|
|
|
| R2011b Documentation → Financial Toolbox | |
Learn more about Financial Toolbox |
|
| Contents | Index |
Credit ratings rank borrowers according to their credit worthiness. Though this ranking is, in itself, useful, institutions are also interested in knowing how likely it is that borrowers in a particular rating category will be upgraded or downgraded to a different rating, and especially, how likely it is that they will default.
Transition probabilities offer one way to characterize the past changes in credit quality of obligors (typically firms), and are cardinal inputs to many risk management applications. Financial Toolbox software supports the estimation of transition probabilities using both cohort and duration (also known as hazard rate or intensity) approaches using transprob and related functions.
Note The sample dataset used throughout this chapter is simulated using a single transition matrix. No attempt is made to match historical trends in transition rates. |
The Data_TransProb.mat file contains sample credit ratings data.
load Data_TransProb
data(1:10,:)
ans =
'00010283' '10-Nov-1984' 'CCC'
'00010283' '12-May-1986' 'B'
'00010283' '29-Jun-1988' 'CCC'
'00010283' '12-Dec-1991' 'D'
'00013326' '09-Feb-1985' 'A'
'00013326' '24-Feb-1994' 'AA'
'00013326' '10-Nov-2000' 'BBB'
'00014413' '23-Dec-1982' 'B'
'00014413' '20-Apr-1988' 'BB'
'00014413' '16-Jan-1998' 'B' The sample data is formatted as a cell array with three columns. Each row contains an ID (column 1), a date (column 2), and a credit rating (column 3). The assigned credit rating corresponds to the associated ID on the associated date. All information corresponding to the same ID must be stored in contiguous rows. In this example, IDs, dates, and ratings are stored in string format, but you also can enter them in numeric format.
In this example, the simplest calling syntax for transprob passes the nRecords-by-3 cell array as the only input argument. The default startDate and endDate are the earliest and latest dates in the data. The default estimation algorithm is the duration method and 1-year transition probabilities are estimated:
transMat0 = transprob(data)
transMat0 =
93.1170 5.8428 0.8232 0.1763 0.0376 0.0012 0.0001 0.0017
1.6166 93.1518 4.3632 0.6602 0.1626 0.0055 0.0004 0.0396
0.1237 2.9003 92.2197 4.0756 0.5365 0.0661 0.0028 0.0753
0.0236 0.2312 5.0059 90.1846 3.7979 0.4733 0.0642 0.2193
0.0216 0.1134 0.6357 5.7960 88.9866 3.4497 0.2919 0.7050
0.0010 0.0062 0.1081 0.8697 7.3366 86.7215 2.5169 2.4399
0.0002 0.0011 0.0120 0.2582 1.4294 4.2898 81.2927 12.7167
0 0 0 0 0 0 0 100.0000It is recommended to provide explicit start and end dates. Otherwise the estimation window for two different datasets can differ, and the estimates might not be comparable. From this point, assume that the time window of interest is the 5-year period from the end of 1995 to the end of 2000. For comparisons, compute the estimates for this time window. First use the duration algorithm (default option), and then the cohort algorithm explicitly set.
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
transMat1 = transprob(data,'startDate',startDate,'endDate',endDate)
transMat2 = transprob(data,'startDate',startDate,'endDate',endDate,...
'algorithm','cohort')
transMat1 =
90.6236 7.9051 1.0314 0.4123 0.0210 0.0020 0.0003 0.0043
4.4780 89.5558 4.5298 1.1225 0.2284 0.0094 0.0009 0.0754
0.3983 6.1164 87.0641 5.4801 0.7637 0.0892 0.0050 0.0832
0.1029 0.8572 10.7918 83.0204 3.9971 0.7001 0.1313 0.3992
0.1043 0.3745 2.2962 14.0954 78.9840 3.0013 0.0463 1.0980
0.0113 0.0544 0.7055 3.2925 15.4350 75.5988 1.8166 3.0860
0.0044 0.0189 0.1903 1.9743 6.2320 10.2334 75.9983 5.3484
0 0 0 0 0 0 0 100.0000
transMat2 =
90.1554 8.5492 0.9067 0.3886 0 0 0 0
4.9512 88.5221 5.1763 1.0503 0.2251 0 0 0.0750
0.2770 6.6482 86.2188 6.0942 0.6233 0.0693 0 0.0693
0.0794 0.8737 11.6759 81.6521 4.3685 0.7943 0.1589 0.3971
0.1002 0.4008 1.9038 15.4309 77.8557 3.4068 0 0.9018
0 0 0.2262 2.4887 17.4208 74.2081 2.2624 3.3937
0 0 0.7576 1.5152 6.0606 10.6061 75.0000 6.0606
0 0 0 0 0 0 0 100.0000By default, the cohort algorithm internally gets yearly snapshots of the credit ratings, but the number of snapshots per year is definable using the parameter/value pair snapsPerYear. To get the estimates using quarterly snapshots:
transMat3 = transprob(data,'startDate',startDate,'endDate',endDate,...
'algorithm','cohort','snapsPerYear',4)
transMat3 =
90.4765 8.0881 1.0072 0.4069 0.0164 0.0015 0.0002 0.0032
4.5949 89.3216 4.6489 1.1239 0.2276 0.0074 0.0007 0.0751
0.3747 6.3158 86.7380 5.6344 0.7675 0.0856 0.0040 0.0800
0.0958 0.7967 11.0441 82.6138 4.1906 0.7230 0.1372 0.3987
0.1028 0.3571 2.3312 14.4954 78.4276 3.1489 0.0383 1.0987
0.0084 0.0399 0.6465 3.0962 16.0789 75.1300 1.9044 3.0956
0.0031 0.0125 0.1445 1.8759 6.2613 10.7022 75.6300 5.3705
0 0 0 0 0 0 0 100.0000Both duration and cohort compute 1-year transition probabilities by default, but the time interval for the transitions is definable using the parameter/value pair transInterval. For example, to get the 2-year transition probabilities using the cohort algorithm with the same snapshot periodicity and estimation window:
transMat4 = transprob(data,'startDate',startDate,'endDate',endDate,...
'algorithm','cohort','snapsPerYear',4,'transInterval',2)
transMat4 =
82.2358 14.6092 2.2062 0.8543 0.0711 0.0074 0.0011 0.0149
8.2803 80.4584 8.3606 2.2462 0.4665 0.0316 0.0030 0.1533
0.9604 11.1975 76.1729 9.7284 1.5322 0.2044 0.0162 0.1879
0.2483 2.0903 18.8440 69.5145 6.9601 1.2966 0.2329 0.8133
0.2129 0.8713 5.4893 23.5776 62.6438 4.9464 0.1390 2.1198
0.0378 0.1895 1.7679 7.2875 24.9444 57.1783 2.8816 5.7132
0.0154 0.0716 0.6576 4.2157 11.4465 16.3455 57.4078 9.8399
0 0 0 0 0 0 0 100.0000The dataset data from Data_TransProb.mat contains sample credit ratings using the default rating scale {'AAA', 'AA','A', 'BBB', 'BB', 'B', 'CCC', 'D'}. It also contains the dataset dataIGSG with ratings investment grade ('IG'), speculative grade ('SG'), and default ('D'). To estimate the transition matrix for this dataset, use the labels argument.
load Data_TransProb
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
dataIGSG(1:10,:)
transMatIGSG = transprob(dataIGSG,'labels',{'IG','SG','D'},...
'startDate',startDate,'endDate',endDate)
ans =
'00011253' '04-Apr-1983' 'IG'
'00012751' '17-Feb-1985' 'SG'
'00012751' '19-May-1986' 'D'
'00014690' '17-Jan-1983' 'IG'
'00012144' '21-Nov-1984' 'IG'
'00012144' '25-Mar-1992' 'SG'
'00012144' '07-May-1994' 'IG'
'00012144' '23-Jan-2000' 'SG'
'00012144' '20-Aug-2001' 'IG'
'00012937' '07-Feb-1984' 'IG'
transMatIGSG =
98.1986 1.5179 0.2835
8.5396 89.4891 1.9713
0 0 100.0000There is another dataset, dataIGSGnum, with the same information as dataIGSG, except the ratings are mapped to a numeric scale where 'IG'=1, 'SG'=2, and 'D'=3. To estimate the transition matrix, use the labels optional argument specifying the numeric scale as a cell array.
dataIGSGnum(1:10,:)
% Note {1,2,3} and num2cell(1:3) are equivalent; num2cell is convenient
% when the number of ratings is larger
transMatIGSGnum = transprob(dataIGSGnum,'labels',{1,2,3},...
'startDate',startDate,'endDate',endDate)
ans =
'00011253' '04-Apr-1983' [1]
'00012751' '17-Feb-1985' [2]
'00012751' '19-May-1986' [3]
'00014690' '17-Jan-1983' [1]
'00012144' '21-Nov-1984' [1]
'00012144' '25-Mar-1992' [2]
'00012144' '07-May-1994' [1]
'00012144' '23-Jan-2000' [2]
'00012144' '20-Aug-2001' [1]
'00012937' '07-Feb-1984' [1]
transMatIGSGnum =
98.1986 1.5179 0.2835
8.5396 89.4891 1.9713
0 0 100.0000Any time the input dataset contains ratings not included in the default rating scale {'AAA', 'AA', 'A', 'BBB', 'BB', 'B', 'CCC', 'D'}, the full rating scale needs to be specified using the labels optional argument. For example, if the dataset contains ratings 'AAA', ..., 'CCC, 'D', and 'NR' (not rated), use labels with this cell array {'AAA', 'AA', 'A','BBB','BB','B','CCC','D','NR'}.
Transition probability estimates are sensitive to the length of the estimation window. When the estimation window is small, the estimates only capture recent credit events, and these can change significantly from one year to the next. These are called point-in-time (PIT) estimates. In contrast, a large time window yields fairly stable estimates that average transition rates over a longer period of time. These are called through-the-cycle (TTC) estimates.
The estimation of PIT probabilities requires repeated calls to transprob with a rolling estimation window. Use transprobprep every time repeated calls to transprob are required. transprobprep performs a preprocessing step on the raw dataset that is independent of the estimation window. The benefits of transprobprep are greater as the number of repeated calls to transprob increases. Also, the performance gains from transprobprep are more significant for the cohort algorithm.
load Data_TransProb
prepData = transprobprep(data);
Years = 1991:2000;
nYears = length(Years);
nRatings = length(prepData.ratingsLabels);
transMatPIT = zeros(nRatings,nRatings,nYears);
sampleTotals(nYears,1) = struct('totalsVec',[],'totalsMat',[],...
'algorithm','cohort');
for t = 1:nYears
startDate = ['31-Dec-' num2str(Years(t)-1)];
endDate = ['31-Dec-' num2str(Years(t))];
[transMatPIT(:,:,t),sampleTotals(t)] = transprob(prepData,...
'startDate',startDate,'endDate',endDate);
end
Here is the PIT transition matrix for 1993. Recall that the sample dataset contains simulated credit migrations so the PIT estimates in this example do not match actual historical transition rates.
transMatPIT(:,:,Years==1993)
ans =
95.3193 4.5999 0.0802 0.0004 0.0002 0.0000 0.0000 0.0000
2.0631 94.5931 3.3057 0.0254 0.0126 0.0002 0.0000 0.0000
0.0237 2.1748 95.5901 1.4700 0.7284 0.0131 0.0000 0.0000
0.0003 0.0372 3.2585 95.2914 1.3876 0.0250 0.0001 0.0000
0.0000 0.0005 0.0657 3.8292 92.7474 3.3459 0.0111 0.0001
0.0000 0.0001 0.0128 0.7977 8.0926 90.4897 0.5958 0.0113
0.0000 0.0000 0.0005 0.0459 0.5026 11.1621 84.9315 3.3574
0 0 0 0 0 0 0 100.0000A structure array stores the sampleTotals optional output from transprob. The sampleTotals structure contains summary information on the total time spent on each rating, and the number of transitions out of each rating, for each year under consideration. For more information on the sampleTotals structure, see Algorithms.
As an example, the sampleTotals structure for 1993 is used here. The total time spent on each rating is stored in the totalsVec field of the structure. The total transitions out of each rating are stored in the totalsMat field. A third field, algorithm, indicates the algorithm used to generate the structure.
sampleTotals(Years==1993).totalsVec
sampleTotals(Years==1993).totalsMat
sampleTotals(Years==1993).algorithm
ans =
144.4411 230.0356 262.2438 204.9671 246.1315 147.0767 54.9562 215.1479
ans =
0 7 0 0 0 0 0 0
5 0 8 0 0 0 0 0
0 6 0 4 2 0 0 0
0 0 7 0 3 0 0 0
0 0 0 10 0 9 0 0
0 0 0 1 13 0 1 0
0 0 0 0 0 7 0 2
0 0 0 0 0 0 0 0
ans =
durationTo get the TTC transition matrix, pass the sampleTotals structure array to transprobbytotals. Internally, transprobbytotals aggregates the information in the sampleTotals structures to get the total time spent on each rating over the 10 years considered in this example, and the total number of transitions out of each rating during the same period. transprobbytotals uses the aggregated information to get the TTC matrix, or average 1-year transition matrix.
transMatTTC = transprobbytotals(sampleTotals)
transMatTTC =
92.8544 6.1068 0.7463 0.2761 0.0123 0.0009 0.0001 0.0032
2.9399 92.2329 3.8394 0.7349 0.1676 0.0050 0.0004 0.0799
0.2410 4.5963 90.3468 3.9572 0.6909 0.0521 0.0025 0.1133
0.0530 0.4729 7.9221 87.2751 3.5075 0.4650 0.0791 0.2254
0.0460 0.1636 1.1873 9.3442 85.4305 2.9520 0.1150 0.7615
0.0031 0.0152 0.2608 1.5563 10.4468 83.8525 1.9771 1.8882
0.0009 0.0041 0.0542 0.8378 2.9996 7.3614 82.4758 6.2662
0 0 0 0 0 0 0 100.0000The same TTC matrix could be obtained with a direct call to transprob, setting the estimation window to the 10 years under consideration. But it is much more efficient to use the sampleTotals structures, whenever they are available. (Note, for the duration algorithm, these alternative workflows can result in small numerical differences in the estimates whenever leap years are part of the sample.)
In Estimate Transition Probabilities, a 1-year transition matrix is estimated using the 5-year time window from 1996 through 2000. This is another example of a TTC matrix and this can also be computed using the sampleTotals structure array.
transprobbytotals(sampleTotals(Years>=1996 Years<=2000))
ans =
90.6239 7.9048 1.0313 0.4123 0.0210 0.0020 0.0003 0.0043
4.4776 89.5565 4.5294 1.1224 0.2283 0.0094 0.0009 0.0754
0.3982 6.1159 87.0651 5.4797 0.7636 0.0892 0.0050 0.0832
0.1029 0.8571 10.7909 83.0218 3.9968 0.7001 0.1313 0.3991
0.1043 0.3744 2.2960 14.0947 78.9851 3.0012 0.0463 1.0980
0.0113 0.0544 0.7054 3.2922 15.4341 75.6004 1.8165 3.0858
0.0044 0.0189 0.1903 1.9742 6.2318 10.2332 75.9990 5.3482
0 0 0 0 0 0 0 100.0000By varying the start and end dates, the amount of data considered for the estimation is changed, but the output still contains, by default, 1-year transition probabilities. You can change the default behavior by specifying the transInterval argument, as illustrated in Estimate Transition Probabilities.
However, when t-year transition probabilities are required for a whole range of values of t, for example, 1-year, 2-year, 3-year, 4-year, and 5-year transition probabilities, it is more efficient to call transprob once to get the optional output sampleTotals. You can use the same sampleTotals structure can be used to get the t-year transition matrix for any transition interval t. Given a sampleTotals structure and a transition interval, you can get the corresponding transition matrix by using transprobbytotals.
load Data_TransProb
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
[~,sampleTotals] = transprob(data,'startDate', ...
startDate, 'endDate',endDate);
DefProb = zeros(7,5);
for t = 1:5
transMatTemp = transprobbytotals(sampleTotals,'transInterval',t);
DefProb(:,t) = transMatTemp(1:7,8);
end
DefProb
DefProb =
0.0043 0.0169 0.0377 0.0666 0.1033
0.0754 0.1542 0.2377 0.3265 0.4213
0.0832 0.1936 0.3276 0.4819 0.6536
0.3992 0.8127 1.2336 1.6566 2.0779
1.0980 2.1189 3.0668 3.9468 4.7644
3.0860 5.6994 7.9281 9.8418 11.4963
5.3484 9.8053 13.5320 16.6599 19.2964transprob also returns the idTotals structure array which contains, for each ID, or company, the total time spent on each rating, and the total transitions out of each rating. For more information on the the idTotals structure, see Algorithms. The idTotals structure is similar to the sampleTotals structures (see Estimate Point-in-Time and Through-the-Cycle Probabilities), but idTotals has the information at an ID level. Because most companies only migrate between very few ratings, the numeric arrays in idTotals are stored as sparse arrays to reduce memory requirements.
You can use the idTotals structure array to estimate confidence intervals for the transition probabilities using a bootstrapping procedure, as the following example demonstrates. To do this, call transprob and keep the third output argument, idTotals. The idTotals fields are displayed for the last company in the sample. Within the estimation window, this company spends almost a year as 'AA' and it is then upgraded to 'AAA'.
load Data_TransProb
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
[transMat,~,idTotals] = transprob(data,...
'startDate',startDate,'endDate',endDate);
% Total time spent on each rating
full(idTotals(end).totalsVec)
% Total transitions out of each rating
full(idTotals(end).totalsMat)
% Algorithm
idTotals(end).algorithm
ans =
4.0820 0.9180 0 0 0 0 0 0
ans =
0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
ans =
durationNext, use bootstrp from Statistics Toolbox with transprobbytotals as the bootstrap function and idTotals as the data to sample from. Each bootstrap sample corresponds to a dataset made of companies sampled with replacement from the original data. However, you do not have to draw companies from the original data, because a bootstrap idTotals sample contains all the information required to compute the transition probabilities. transprobbytotals aggregates all structures in each bootstrap idTotals sample and finds the corresponding transition matrix.
To estimate 95% confidence intervals for the transition matrix and display the probabilities of default together with its upper and lower confidence bounds:
PD = transMat(1:7,8);
bootstat = bootstrp(100,@(totals)transprobbytotals(totals),idTotals);
ci = prctile(bootstat,[2.5 97.5]); % 95% confidence
CIlower = reshape(ci(1,:),8,8);
CIupper = reshape(ci(2,:),8,8);
PD_LB = CIlower(1:7,8);
PD_UB = CIupper(1:7,8);
[PD_LB PD PD_UB]
ans =
0.0004 0.0043 0.0106
0.0028 0.0754 0.2192
0.0126 0.0832 0.2180
0.1659 0.3992 0.6617
0.5703 1.0980 1.7260
1.7264 3.0860 4.7602
1.7678 5.3484 9.5055Credit rating scales can be more or less granular. For example, there are ratings with qualifiers (such as 'AA+', 'BB-', etc.), whole ratings ('AA', 'BB', etc.), and investment or speculative grade ('IG', 'SG') categories. Given a dataset with credit ratings at a more granular level, transition probabilities for less granular categories can be of interest. For example, you might be interested in a transition matrix for investment and speculative grades given a dataset with whole ratings. Use transprobgrouptotals for this evaluation, as illustrated in the following examples. The sample dataset data has whole credit ratings:
load Data_TransProb
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
data(1:5,:)
ans =
'00010283' '10-Nov-1984' 'CCC'
'00010283' '12-May-1986' 'B'
'00010283' '29-Jun-1988' 'CCC'
'00010283' '12-Dec-1991' 'D'
'00013326' '09-Feb-1985' 'A' A call to transprob returns the transition matrix and totals structures for the eight ('AAA' to 'D') whole credit ratings. The array with number of transitions out of each credit rating is displayed after the call to transprob:
[transMat,sampleTotals,idTotals] = transprob(data,'startDate',startDate,...
'endDate',endDate);
sampleTotals.totalsMat
ans =
0 67 7 3 0 0 0 0
67 0 68 15 3 0 0 1
4 101 0 93 11 1 0 1
1 7 163 0 62 10 2 5
1 3 16 168 0 37 0 11
0 0 2 10 83 0 10 14
0 0 0 2 8 16 0 7
0 0 0 0 0 0 0 0Next, use transprobgrouptotals to group whole ratings into investment and speculative grades. This function takes a totals structure as the first argument. The second argument indicates the edges between rating categories. In this case, ratings 1 through 4 ('AAA' through 'BBB') correspond to the first category ('IG'), ratings 5 through 7 ('BB' through 'CCC') to the second category ('SG'), and rating 8 ('D') is a category of its own. transprobgrouptotals adds up the total time spent on ratings that belong to the same category. For example, total times spent on 'AAA' through 'BBB' are added up as the total time spent on 'IG'. transprobgrouptotals also adds up the total number of transitions between any 'IG' rating and any 'SG' rating, for example, a credit migration from 'BBB' to 'BB'.
The grouped totals can then be passed to transprobbytotals to obtain the transition matrix for investment and speculative grades. Note that both totalsMat and the new transition matrix are both 3-by-3, corresponding to the grouped categories 'IG', 'SG', and 'D'.
sampleTotalsIGSG = transprobgrouptotals(sampleTotals,[4 7 8])
transMatIGSG = transprobbytotals(sampleTotalsIGSG)
sampleTotalsIGSG =
totalsVec: [4.8591e+003 1.5034e+003 1.1621e+003]
totalsMat: [3x3 double]
algorithm: 'duration'
transMatIGSG =
98.1591 1.6798 0.1611
12.3228 85.6961 1.9811
0 0 100.0000When a totals structure array is passed to transprobgrouptotals, this function groups each structure in the array individually and preserves sparsity, if the fields in the input structures are sparse. One way to exploit this feature is to compute confidence intervals for the investment grade default rate and the speculative grade default rate (see also Estimate Bootstrap Confidence Intervals).
PDIGSG = transMatIGSG(1:2,3);
idTotalsIGSG = transprobgrouptotals(idTotals,[4 7 8]);
bootstat = bootstrp(100,@(totals)transprobbytotals(totals),idTotalsIGSG);
ci = prctile(bootstat,[2.5 97.5]); % 95% confidence
CIlower = reshape(ci(1,:),3,3);
CIupper = reshape(ci(2,:),3,3);
PDIGSG_LB = CIlower(1:2,3);
PDIGSG_UB = CIupper(1:2,3);
[PDIGSG_LB PDIGSG PDIGSG_UB]
ans =
0.0603 0.1611 0.2538
1.3470 1.9811 2.6195Transition probabilities and the number of transitions between ratings are usually reported without the 'D' ('Default') row. For example, a credit report can contain the following table, indicating the number of issuers starting in each rating (first column), and the number of transitions between ratings (remaining columns):
Initial AAA AA A BBB BB B CCC D
AAA 98 88 9 1 0 0 0 0 0
AA 389 0 368 19 2 0 0 0 0
A 1165 1 21 1087 56 0 0 0 0
BBB 1435 0 2 89 1289 45 8 0 2
BB 915 0 0 1 60 776 73 2 3
B 867 0 0 1 7 88 715 39 17
CCC 112 0 0 0 1 3 34 61 13You can store the information in this table in a totals structure compatible with the cohort algorithm. For more information on the cohort algorithm and the totals structure, see Algorithms. Note that the totalsMat field is a nonsquare array in this case.
% Define totals structure
totals.totalsVec = [98 389 1165 1435 915 867 112];
totals.totalsMat = [
88 9 1 0 0 0 0 0;
0 368 19 2 0 0 0 0;
1 21 1087 56 0 0 0 0;
0 2 89 1289 45 8 0 2;
0 0 1 60 776 73 2 3;
0 0 1 7 88 715 39 17;
0 0 0 1 3 34 61 13];
totals.algorithm = 'cohort';
transprobbytotals and transprobgrouptotals accept totals inputs with nonsquare totalsMat fields. To get the transition matrix corresponding to the previous table, and to group ratings into investment and speculative grade with the corresponding matrix:
transMat = transprobbytotals(totals)
% Group into IG/SG and get IG/SG transition matrix
totalsIGSG = transprobgrouptotals(totals,[4 7]);
transMatIGSG = transprobbytotals(totalsIGSG)
transMat =
89.7959 9.1837 1.0204 0 0 0 0 0
0 94.6015 4.8843 0.5141 0 0 0 0
0.0858 1.8026 93.3047 4.8069 0 0 0 0
0 0.1394 6.2021 89.8258 3.1359 0.5575 0 0.1394
0 0 0.1093 6.5574 84.8087 7.9781 0.2186 0.3279
0 0 0.1153 0.8074 10.1499 82.4683 4.4983 1.9608
0 0 0 0.8929 2.6786 30.3571 54.4643 11.6071
transMatIGSG =
98.2183 1.7169 0.0648
3.6959 94.5618 1.7423
The idTotals output from transprob can also be exploited to update the transition probability estimates after removing some outlier information. For more information on idTotals, see Algorithms. For example, if you know that the credit rating migration information for the 4th and 27th companies in the data have problems, you can remove those companies and efficiently update the transition probabilities as follows:
load Data_TransProb
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
[transMat,~,idTotals] = transprob(data,'startDate', ...
startDate, 'endDate',endDate);
transMat
transMat =
90.6236 7.9051 1.0314 0.4123 0.0210 0.0020 0.0003 0.0043
4.4780 89.5558 4.5298 1.1225 0.2284 0.0094 0.0009 0.0754
0.3983 6.1164 87.0641 5.4801 0.7637 0.0892 0.0050 0.0832
0.1029 0.8572 10.7918 83.0204 3.9971 0.7001 0.1313 0.3992
0.1043 0.3745 2.2962 14.0954 78.9840 3.0013 0.0463 1.0980
0.0113 0.0544 0.7055 3.2925 15.4350 75.5988 1.8166 3.0860
0.0044 0.0189 0.1903 1.9743 6.2320 10.2334 75.9983 5.3484
0 0 0 0 0 0 0 100.0000
nIDs = length(idTotals);
keepInd = setdiff(1:nIDs,[4 27]);
transMatNoOutlier = transprobbytotals(idTotals(keepInd))
transMatNoOutlier =
90.6241 7.9067 1.0290 0.4124 0.0211 0.0020 0.0003 0.0043
4.4917 89.5918 4.4779 1.1240 0.2288 0.0094 0.0009 0.0756
0.3990 6.1220 87.0530 5.4841 0.7643 0.0893 0.0050 0.0833
0.1030 0.8576 10.7909 83.0207 3.9971 0.7001 0.1313 0.3992
0.1043 0.3746 2.2960 14.0955 78.9840 3.0013 0.0463 1.0980
0.0113 0.0544 0.7054 3.2925 15.4350 75.5988 1.8166 3.0860
0.0044 0.0189 0.1903 1.9743 6.2320 10.2334 75.9983 5.3484
0 0 0 0 0 0 0 100.0000Deciding which companies to remove is a case-by-case situation. Reasons to remove a company can include a typo in one of the ratings history, or an unusual migration between ratings whose impact on the transition probability estimates must be measured. Note that transprob does not reorder the companies in any way. The ordering of companies in the input data is the same as the ordering in the idTotals array.
You can use idTotals efficiently to get estimates over different segments of the sample. For more information on idTotals, see Algorithms. For example, assume that the companies in the example are grouped into three geographic regions and that the companies were grouped by geographic regions previously, so that the first 340 companies correspond to the first region, the next 572 companies to the second region, and the rest to the third region. You can efficiently get transition probabilities for each region as follows:
load Data_TransProb
startDate = '31-Dec-1995';
endDate = '31-Dec-2000';
[~,~,idTotals] = transprob(data,'startDate', ...
startDate, 'endDate',endDate);
n1 = 340;
n2 = 572;
transMatG1 = transprobbytotals(idTotals(1:n1))
transMatG2 = transprobbytotals(idTotals(n1+1:n1+n2))
transMatG3 = transprobbytotals(idTotals(n1+n2+1:end))
transMatG1 =
90.8299 7.6501 0.3178 1.1700 0.0255 0.0044 0.0021 0.0002
4.3572 89.0262 5.7838 0.8039 0.0245 0.0029 0.0013 0.0001
0.7066 6.7567 86.6320 5.4950 0.3721 0.0252 0.0101 0.0023
0.0626 1.3688 10.3895 83.5022 3.6823 0.6466 0.3084 0.0396
0.0256 0.7884 2.6970 13.7857 78.8321 2.8310 0.0561 0.9842
0.0026 0.1095 0.4280 3.5204 21.1437 72.9230 1.6456 0.2273
0.0005 0.0216 0.0730 0.4574 4.9586 4.2821 80.3062 9.9006
0 0 0 0 0 0 0 100.0000
transMatG2 =
90.5798 8.4877 0.8202 0.0884 0.0132 0.0011 0.0000 0.0096
4.1999 90.0371 3.8657 1.4744 0.2144 0.0128 0.0001 0.1956
0.3022 5.9869 86.7128 5.5526 1.0411 0.1902 0.0015 0.2127
0.0204 0.5606 10.9342 82.9195 4.0123 0.7398 0.0059 0.8073
0.0089 0.3338 2.1185 16.6496 76.2395 3.1241 0.0261 1.4995
0.0013 0.0465 0.6710 2.4731 14.7281 76.7378 1.2993 4.0428
0.0002 0.0080 0.0681 0.4598 4.1324 8.4380 80.9092 5.9843
0 0 0 0 0 0 0 100.0000
transMatG3 =
90.5655 7.5408 1.5288 0.3369 0.0258 0.0015 0.0003 0.0004
4.8073 89.3842 4.4865 0.9582 0.3509 0.0095 0.0009 0.0025
0.3153 5.8771 87.6353 5.4101 0.7160 0.0322 0.0052 0.0088
0.1995 0.8625 10.8682 82.8717 4.1423 0.6903 0.1565 0.2090
0.2465 0.1091 2.1558 12.0289 81.5803 3.0057 0.0616 0.8122
0.0227 0.0400 0.9380 4.3175 12.3632 75.9429 2.5766 3.7991
0.0149 0.0180 0.3414 3.6918 8.1414 13.6010 70.7254 3.4661
0 0 0 0 0 0 0 100.0000This example shows how to aggregate estimates from two (or more) datasets. It is possible that two datasets, coming from two different databases, must be considered for the estimation of the transition probabilities. Also, if a dataset is too large and cannot be loaded into memory, the dataset can be split into two (or more) datasets. In these cases, it is simple to apply transprob to each individual dataset, and then get the final estimates corresponding to the aggregated data with a call to transprobbytotals at the end.
For demonstration, the dataset data is artificially split into two sections in this example. In practice the two datasets would come from different files or databases. When aggregating multiple datasets, the history of a company cannot be split across datasets. You can verify that this condition is satisfied for the arbitrarily chosen cut-off point.
load Data_TransProb
cutoff = 2099;
data(cutoff-5:cutoff,:)
data(cutoff+1:cutoff+6,:)
ans =
'00011166' '24-Aug-1995' 'BBB'
'00011166' '25-Jan-1997' 'A'
'00011166' '01-Feb-1998' 'AA'
'00014878' '15-Mar-1983' 'B'
'00014878' '21-Sep-1986' 'BB'
'00014878' '17-Jan-1998' 'BBB'
ans =
'00012043' '09-Feb-1985' 'BBB'
'00012043' '03-Jan-1988' 'A'
'00012043' '15-Jan-1994' 'AAA'
'00011157' '24-Jun-1984' 'A'
'00011157' '09-Dec-1999' 'BBB'
'00011157' '28-Mar-2001' 'A' When working with multiple datasets, it is important to set the start and end dates explicitly. Otherwise, the estimation window will differ for each dataset because the default start and end dates used by transprob are the earliest and latest dates found in the input data.
startDate = '31-Dec-1995'; endDate = '31-Dec-2000';
In practice, this is the point where you can read in the first dataset. Now, the dataset is already obtained. Call transprob with the first dataset and the explicit start and end dates. Keep only the sampleTotals output. For details on sampleTotals, seeAlgorithms.
[~,sampleTotals(1)] = transprob(data(1:cutoff,:),... 'startDate',startDate,'endDate',endDate);
Repeat for the remaining datasets. Note the different sampleTotals structures are stored in a structured array.
[~,sampleTotals(2)] = transprob(data(cutoff+1:end,:),... 'startDate',startDate,'endDate',endDate);
To get the transition matrix corresponding to the aggregated dataset, use transprobbytotals. When the totals input is a structure array, transprobbytotals aggregates the information over all structures, and returns a single transition matrix.
transMatAggr = transprobbytotals(sampleTotals)
transMatAggr =
90.6236 7.9051 1.0314 0.4123 0.0210 0.0020 0.0003 0.0043
4.4780 89.5558 4.5298 1.1225 0.2284 0.0094 0.0009 0.0754
0.3983 6.1164 87.0641 5.4801 0.7637 0.0892 0.0050 0.0832
0.1029 0.8572 10.7918 83.0204 3.9971 0.7001 0.1313 0.3992
0.1043 0.3745 2.2962 14.0954 78.9840 3.0013 0.0463 1.0980
0.0113 0.0544 0.7055 3.2925 15.4350 75.5988 1.8166 3.0860
0.0044 0.0189 0.1903 1.9743 6.2320 10.2334 75.9983 5.3484
0 0 0 0 0 0 0 100.0000As a sanity check, for this example you can verify that the aggregation procedure yields the same estimates (up to numerical differences) as estimating the probabilities directly over the entire sample:
transMatWhole = transprob(data,'startDate',startDate,'endDate',endDate)
aggError = max(max(abs(transMatAggr - transMatWhole)))
transMatWhole =
90.6236 7.9051 1.0314 0.4123 0.0210 0.0020 0.0003 0.0043
4.4780 89.5558 4.5298 1.1225 0.2284 0.0094 0.0009 0.0754
0.3983 6.1164 87.0641 5.4801 0.7637 0.0892 0.0050 0.0832
0.1029 0.8572 10.7918 83.0204 3.9971 0.7001 0.1313 0.3992
0.1043 0.3745 2.2962 14.0954 78.9840 3.0013 0.0463 1.0980
0.0113 0.0544 0.7055 3.2925 15.4350 75.5988 1.8166 3.0860
0.0044 0.0189 0.1903 1.9743 6.2320 10.2334 75.9983 5.3484
0 0 0 0 0 0 0 100.0000
aggError =
2.8422e-014
 | Credit Rating | Forecast Corporate Default Rates |  |
View demos and recorded presentations led by industry experts.
Now On Demand
Network with industry peers and learn the latest applications of the leading software product for computational finance.
| © 1984-2012- The MathWorks, Inc. - Site Help - Patents - Trademarks - Privacy Policy - Preventing Piracy - RSS |
