Scene 1

Scene 2

Credit Card fraud detection is a necessity for banks in the current time to reduce time-consuming manual reviews, costly chargebacks and fees, and denial of legitimate transactions. We need to develop a machine learning model to detect fraudulent transactions based on the historical transactional data of customers with a pool of merchants. We have to analyse the business impact of these fraudulent transactions and recommend the optimal ways that the bank can adopt to mitigate the fraud risks. Also we need to do the cost benefit analysis of the model we have created.

Scene 3

The various steps used in the case study to reach to the goal are as follows: 1 1. Understanding the data set and data preparation. 2 2. Performing     Exploratory   data analysis and making interpretations 3 3.  Balancing the data as the data is highly imbalanced 4 4. Performing Test- train split and building the logistic regression model. 5 5. Calculating all the metrics score and comparing the train and test set 6 6. We will also create the decision tree model and evaluate the model. 7 7. Lastly we will do the Cost benefit Analysis and we will find out how much are we saving because of the model.  8

Scene 4

is_fraud count percentage 0 0 1842743 99.478999 1 1 9651 0.521001

The fraud and non- fraud distribution of transaction

We see that the data is highly Imbalanced and the no. of  non-fraud transactions are  much more.

Even visualizing we see the imbalance of the data.

Distribution of fraud and non-fraud transactions.

Scene 5

ws_transport
wocery_pos
Mme
opping_pos
entertainment
Bod_dining
JErsonaI_care
tealth fitness
m sc_pos
msc net
wocery_net
%nday
Sunday
Tuesday
Friday
Th ursday
'Bhdnesday

We see that the highest credit card transactions are of category gas transport and then groceries. Also Monday and Sunday of the week have the highest transactions.

2019-12
2020-12
2020-08
2020-06
2019-08
2019-07
2019-06
2020-07
2020-05
2020-03
2020-11
2019-05
2019-03
2019-09
2019-11
2020-09
2020-10
2019-10
2019-04
2020-04
2019-01
2020-01
2019-02
2020-02

Females do more transaction than the males. We see that maximum transactions are done in the month of December in both the years.

Category and day distribution of transactions

Scene 6

Most of the customers are the age of 33 to 57. The minimum age is 14 and the maximum age is 96 of the customers using credit card

We see that the ratio of female customers are more than the male customers using credit card

Count 1852394 Mean 46.26 Std 17.41 Min 14 25% 33 50% 44 75% 57 Max 96

Gender gender_count percent 0 F 1014749 54.78 1 M 837645 45.22

Scene 7

The mean of fraud transactions is way more than non- fraud transactions. This shows that fraud transactions need to be taken care of. The mean of  amount of fraud transactions is $ 530 where as that of non fraud is $ 67

700000
00000
tooooo
400000
moooo
100000
200
Overall Amt Distribution
400 ID00
Transaction Amount
700000
00000
tcoooo
400000
moooo
100000
1000
1200
1400
200
Non Fraud Amt Distribution
400 ID00
Transaction Amount
1200
1400
200
Fraud Amt Distribution
400 800 ID00
Transaction Amount
1200
1400

We see that the non fraud transactions ranges between amount of 1 - 350. While the fraud transactions ranges between the amount of 1 - 1300 approximately

Row Type Overall Amt Distribution Non Fraud Amt Distribution Fraud Amt Distribution count 1852394 1842743 9651 mean 70.06 67.65 530.66 std 159.25 153.55 391.03 min 1 1 1.06 50% 47.45 47.24 390 95% 195.34 189.59 1084.09 99.90% 1517.24 1519.62 1293.13 max 28948.9 28948.9 1376.04

Scene 8

When we checked the value of the outliers they were all non fraud transactions and spent on travel.

025002
003 0.017 16 0074
002
o au 4,36 0.2 034 018 022 0.21 0170.081 0.1900140m2 002-oooan013
0 0180038-0078001800042014002m002m02800160017-0017002mmswoouoog.m10003.002D012.029
03000510144.01
0 0140022-0.36 00380.09
0020019 02 00780.05200
0 00180089-0.11
002Æ.m3N
0.0100014m"6e-W001a.001.ooum0000350.01000az01
0 0640,'1-0089-0.11 0.1
csao .0510 062005900430 0520.0540 -o 030 .moermamoo 036001 0050020003
086 01 -0.1 00210033
0014-0.089011 01 -0.0
002000030.21 00220085 0059 01 .0082
0860.0820060068007M'.0760_0840048001
0.1
o 07-0 0850.0880097-0 0026 0090mow.01
00140 m250.170mzm_06300760N30.0740mo.073007
0 01 n
038.0001n19
o 0.07300810 .00 oooaoo
07800860
O 09.0 2
.0_06 01 .0084 01 0.097.o_cm00810086009
o 0022 0.05
ocxjøtmtaooon 001" 001 130 21
0.17 4.14
o 0560 oow«x
0 01000 0160 006
0032m1am3000.coouW300.18
021 017-0.18 -02 43240.Ø4401öo.ooso.cx)30.00R
0 14
o '2 0.14
k_Th
.ek_SLmday
24 h
15-0 12
017-0
.14
-0.2 -0.16
) 00800180093 300010010.
00010.0180081 016 0120.04101343080035
03400320.04
0020020.
002
o 0740.034.mem.07000BÄ 0770028002 omoooo 130.0000014001 0075101 oow m20c070 0030 mid.crg 0002mmu.m28 0320.0210092
0 2 looamoouM6ems0008ajooocxj.4D005D0020cx)3W048
0 084
D 084

Hist_fraud_trans_24h is highly correlated with  is_fraud  - 0.77 Hist_trans_24h is also correlated with hist_trans_60d - 0.56

Scene 9

When we plot the data among the various locations we see that the maximum fraud transactions are happening around East USA

Scene 10

419598
41378
Predicted Fraud
133278
511445
Predicted Not Fraud
400000
300000
200000
100000

After building the model we get the following above result. We see that our model have been quite good in detecting the fraud cases.

In the logistic regression model we get a good accuracy, precision , recall and F-1 score. And it performed well on both train and test set.

Our decision Tree model has worked well and give us a 100 percent effectiveness of the model.

Predicted Fraud Predicted Non- Fraud Fraud 415958 133278 Not Fraud 41378 511445

Logistic Regression Model Metrics Train set Test Set       Accuracy 84.20% 84.20% Precision 90.96% 91.02% Recall 75.95% 75.89% F-1 Score 82.78% 82.77%

Decision Tree Model Metrics Train set Test Set       Accuracy 100.00% 100.00% Precision 100.00% 100.00% Recall 100.00% 100.00% F-1 Score 100.00% 100.00%

Scene 11

Cost Benefit Analysis S. No Questions Answer a Average number of transactions per month 77183 b Average number of fraudulent transaction per month 402 c Average amount per fraud transaction  $                                                                                              530.00

Cost Benefit Analysis S. No Questions Answer 1 Cost incurred per month before the model was deployed (b*c)  $                                                                                  2,13,060.00  2 Average number of transactions per month detected as fraudulent by the model (TF) 34966.5 3 Cost of providing customer executive support per fraudulent transaction detected by the model  $                                                                                                    1.50  4 Total cost of providing customer support per month for fraudulent transactions detected by the model (TF*$1.5)  $                                                                                       52,449.75  5 Average number of transactions per month that are fraudulent but not detected by the model (FN) 11106.5 6 Cost incurred due to fraudulent transactions left undetected by the model (FN*c)  $                                                                                       16,659.75  7 Cost incurred per month after the model is built and deployed (4+6)  $                                                                                       69,109.50  8 Final savings = Cost incurred before - Cost incurred after(1-7)  $                                                                                  1,43,950.50

Performing the Cost-benefit Analysis we see that due to our model we created the company can save up to $ 1,43,950  by accurately detecting fraudulent transactions.

Credit Card Fraud Detection Model