Implications of macroeconomic conditions on Romanian portfolio credit risk. A cost-sensitive ensemble learning methods comparison

Figures & data

Table 1. Confusion matrix (Source: Siddiqi, 2017).

		Actual
		Default	Non-default
Predicted	Default	True positive (TP)	False positive (FP)
	Non-default	False negative (FN)	True negative (TN)

Table 2. Cost functions used in the assessment (Source: authors’ computation).

	$\alpha$	$\beta$	Scenario
$f_0$	1	1		Baseline
$f_1$	1	0.5	Risk averse	The cost of misclassifying a default as a non-default becomes twice severe than cost of misclassifying a non-default as default
$f_2$	0.5	1	Risk tolerant	The cost of misclassifying a default as a non-default becomes twice smaller than cost of misclassifying a non-default as default
$f_3$	0.25	1	Risk taker	The cost of misclassifying a default as a non-default is equal to a 25% cost of misclassifying a non-default as default
$f_4$	1	0.25	Severe risk averse	The cost of misclassifying a default as a non-default becomes four times higher than cost of misclassifying a non-default as default

Figure 1. Evolution of interest rate for mortgage loans in RON and EUR (Source: National Bank of Romania).

Table 3. Portfolio allocation based on interest rate regime (Source: authors’ computation).

	Regime I	Regime II	Total
Credit Policy I	4999	3424	8423
Credit Policy II	3761	5336	9097
Total	8760	8760	17,520

Figure 2. The default profile in Regime I and Regime II (Source: authors’ computation).

Figure 3. Minimum classification error in Regime I (Source: authors’ computation).

Table 4. Champion model results in Regime I (Source: authors’ computation).

Cost functions	Ensemble method	Nr of learners	Max number of splits	Learning rate	TP	FN	FP	TN	Lost function
$f_0$	Logit Boost	481	30	0.01737	4330	50	93	4287	143.00
$f_1$	Random Forest	10	65		4331	49	106	4274	102.00
$f_2$	Random Forest	364	502		4285	95	65	4315	112.50
$f_3$	Gentle Boost	10	21	0.001516	4200	180	52	4328	97.00
$f_4$	Random Forest	58	110		4338	42	123	4257	72.75

Figure 4. Sensitivity and specificity in Regime I (Source: authors’ computation).

Table 5. Champion model results in Regime II (Source: authors’ computation).

Cost functions	Ensemble method	Nr of learners	Max number of splits	Learning rate	TP	FN	FP	TN	Lost function
$f_0$	Gentle Boost	456	379	0.73486	4058	322	96	4284	418
$f_1$	Gentle Boost	78	884	0.15407	4066	314	109	4271	368.5
$f_2$	Random Forest	497	219		4005	375	63	4317	250.5
$f_3$	Random Forest	11	19		3868	512	18	4362	146
$f_4$	Random Forest	10	3741		4128	252	331	4049	334.75

Figure 5. Sensitivity and specificity in Regime II (Source: authors’ computation).

Figure 6. True positive and true negative rate across scenarios for both regimes (Source: authors’ computation).

Table 6. Random Forest across scenarios for both regimes (Source: authors’ computation).

Cost functions	Ensemble method	Nr of learners	Max number of splits	TP	FN	FP	TN	Lost function
$f_0$	Random Forest	495	1843	3836	544	180	4200	724
$f_1$	Random Forest	486	817	3986	394	287	4093	537.5
$f_2$	Random Forest	492	2759	3753	627	103	4227	416.6
$f_3$	Random Forest	497	109	2155	2225	29	4351	585.2
$f_4$	Random Forest	476	403	4067	313	461	3919	428.25

Figure 7. Sensitivity and specificity in Portfolio (Source: authors’ computation).

Table 7. Accuracy and loss function (Source: authors’ computation).

	Regime I	Regime II	Portfolio	Regime I	Regime II	Portfolio
Accuracy
1. Risk taker	97.35%	93.95%	74.27%	3. Baseline	3. Baseline	4. Risk averse
2. Risk tolerant	98.17%	95.00%	91.62%	4. Risk averse	4. Risk averse	3. Baseline
3. Baseline	98.37%	95.23%	91.74%	2. Risk tolerant	2. Risk tolerant	2. Risk tolerant
4. Risk averse	98.23%	95.17%	92.23%	5. Severe risk averse	1. Risk taker	5. Severe risk averse
5. Severe risk averse	98.12%	93.34%	91.16%	1. Risk taker	5. Severe risk averse	1. Risk taker
Loss function
1. Risk taker	97.0	146.0	585.2	5. Severe risk averse	1. Risk taker	2. Risk tolerant
2. Risk tolerant	112.5	250.5	416.6	1. Risk taker	2. Risk Tolerant	5. Severe risk averse
3. Baseline	143.0	418.0	724.0	4. Risk averse	5. Severe Risk Averse	4. Risk averse
4. Risk averse	102.0	368.5	537.5	2. Risk tolerant	4. Risk Averse	1. Risk taker
5. Severe risk averse	72.8	334.8	428.3	3. Baseline	3. Baseline	3. Baseline

Siddiqi, N. (2017). Intelligent credit scoring. Building and implementing better credit risk scorecards (2nd ed.). Wiley Publishing House.

 Google Scholar