) Current assets Final year 1 year before 2 years before 3 years before 4 years before Y on Y change rate 4,977,284 3,921,740 6,208,416 40,249,410 981,290 ① 9.77 0.27 -0.37 -0.85 40.02 ② −0.37 ③ −0.85 ④ 39.07 ⑤ −0.25 Note: Key to the numbers in a circle: ① Change rate: Arithmetic mean; ② Absolute minimum of negative change rate; ③ Absolute maximum of negative change rate; ④ Sum of negative change rate; ⑤ Year-equivalent change rate between 4 years before and the final year. d) Clustering (Machine learning) model We take the FVs obtained for the eight previously described bankruptcy prediction indexes and create matrix data, as illustrated in Table 5, which will be fed into the machine learning (clustering). Note that the rows are equal to the number of sample datasets, and the names of sampled companies (both bankrupt and non-bankrupt) will appear in the first column. Table 5: Clustering model Company Explanatory variable Current ratio Operatingcash flow · · · Equity/Total liabilities A FV FV FV FV B FV FV FV FV … FV FV FV FV n FV FV FV FV The data in Table 5 are the distance matrix and not the adjacency matrix. The adjacency matrix cannot be used for clustering because distance data are generated between samples (companies). For this reason, we selected a clustering method based on the distance matrix, as advanced by Otsuki [16]. According to this method, the Euclidean distance is obtained based on the principal component scores calculated until the cumulative contribution ratio surpasses 90%, forming a matrix of principal component scores. A silhouette analysis, as shown in (3), is then run on this principal component score matrix, and the number of clusters K is taken at the highest silhouette value at which clustering takes place. 13 Global Journal of Management and Business Research Volume XXII Issue I Version I Year 2022 ( ) A © 2022 Global Journals A Study on Machine Learning Prediction Model for Company Bankruptcy using Features in Time Series Financial Data