Loading [MathJax]/jax/output/SVG/jax.js
Research article Special Issues

Benchmarking alternative interpretable machine learning models for corporate probability of default

  • In this study we investigate alternative interpretable machine learning ("IML") models in the context of probability of default ("PD") modeling for the large corporate asset class. IML models have become increasingly prominent in highly regulated industries where there are concerns over the unintended consequences of deploying black box models that may be deemed conceptually unsound. In the context of banking and in wholesale portfolios, there are challenges around using models where the outcomes may not be explainable, both in terms of the business use case as well as meeting model validation standards. We compare various IML models (deep neural networks and explainable boosting machines), including standard approaches such as logistic regression, using a long and robust history of corporate borrowers. We find that there are material differences between the approaches in terms of dimensions such as model predictive performance and the importance or robustness of risk factors in driving outcomes, including conflicting conclusions depending upon the IML model and the benchmarking measure considered. These findings call into question the value of the modest pickup in performance with the IML models relative to a more traditional technique, especially if these models are to be applied in contexts that must meet supervisory and model validation standards.

    Citation: Michael Jacobs, Jr. Benchmarking alternative interpretable machine learning models for corporate probability of default[J]. Data Science in Finance and Economics, 2024, 4(1): 1-52. doi: 10.3934/DSFE.2024001

    Related Papers:

    [1] Yiqin Bao, Qiang Zhao, Jie Sun, Wenbin Xu, Hongbing Lu . An edge cloud and Fibonacci-Diffie-Hellman encryption scheme for secure printer data transmission. Mathematical Biosciences and Engineering, 2024, 21(1): 96-115. doi: 10.3934/mbe.2024005
    [2] Takao Komatsu, Haotian Ying . The p-Frobenius and p-Sylvester numbers for Fibonacci and Lucas triplets. Mathematical Biosciences and Engineering, 2023, 20(2): 3455-3481. doi: 10.3934/mbe.2023162
    [3] Zhongxue Yang, Yiqin Bao, Yuan Liu, Qiang Zhao, Hao Zheng, Wenbin Xu . Lightweight blockchain fuzzy decision scheme through MQTT and Fibonacci for sustainable transport. Mathematical Biosciences and Engineering, 2022, 19(12): 11935-11956. doi: 10.3934/mbe.2022556
    [4] Yan Wang, Guichen Lu, Jiang Du . Calibration and prediction for the inexact SIR model. Mathematical Biosciences and Engineering, 2022, 19(3): 2800-2818. doi: 10.3934/mbe.2022128
    [5] Anh Quang Tran, Tien-Anh Nguyen, Van Tu Duong, Quang-Huy Tran, Duc Nghia Tran, Duc-Tan Tran . MRI Simulation-based evaluation of an efficient under-sampling approach. Mathematical Biosciences and Engineering, 2020, 17(4): 4048-4063. doi: 10.3934/mbe.2020224
    [6] Yong Tao, Fan Ren, Youdong Chen, Tianmiao Wang, Yu Zou, Chaoyong Chen, Shan Jiang . A method for robotic grasping based on improved Gaussian mixture model. Mathematical Biosciences and Engineering, 2020, 17(2): 1495-1510. doi: 10.3934/mbe.2020077
    [7] Zhipeng Zhao, Zhenyu Hu, Zhiyu Zhao, Xiaoyu Du, Tianfei Chen, Lijun Sun . Fault-tolerant Hamiltonian cycle strategy for fast node fault diagnosis based on PMC in data center networks. Mathematical Biosciences and Engineering, 2024, 21(2): 2121-2136. doi: 10.3934/mbe.2024093
    [8] Auwalu Saleh Mubarak, Zubaida Said Ameen, Fadi Al-Turjman . Effect of Gaussian filtered images on Mask RCNN in detection and segmentation of potholes in smart cities. Mathematical Biosciences and Engineering, 2023, 20(1): 283-295. doi: 10.3934/mbe.2023013
    [9] Sergio Bermudo, Robinson A. Higuita, Juan Rada . k-domination and total k-domination numbers in catacondensed hexagonal systems. Mathematical Biosciences and Engineering, 2022, 19(7): 7138-7155. doi: 10.3934/mbe.2022337
    [10] Lal Hussain, Wajid Aziz, Ishtiaq Rasool Khan, Monagi H. Alkinani, Jalal S. Alowibdi . Machine learning based congestive heart failure detection using feature importance ranking of multimodal features. Mathematical Biosciences and Engineering, 2021, 18(1): 69-91. doi: 10.3934/mbe.2021004
  • In this study we investigate alternative interpretable machine learning ("IML") models in the context of probability of default ("PD") modeling for the large corporate asset class. IML models have become increasingly prominent in highly regulated industries where there are concerns over the unintended consequences of deploying black box models that may be deemed conceptually unsound. In the context of banking and in wholesale portfolios, there are challenges around using models where the outcomes may not be explainable, both in terms of the business use case as well as meeting model validation standards. We compare various IML models (deep neural networks and explainable boosting machines), including standard approaches such as logistic regression, using a long and robust history of corporate borrowers. We find that there are material differences between the approaches in terms of dimensions such as model predictive performance and the importance or robustness of risk factors in driving outcomes, including conflicting conclusions depending upon the IML model and the benchmarking measure considered. These findings call into question the value of the modest pickup in performance with the IML models relative to a more traditional technique, especially if these models are to be applied in contexts that must meet supervisory and model validation standards.



    The Fibonacci sequence is defined Fn=Fn1+Fn2 for n2 with the initial conditions F0=0,F1=1. The Fibonacci sequence and the golden ratio are used in many fields such as cryptology, coding theory and quantum physics in [1,2,3,4,5,6,7,8,9,10,11]. Horadam defined the Gaussian Fibonacci numbers in [12,13] and gave some general identities about the Gaussian Fibonacci numbers. Jordan generalized the Gaussian Fibonacci numbers with a similar definition in [14]. The Gaussian Fibonacci sequence is defined as GFn=GFn1+GFn2 with n>1 where GF0=i,GF1=1 in [14]. It can also be easily seen Fn that GFn=Fn+iFn1 where Fn is the nth Fibonacci number.

    Asci and Gurel generalized these studies and defined the k-order Gaussian Fibonacci numbers in [15] by the following recurrence relation

    GF(k)n=kj=1GF(k)nj,forn>0andk2

    with boundary conditions for 1kn0,

    GF(k)n={1i,k=1ni,k=2n0,otherwise.

    It can also be seen that GF(k)n=F(k)n+iF(k)n1 where F(k)n is the nth k-order Fibonacci number.

    In [16], Asci and Aydinyuz defined the k-order Gaussian Fibonacci polynomials and gave some important results. The k-order Gaussian Fibonacci polynomials {GFn(k)}n=0 are

    GF(k)n(x)=kj=1xkjGF(k)nj(x)

    for n>0 and k2 with the initial conditions for 1kn0,

    GF(k)n(x)={1ix,k=1ni,k=2n0,otherwise.

    It can be seen that

    GF(k)n(x)=F(k)n(x)+iF(k)n1(x)

    where F(k)n(x) is the nth k-order Fibonacci polynomial.

    In order to ensure information security in terms of data transfer over communication channels, a lot of work has been done on this subject and continues to be done. Therefore, coding/decoding algorithms play an important role to ensure information security. Especially, Fibonacci coding theory is one of the most preferred in this field. We can see examples of these in many studies. For example, in [17], Stakhov gave a new approach to a coding theory using the generalization of the Cassini formula for Fibonacci p-numbers and Qp matrices in 2006. In 2009, Basu and Prasad in [18] presented the generalized relations among the code elements for the Fibonacci coding theory. Also, Basu and Das introduced a new coding theory for Tribonacci matrices in [19] and in 2014 they defined the coding theory for Fibonacci n-step numbers by generalizing the coding theory for Tribonacci numbers in [20]. Esmaeili in [21] also described a Fibonacci polynomial-based coding method with error detection and correction.

    In this paper, we remember the coding theory for k-order Gaussian Fibonacci polynomials given in [16] by taking x=1 for k-order Gaussian Fibonacci numbers. We describe the coding theory for the k-order Gaussian Fibonacci numbers. We give illustrative examples. The interesting relation between the elements of the code-message matrix is analyzed and derived. The most important of this paper is that we deal with error detection and error correction methods and examples of error probability is given.

    When we take x=1 in the coding theory defined in [16], we obtain the k-order Gaussian Fibonacci coding method. First, we define the Qk, Rk and En(k) matrices, which plays an important role in this coding theory. Qk, Rk and En(k) are defined the k×k matrices as the following:

    Qk=[111111000001000000000010]k×k,Rk=[GF(k)k1GF(k)k2GF(k)k3GF(k)2GF(k)10GF(k)k2GF(k)k3GF(k)k4GF(k)100GF(k)k3GF(k)k4GF(k)k5000GF(k)2GF(k)1000GF(k)10000i0000i1i]k×k

    and

    E(k)n=[GF(k)n+k1GF(k)n+k2GF(k)n+1GF(k)nGF(k)n+k2GF(k)n+k3GF(k)nGF(k)n1GF(k)n+1GF(k)nGF(k)nk+3GF(k)nk+2GF(k)nGF(k)n1GF(k)nk+2GF(k)nk+1]k×k

    where GF(k)n is the nth k-order Gaussian Fibonacci number.

    Theorem 2.1. For n1, we get in [15] as follow:

    QnkRk=E(k)n. (2.1)

    Corollary 2.1. For k=2, we get

    QnR=[1110]n[1ii1i]
    =[GFn+1GFnGFnGFn1]

    where GFn is the nth usual Gaussian Fibonacci number in [15].

    Corollary 2.2. For k=3, we get

    Qn3R3=[111100010]n[1+i1010i0i1i]
    =[GTn+2GTn+1GTnGTn+1GTnGTn1GTnGTn1GTn2]

    where GTn is the nth Gaussian Tribonacci number in [23].

    In this section, we redefine k-order Gaussian Fibonacci coding theory using k-order Gaussian Fibonacci numbers and Qk, Rk and En(k) play very important role in the construction of k-order Gaussian Fibonacci coding theory.

    We now obtain the matrix En(k) using Qk, Rk matrices for the k=2,k=3 values and examine the inverse.

    • For k=2, introducing the square matrix Q2 of order 2 as:

    Q2=[1110]

    and the square matrix R2 of order 2 as:

    R2=[1ii1i]

    for n=1, we can use (2.1)

    E(2)1=Q2.R2
    =[1+i11i]
    =[GF(2)2GF(2)1GF(2)1GF(2)0]

    such that

    detE(2)1=det(Q2.R2)
    =(1).(2i)
    =2+i.

    The inverse of E(2)1 is as:

    (E(2)1)1=[1525i25+15i25+15i1535i]
    =1detE(2)1[GF0GF1GF1GF2]
    =12+i[GF0GF1GF1GF2]

    such that

    det(E(2)1)1=12+i
    =2515i.

    Also, by (2.1) for n=2, we can get E(2)2 as follows:

    E(2)2=Q22.R2
    =[2+i1+i1+i1]
    =[GF(2)3GF(2)2GF(2)2GF(2)1]

    such that

    detE(2)2=det(Q22.R2)
    =(1)2.(2i)
    =2i.

    The inverse of E(2)2 is as:

    (E(2)2)1=[25+15i1535i1535i35+45i]
    =1detE(2)2[GF(2)1GF(2)2GF(2)2GF(2)3]
    =12i[GF(2)1GF(2)2GF(2)2GF(2)3]

    such that

    det(E(2)2)1=12i
    =25+15i.

    Theorem 2.1.1. E(2)n=[GF(2)n+1GF(2)nGF(2)nGF(2)n1] where E(2)1=[1+i11i].

    Theorem 2.1.2. (E(2)n)1=1detE(2)n[GF(2)n1GF(2)nGF(2)nGF(2)n+1]

    =1(detQ2)n.detR2[GF(2)n1GF(2)nGF(2)nGF(2)n+1].

    • For k=3, introducing the square matrix Q3 of order 3 as:

    Q3=[111100010]

    and the square matrix R3 of order 3 as:

    R3=[1+i1010i0i1i]

    for n=1, we can use (2.1)

    E(3)1=Q3.R3
    =[2+i1+i11+i1010i]
    =[GF(3)3GF(3)2GF(3)1GF(3)2GF(3)1GF(3)0GF(3)1GF(3)0GF(3)1]

    such that

    detE(3)1=det(Q3.R3)
    =1.2i
    =2i.

    The inverse of E(3)1 is as:

    (E(3)1)1=[121212i12i1212i1+i1212i12i1212i12i]
    =1detE(3)1[GF(3)1GF1(3)(GF(3)0)2GF(3)0GF1(3)GF(3)1GF2(3)GF(3)0GF2(3)(GF(3)1)2GF(3)0GF1(3)GF(3)1GF2(3)GF(3)3GF1(3)(GF(3)1)2GF(3)1GF2(3)GF(3)0GF3(3)(GF(3)1)2GF(3)0GF2(3)GF(3)1GF2(3)GF(3)0GF3(3)GF(3)1GF3(3)(GF(3)2)2]
    =1detR3[GF(3)1GF1(3)(GF(3)0)2GF(3)0GF1(3)GF(3)1GF2(3)GF(3)0GF2(3)(GF(3)1)2GF(3)0GF1(3)GF(3)1GF2(3)GF(3)3GF1(3)(GF(3)1)2GF(3)1GF2(3)GF(3)0GF3(3)(GF(3)1)2GF(3)0GF2(3)GF(3)1GF2(3)GF(3)0GF3(3)GF(3)1GF3(3)(GF(3)2)2]

    such that

    det(E(3)1)1=12i
    =i2.

    Also, by (2.1) for n=2, we can get E(3)2 as follows:

    E(3)2=Q23.R3
    =[4+2i2+i1+i2+i1+i11+i10]
    =[GF(3)4GF(3)3GF(3)2GF(3)3GF(3)2GF(3)1GF(3)2GF(3)1GF(3)0]

    such that

    detE(3)2=det(Q23.R3)
    =12.2i
    =2i.

    The inverse of E(3)2 is as:

    (E(3)2)1=[12i1212i12i1212i112+32i12i12+32i1+12i]
    =1detE(3)2[GF(3)0GF(3)2(GF(3)1)2GF(3)1GF(3)2GF(3)0GF(3)3GF(3)1GF(3)3(GF(3)2)2GF(3)1GF(3)2GF(3)0GF(3)3GF(3)0GF(3)4(GF(3)2)2GF(3)2GF(3)3GF(3)1GF(3)4GF(3)1GF(3)3(GF(3)2)2GF(3)2GF(3)3GF(3)1GF(3)4GF(3)2GF(3)4(GF(3)3)2]
    =1detR3[GF(3)0GF(3)2(GF(3)1)2GF(3)1GF(3)2GF(3)0GF(3)3GF(3)1GF(3)3(GF(3)2)2GF(3)1GF(3)2GF(3)0GF(3)3GF(3)0GF(3)4(GF(3)2)2GF(3)2GF(3)3GF(3)1GF(3)4GF(3)1GF(3)3(GF(3)2)2GF(3)2GF(3)3GF(3)1GF(3)4GF(3)2GF(3)4(GF(3)3)2].

    Theorem 2.1.3. E(3)n=[GF(3)n+2GF(3)n+1GF(3)nGF(3)n+1GF(3)nGF(3)n1GF(3)nGF(3)n1GF(3)n2] where E(3)1=[2+i1+i11+i1010i].

    Theorem 2.1.4.

    (E(3)n)1=1detR3[GF(3)n2GF(3)n(GF(3)n1)2GF(3)n1GF(3)nGF(3)n2GF(3)n+1GF(3)n1GF(3)n+1(GF(3)n)2GF(3)n1GF(3)nGF(3)n2GF(3)n+1GF(3)n2GF(3)n+2(GF(3)n)2GF(3)nGF(3)n+1GF(3)n1GF(3)n+2GF(3)n1GF(3)n+1(GF(3)n)2GF(3)nGF(3)n+1GF(3)n1GF(3)n+2GF(3)nGF(3)n+2(GF(3)n+1)2].

    For arbitrary k-positive integers, the square matrix E(k)n of order k and inverses can be found similarly.

    In this section, we describe a new k-order Gaussian Fibonacci coding theory. We put our message in a matrix of M and let us represent the initial message in the form of the square matrix M of order k. We take the k-order Gaussian Fibonacci matrix E(k)n as a coding matrix and its inverse matrix (E(k)n)1 as a decoding matrix for an arbitrary positive integer k. The transformation M×E(k)n=C is called k-order Gaussian Fibonacci coding and we name the transformation C×(E(k)n)1=M as k-order Gaussian Fibonacci decoding. We define C as a code matrix.

    The given example is solved using the alphabet table below.

    Using the arbitrary value of s0, we write the following alphabet according to mod30. We can extend the characters in the table according to our wishes. We begin the “s” for the first character in Table 1.

    Table 1.  Alphabet table.
    A B C D E F G H I J
    s s+1 s+2 s+3 s+4 s+5 s+6 s+7 s+8 s+9
    K L M N O P Q R S T
    s+10 s+11 s+12 s+13 s+14 s+15 s+16 s+17 s+18 s+19
    U V W X Y Z 0 ! ? .
    s+20 s+21 s+22 s+23 s+24 s+25 s+26 s+27 s+28 s+29

     | Show Table
    DownLoad: CSV

    Example 2.3.1. Let us consider the message matrix for the following message text:

    âœCODE❠(2.2)

    Step 1: Let’s create the message matrix using the message text:

    M=[CODE]2×2

    Step 2: Let’s write the message matrix M according to the alphabet table for the arbitrary value “s” we choose. For s=2;

    M=[41656]

    Step 3: For k=2, n=2, we use (2.1);

    E(2)2=Q22.R2
    =[2+i1+i1+i1]

    Step 4: The code message is:

    C=M×E(2)2
    =[41656][2+i1+i1+i1]
    =[24+20i20+4i16+11i11+5i]

    Step 5: The decode message is:

    M=C×(E(2)2)1
    =[24+20i20+4i16+11i11+5i][25+15i1535i1535i35+45i]
    =[41656]
    =[CODE].

    Example 2.3.2. Let us consider the message matrix for the following message text:

    "PUBLICKEY"

    Step 1: Let’s create the message matrix using the message text:

    M=[PUBLICKEY]3×3

    Step 2: Let’s write the message matrix M according to the alphabet table for the arbitrary value “s” we choose. For s=5:

    M=[202561613715929]

    Step 3: For k=3, n=6, we use (2.1):

    E(3)6=Q63.R3
    =[44+24i24+13i13+7i24+13i13+7i7+4i13+7i7+4i4+2i]

    Step 4: The code message is:

    C=M×E(3)6
    =[1558+847i847+459i459+252i1107+602i602+327i327+178i1253+680i680+374i374+199i]

    Step 5: The decode message is:

    M=C×(E(6)3)1
    =[PUBLICKEY]
    =[202561613715929].

    In this section, we consider the k-order Gaussian Fibonacci coding/decoding method for k=2. We have an interesting relation among the elements of a code matrix C that has a crucial role in the error-correction process, outlined as follows:

    C=M×E(2)n=[m1m2m3m4]×[GF(2)n+1GF(2)nGF(2)nGF(2)n1] (3.1)
    =[c1c2c3c4] (3.2)

    and

    M=C×(E(2)n)1 (3.3)
    =[c1c2c3c4]×(1(det(Q2))ndetR2[GF(2)n1GF(2)nGF(2)nGF(2)n+1]) (3.4)
    =[m1m2m3m4]

    For the case for an even integer n=2m, we obtain the following equation

    [m1m2m3m4]=[c1c2c3c4]×(12i[GF(2)n1GF(2)nGF(2)nGF(2)n+1]) (3.5)

    since the det(Q2)=1 and det(R2)=2i.

    It follows from (3.5) that the elements of the matrix M can be obtained according to the following formulas:

    m1=12i(GF(2)n1c1GF(2)nc2) (3.6)
    m2=12i(GF(2)nc1+GF(2)n+1c2) (3.7)
    m3=12i(GF(2)n1c3GF(2)nc4) (3.8)
    m4=12i(GF(2)nc3+GF(2)n+1c4) (3.9)

    Since the s0, the elements of the matrix M are

    m10,m20,m30andm40 (3.10)

    Because of the condition (3.10) we can write the equalities (3.6)–(3.9) as:

    m1=12i(GF(2)n1c1GF(2)nc2)0 (3.11)
    m2=12i(GF(2)nc1+GF(2)n+1c2)0 (3.12)
    m3=12i(GF(2)n1c3GF(2)nc4)0 (3.13)
    m4=12i(GF(2)nc3+GF(2)n+1c4)0 (3.14)

    From the Eqs (3.11) and (3.12), we can get

    GF(2)nGF(2)n1c2c1GF(2)n+1GF(2)nc2 (3.15)

    or

    GF(2)nGF(2)n1c1c2GF(2)n+1GF(2)n (3.16)

    Similarly, from (3.13) and (3.14), we can obtain:

    GF(2)nGF(2)n1c4c3GF(2)n+1GF(2)nc4 (3.17)

    or

    GF(2)nGF(2)n1c3c4GF(2)n+1GF(2)n (3.18)

    Since two consecutive Gaussian Fibonacci numbers approach the golden ratio, we also obtain the following equations in (3.16) and (3.18) that connect the elements of the code matrix in (3.1):

    c1τc2 (3.19)
    c3τc4 (3.20)

    where τ=1+52 is the “golden ratio”.

    Similarly, for an odd integer n=2m+1, we can write the same approximate equalities (3.19) and (3.20) that connect the elements c1,c2,c3 and c4 of the code matrix in (3.1).

    Thus, we have generated some important identities that connect the elements of the code matrix in (3.1) for the case k=2.

    If we consider the coding/decoding method for k=3, we can obtain interesting identities as in k=2. Since two consecutive Gaussian Tribonacci numbers approach Φ where Φ=1.8393. However, for general case of k we can find mathematical identities that connect the code matrix elements similar to the (3.19) and (3.20).

    Example 3.1. Suppose k=2. Thus, we have τ=1+521.618. Assume that the following message-matrix is to be transmitted in (2.2):

    For s=2, we can get

    M=[CODE]=[41656]

    • If n=2 then

    C=[24+20i20+4i16+11i11+5i]=[c1c2c3c4]

    In this case, we have the following numbers rounded off to their first five digits:

    c1c2=1.3462+0.7307i,c3c4=1.5822+0.2808i

    • For n=5, we have

    C=[112+68i68+44i70+43i43+27i]

    In this case

    c1c2=1.61710.0463i,c3c4=1.61790.0159i

    • For n=10, we have

    C=[1236+764i764+472i775+479i479+296i]

    In this case

    c1c2=1.618+0.0003i,c3c4=1.618+0.0001i

    • For n=15, we have

    C=[13708+8472i8472+5236i8595+5312i5312+3283i]

    In this case

    c1c2=1.618,c3c4=1.618

    These show that for n=15 the relation (3.19) and (3.20) holds very well.

    Therefore, we select the value of n large enough, as seen in Examples 2.3.1 we get a good approximation of τ. If we apply similar operation in Example 2.3.2, when n is chosen large enough we get a good approximation of Φ where Φ=1.8393 and Φ is the ratio between two Gaussian Tribonacci numbers.

    The k-order Gaussian Fibonacci coding/decoding method considered above provides an interesting possibility to detect and correct “errors” in code message C. Error detection and correction for k=2 depends on the states of the determinant of the matrix C and the connections with the code matrix elements at (3.19) and (3.20). These mathematical relations given by (3.19), (3.20) and the determinant of the code matrix C play a role in checking relations of the k-order Gaussian Fibonacci coding/decoding method. The error-correction algorithm for Fibonacci coding has been explained in [10] and [13]. This defined algorithm is valid and applicable to the k-order Gaussian Fibonacci numbers we have given in this study.

    First, let’s calculate the determinant of our message matrix M and send it to the communication channel right after the code matrix elements. Here, the determinant of M will be the control element for the C code matrix received from the communication channel. After taking the determinant of the code matrix C and its control element M, we can calculate the determinant of the code matrix C and compare the determinant of M with the determinant of C with respect to the control relation. If the operations we have done after this check are exactly appropriate, we can conclude that the elements of the code matrix C are transmitted on the communication channel without errors. If the opposite happens, we can say that there are errors in the elements of the code matrix C or the determinant of the control element M.

    Suppose the elements of the C code matrix errors. This matrix can have a one-fold error, double-fold error or k2 fold error. To explain how to correct these errors, we consider the 2×2 type matrix obtained for k=2. These cases are considered.

    Our first hypothesis is that we have the case of a “single” error in the code matrix C received from the communication channel. It is clearly seen that there are four types of single-fold errors in the C code matrix:

    (a)[xc2c3c4](b)[c1yc3c4](c)[c1c2zc4](d)[c1c2c3t] (4.1)

    where x,y,z,t are possible “destroyed” elements.

    Since it is C=M×E(2)n, if we take the determinant of both sides, we get the following equations:

    xc4c2c3=det(M)(1)n(2i)(apossibleâœsingleerrorâisintheelementc1) (4.2)
    c1c4yc3=det(M)(1)n(2i)(apossibleâœsingleerrorâisintheelementc2) (4.3)
    c1c4c2z=det(M)(1)n(2i)(apossibleâœsingleerrorâisintheelementc3) (4.4)
    c1tc2c3=det(M)(1)n(2i)(apossibleâœsingleerrorâisintheelementc4) (4.5)

    or equivalently

    x=det(M)(1)n(2i)+c2c3c4(a) (4.6)
    y=det(M)(1)n(2i)+c1c4c3(b) (4.7)
    z=det(M)(1)n(2i)+c1c4c2(c) (4.8)
    t=det(M)(1)n(2i)+c2c3c1(d) (4.9)

    The formulas (4.6)–(4.9) give four possible states of a “single error”. However, we need to choose the correct case from among x,y,z,t integer solutions. We have to choice such solutions, which satisfies to the additional “checking relations” (3.19) and (3.20). If calculations by formulas (4.6)–(4.9) do not give an integer result, we can conclude that our “single error” hypothesis is incorrect or that the determinant of control element M is “error”. In the second case, we can use approximate Eqs (3.19) and (3.20) to check the correctness of the C code matrix.

    Similarly, we can easily check “double errors” of the C code matrix. Suppose that the code matrix C has elements x and y error as shown below:

    [xyc3c4] (4.10)

    Then,

    xc4yc3=det(M)(1)n(2i) (4.11)

    However, according to the (3.19) there is the following relation between x and y

    xτy (4.12)

    Again, only integer solutions are acceptable. If no integer solution is obtained, then two-fold errors have not occurred.

    Hence, we can show that with this type of approach, there is the possibility to correct all possible triple-fold errors in the C code matrix.

    [xyzc4].

    Therefore, our method of error correction depends on the verification of different hypotheses regarding errors in the code matrix using the determinant of the control relationship C matrix, (3.19) and (3.20), and the elements of the code matrix as integers. If all our solutions do not reach integer solutions, it means that the determinant of the control item M is wrong or there is a four-fold error in the C code matrix, and we should reject it. The code matrix C is defective and not correctable.

    As a result, there are 15 error conditions in the C code matrix. According to the method given in [17], it means that the correctable probability of the method is equal since 14 cases between them can be corrected and

    Scor=1415=0.9333=93.33%

    If we generalize this equation as in [20], since only k2 fold errors cannot be corrected with this method, the error correction capacity of the method is

    2k212k2

    where k is the order of the message-matrix. So far sufficiently large values of k, the probability of decoding error is almost zero.

    Consequently, for a sufficiently large value of k, in the case of n=k, the correct ability of this method is

    2k222k21.

    Therefore, for large value of k, the correct possibility of the method is

    2k222k211=100%.

    In this paper, we obtained the coding theory for k-order Gaussian Fibonacci polynomials given in [16] by taking x=1 for k-order Gaussian Fibonacci numbers and gave illustrative examples. This coding method differs from classical algebraic coding methods in this respect. We can show this difference with the following features:

    1)      Since the n,k and s0 values are arbitrarily chosen in the k-order Gaussian Fibonacci coding theory, it is difficult to estimate the information transmission between the two channels by a third channel, which increases the reliability of the information transmission. In addition, this method depends on matrix multiplication and can be performed quickly and easily by today’s computers.

    2)      With k-order Gaussian Fibonacci coding theory, we can encrypt and send messages of the desired length by enlarging the k value sufficiently.

    3)      The main practical feature of this method is that matrix elements are error detection and correction objects.

    4)      The Gaussian Fibonacci coding method, which is the simplest coding method obtained for the k=2 value, has been redefined by handling the 2×2 type matrices with “single-fold, double-fold and triple-fold errors”.

    5)      In the simplest case, for k=2, the correct capability of the method is essentially equal to 93.33%, exceeding all well-known correction codes.

    6)      This error correction method given for k=2 is generalized and the correct ability of the errors of this method increases as k increases, and the correct ability approaches 100% for a sufficiently large value of k.

    7)      This article is just a brief outline of a new coding theory based on the k-order Gaussian Fibonacci matrices of the articles in [17,18,20,21] and [22].

    The authors declare there is no conflict of interest.



    [1] Abdulrahman UFI, Panford JK, Hayfron-Acquah JB (2014) Fuzzy logic approach to credit scoring for micro finances in Ghana: a case study of KWIQPUS money lending. Int J Comput Appl 94: 11–18. https://doi.org/10.5120/16362-5772 doi: 10.5120/16362-5772
    [2] Vahid PR, Ahmadi A (2016) Modeling corporate customers' credit risk considering the ensemble approaches in multiclass classification: evidence from Iranian corporate credits. J Credit Risk 12: 71–95.
    [3] Allen L, Peng L, Shan Y (2020) Social networks and credit allocation on fintech lending platforms. working paper, social science research network. Available from: https://papers.ssrn.com/sol3/papers.cfm?abstract_id = 3537714.
    [4] Altman EI (1968) Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. J Financ 23: 589–609. Available from: https://pdfs.semanticscholar.org/cab5/059bfc5bf4b70b106434e0cb665f3183fd4a.pdf.
    [5] Altman EI, Narayanan P (1997) An international survey of business failure classification models. in financial markets, institutions and instruments. New York: New York University Salomon Center, 6. Available from: https://onlinelibrary.wiley.com/doi/abs/10.1111/1468-0416.00010.
    [6] American Banking Association (2018) New credit score unveiled drawing on bank account data. ABA Bank J, October 22.
    [7] Anagnostou I, Kandhai D, Sanchez Rivero J, et al. (2020) Contagious defaults in a credit portfolio: a Bayesian network approach. J Credit Risk 16: 1–26. https://dx.doi.org/10.2139/ssrn.3446615 doi: 10.2139/ssrn.3446615
    [8] Bjorkegren D, Grissen D (2020) Behavior revealed in mobile phone usage predicts credit repayment. World Bank Econ Rev 34: 618–634. https://doi.org/10.1093/wber/lhz006 doi: 10.1093/wber/lhz006
    [9] Bonds D (1999) Modeling term structures of defaultable bonds. Rev Financ Stud 12: 687–720. Available from: https://academic.oup.com/rfs/article-abstract/12/4/687/1578719?redirectedFrom = fulltext.
    [10] Breeden J (2021) A survey of machine learning in credit risk. J Risk Model Validat 17: 1–62.
    [11] Chava S, Jarrow RA (2004) Bankruptcy prediction with industry effects. Rev Financ 8: 537–69. Available from: https://papers.ssrn.com/sol3/papers.cfm?abstract_id = 287474.
    [12] Clemen R (1989) Combining forecasts: a review and annotated bibliography. Int J Forecast 5: 559–583 Available from: https://people.duke.edu/~clemen/bio/Published%20Papers/13.CombiningReview-Clemen-IJOF-89.pdf.
    [13] Coats PK, Fant LF (1993) Recognizing financial distress patterns using a neural network tool. Financ Manage 22: 142–155. Available from: https://ideas.repec.org/a/fma/fmanag/coats93.html.
    [14] Collobert R, Weston J (2008) A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th International Conference on Neural Information Processing Systems, 1: 160–167. Association for Computing Machinery, New York. https://doi.org/10.1145/1390156.1390177
    [15] Duffie D, Singleton KJ (1999) Simulating correlated defaults. Paper presented at the Bank of England Conference on Credit Risk Modeling and Regulatory Implications Working Paper, Stanford University. Available from: https://kenneths.people.stanford.edu/sites/g/files/sbiybj3396/f/duffiesingleton1999.pdf.
    [16] Dwyer DW, Kogacil AE, Stein RM (2004) Moody's KMV RiskCalcTM v2.1 Model. Moody's Analytics. Available from: https://www.moodys.com/sites/products/productattachments/riskcalc%202.1%20whitepaper.pdf.
    [17] Harrell FE Jr (2018) Road Map for Choosing Between Statistical Modeling and Machine Learning, Stat Think. Available from: http://www.fharrell.com/post/stat-ml.
    [18] Jacobs Jr M (2022a) Quantification of model risk with an application to probability of default estimation and stress testing for a large corporate portfolio. J Risk Model Validat 15: 1–39. https://doi.org/10.21314/JRMV.2022.023 doi: 10.21314/JRMV.2022.023
    [19] Jacobs Jr M (2022b) Validation of corporate probability of default models considering alternative use cases and the quantification of model risk. Data Sci Financ Econ 2: 17–53. https://doi.org/10.3934/DSFE.2022002 doi: 10.3934/DSFE.2022002
    [20] Jarrow RA, Turnbull SM (1995) Pricing derivatives on financial securities subject to credit risk. J Fnanc 50: 53–85. https://doi.org/10.1111/j.1540-6261.1995.tb05167.x doi: 10.1111/j.1540-6261.1995.tb05167.x
    [21] Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems, 1: 1097–1105. Available from: https://cse.iitk.ac.in/users/cs365/2013/hw2/krizhevsky-hinton-12-imagenet-convolutional-NN-deep.pdf.
    [22] Kumar IE, Venkatasubramanian S, Scheidegger C, et al. (2020) Problems with Shapley-Value-Based Explanations as Feature Importance Measures. In International Conference on Machine Learning Research: 5491–5500. Available from: https://proceedings.mlr.press/v119/kumar20e.html.
    [23] Lessmann S, Baesens B, Seow HV, et al. (2015) Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research. Europ J Oper Res 247: 124–136. https://doi.org/10.1016/j.ejor.2015.05.030 doi: 10.1016/j.ejor.2015.05.030
    [24] Li K, Niskanen J, Kolehmainen M, et al. (2016) Financial innovation: Credit default hybrid model for SME lending. Expert Syst Appl 61: 343–355. https://doi.org/10.1016/j.eswa.2016.05.029 doi: 10.1016/j.eswa.2016.05.029
    [25] Li X, Liu S, Li Z, et al. (2020) Flowscope: spotting money laundering based on graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 34: 4731–4738. https://doi.org/10.1609/aaai.v34i04.5906
    [26] Lou Y, Caruana R, Gehrke J, et al. (2013) Accurate intelligible models with pairwise interactions. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: 623–631. https://dl.acm.org/doi/abs/10.1145/2487575.2487579
    [27] Lundberg SM, Lee SI (2017) A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, 4768–4777. Available from: https://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html.
    [28] Mckee TE (2000) Developing a bankruptcy prediction model via rough sets theory. Intell Syst Account Financ Manage 9: 159–173. https://doi.org/fv22ks
    [29] Merton RC (1974) On the pricing of corporate debt: The risk structure of interest rates. J Fnanc 29: 449–470.
    [30] Mester LJ (1997) What's the point of credit scoring? Federal Reserve Bank of Philadelphia. Bus Rev 3: 3–16. Available from: https://fraser.stlouisfed.org/files/docs/historical/frbphi/businessreview/frbphil_rev_199709.pdf.
    [31] Min JH, Lee YC (2005) Bankruptcy prediction using support vector machine with optimal choice of kernel function parameters. Expert Syst Appl 28: 603–614 https://doi.org/10.1016/j.eswa.2004.12.008 doi: 10.1016/j.eswa.2004.12.008
    [32] Molnar C, König G, Herbinger J, et al. (2020) Pitfalls to avoid when interpreting machine learning models. Working Paper, University of Vienna. Available from: http://eprints.cs.univie.ac.at/6427/.
    [33] Odom MD, Sharda R (1990) A neural network model for bankruptcy prediction. In Joint Conference on Neural Networks, 163–168. IEEE Press, Piscataway, NJ. Available from: https://ieeexplore.ieee.org/abstract/document/5726669.
    [34] Opitz D, Maclin R (1999) Popular ensemble methods: An empirical study. J Artif Intell Res 11: 169–198.
    [35] Ribeiro MT, Singh S, Guestrin C (2016) "Why should I trust you?" Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: 1135–1144. Available from: https://dl.acm.org/doi/abs/10.1145/2939672.2939778.
    [36] Rudin C (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell 1: 206–215. https://doi.org/10.1038/s42256-019-0048-x doi: 10.1038/s42256-019-0048-x
    [37] Slack D, Hilgard S, Jia E, et al. (2020) Fooling LIME and SHAP: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 180–186. https://dl.acm.org/doi/abs/10.1145/3375627.3375830
    [38] Sudjianto A, Knauth W, Rahul S, et al. (2020) Unwrapping the black box of deep ReLU networks: Interpretability, diagnostics, and simplification. arXiv preprin, Cornell University. https://arXiv.org/abs/2011.04041
    [39] Sudjianto A, Zhang A (2021) Designing inherently interpretable machine learning models. In Proceedings of ACM ICAIF 2021 Workshop on Explainable AI in Finance. ACM, New York. https://arXiv.org/abs/2111.01743
    [40] Sudjianto A, Zhang A, Yang Z, et al. (2023) PiML toolbox for interpretable machine learning model development and validation. arXiv preprint arXiv. https://doi.org/10.48550/arXiv.2305.04214
    [41] U.S. Banking Regulatory Agencies (2011) The U.S. Office of the Comptroller of the Currency and the Board of Governors of Federal Reserve System. SR 11-7/OCC11-12: Supervisory Guidance on Model Risk Management. Washington, D.C. Available from: https://www.federalreserve.gov/supervisionreg/srletters/sr1107a1.pdf.
    [42] U.S. Banking Regulatory Agencies (2021) The U.S. Office of Comptroller of the Currency, the Board of Governors of the Federal Reserve System, the Federal Deposit Insurance Corporation, the Consumer Financial Protection Bureau, and the National Credit Union Administration. Request for Information and Comment on Financial Institutions' Use of Artificial Intelligence, Including Machine Learning. Washington, D.C. Available from: https://www.federalregister.gov/documents/2021/03/31/2021-06607/requestfor-information-and-comment-on-financial-institutions-use-of-artificialintelligence.
    [43] U.S. Office of the Comptroller of the Currency (2021) Comptroller's Handbook on Model Risk Management. Washington, D.C. Available from: https://www.occ.treas.gov/publicationsand-resources/publications/comptrollers-handbook/files/model-riskmanagement/index-model-risk-management.html.
    [44] Vassiliou PC (2013) Fuzzy semi-Markov migration process in credit risk. Fuzzy Sets and Syst 223: 39–58. https://doi.org/10.1016/j.fss.2013.02.016 doi: 10.1016/j.fss.2013.02.016
    [45] Yang Z, Zhang A, Sudjianto A (2021a) Enhancing explainability of neural networks through architecture constraints. IEEE T Neur Net Learn Syst 32: 2610–2621. https://doi.org/10.1109/TNNLS.2020.3007259 doi: 10.1109/TNNLS.2020.3007259
    [46] Yang Z, Zhang A, Sudjianto A (2021) GAMI-Net: An explainable neural network based on generalized additive models with structured interactions. Pattern Recogn 120: 108192. https://doi.org/10.1016/j.patcog.2021.10819 doi: 10.1016/j.patcog.2021.10819
    [47] Zhu Y, Xie C, Wang G J, et al. (2017) Comparison of individual, ensemble and integrated ensemble machine learning methods to predict China's SME credit risk in supply chain finance. Neural Comput Appl 28: 41–50. https://doi.org/10.1007/s00521-016-2304-x doi: 10.1007/s00521-016-2304-x
  • This article has been cited by:

    1. Öznur ÖZTUNÇ KAYMAK, Coding theory for h(x)-Fibonacci polynomials, 2024, 26, 1301-7985, 226, 10.25092/baunfbed.1347379
    2. Bahar Demirtürk, Combinatorial Analysis of k-Oresme and k-Oresme–Lucas Sequences, 2025, 17, 2073-8994, 697, 10.3390/sym17050697
  • Reader Comments
  • © 2024 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2787) PDF downloads(210) Cited by(1)

Figures and Tables

Figures(67)  /  Tables(7)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog