Investigation into Vehicle Insurance Claims Fraud

Authors : Killian Hughes, Lauren Gorst, Peter Scaife, Taslima Hossain 08/01/2024

Overview

This reports seeks to investigate vehicle claims in the United States of America, specifically the fraud tendancy of different profiles' claims insurance policies. The report will proceed in the following way: data is extracted in its raw form from the csv file 'insurance_claims.csv' taken from the Mendeley data website, then the data will be cleaned for the purpose of processing for data visualisations and further machine learning models to aid early predictive fraud detection.

Data Source : insurace_claims.csv

Referenced Dashboard :

The Purpose of this report is esential to aid insurance companies in the early detection of vehicle insurance fraud based on historical data. Data visualistions will insure an accurate overview of current trends in this domain.

Report Contents

• Section 1 : Data Overview
• Section 2 : Data Cleaning
• Section 3 : Data Visualisations
• Section 4 : Evaluations and Further Insights
• Section 5 : Conclusion

Section 1 : Data Overview

The data is first imported using Pandas to perform an initial investigation. Below are the imports necessary for the report.

# Imports 

import pandas as pd 
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np

df = pd.read_csv("insurance_claims.csv")
display(df)

	months_as_customer	age	policy_number	policy_bind_date	policy_state	policy_csl	policy_deductable	policy_annual_premium	umbrella_limit	insured_zip	...	police_report_available	total_claim_amount	injury_claim	property_claim	vehicle_claim	auto_make	auto_model	auto_year	fraud_reported	_c39
0	328	48	521585	2014-10-17	OH	250/500	1000	1406.91	0	466132	...	YES	71610	6510	13020	52080	Saab	92x	2004	Y	NaN
1	228	42	342868	2006-06-27	IN	250/500	2000	1197.22	5000000	468176	...	?	5070	780	780	3510	Mercedes	E400	2007	Y	NaN
2	134	29	687698	2000-09-06	OH	100/300	2000	1413.14	5000000	430632	...	NO	34650	7700	3850	23100	Dodge	RAM	2007	N	NaN
3	256	41	227811	1990-05-25	IL	250/500	2000	1415.74	6000000	608117	...	NO	63400	6340	6340	50720	Chevrolet	Tahoe	2014	Y	NaN
4	228	44	367455	2014-06-06	IL	500/1000	1000	1583.91	6000000	610706	...	NO	6500	1300	650	4550	Accura	RSX	2009	N	NaN
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
995	3	38	941851	1991-07-16	OH	500/1000	1000	1310.80	0	431289	...	?	87200	17440	8720	61040	Honda	Accord	2006	N	NaN
996	285	41	186934	2014-01-05	IL	100/300	1000	1436.79	0	608177	...	?	108480	18080	18080	72320	Volkswagen	Passat	2015	N	NaN
997	130	34	918516	2003-02-17	OH	250/500	500	1383.49	3000000	442797	...	YES	67500	7500	7500	52500	Suburu	Impreza	1996	N	NaN
998	458	62	533940	2011-11-18	IL	500/1000	2000	1356.92	5000000	441714	...	YES	46980	5220	5220	36540	Audi	A5	1998	N	NaN
999	456	60	556080	1996-11-11	OH	250/500	1000	766.19	0	612260	...	?	5060	460	920	3680	Mercedes	E400	2007	N	NaN

1000 rows × 40 columns

The data includes 1000 rows and 40 columns. For the purpose of the investgation we have decided to limit the number of columns choosing to omit columns: Umbrella_limit, insured_relationship, Property_damage, _c39. Namely c_39 represented a completely void column with no description.

We can also note that our data relates to three policy states in particular: Illinois, Ohio, Indiana.

missing_values = df.isnull().sum()
missing_values

months_as_customer                0
age                               0
policy_number                     0
policy_bind_date                  0
policy_state                      0
policy_csl                        0
policy_deductable                 0
policy_annual_premium             0
umbrella_limit                    0
insured_zip                       0
insured_sex                       0
insured_education_level           0
insured_occupation                0
insured_hobbies                   0
insured_relationship              0
capital-gains                     0
capital-loss                      0
incident_date                     0
incident_type                     0
collision_type                    0
incident_severity                 0
authorities_contacted            91
incident_state                    0
incident_city                     0
incident_location                 0
incident_hour_of_the_day          0
number_of_vehicles_involved       0
property_damage                   0
bodily_injuries                   0
witnesses                         0
police_report_available           0
total_claim_amount                0
injury_claim                      0
property_claim                    0
vehicle_claim                     0
auto_make                         0
auto_model                        0
auto_year                         0
fraud_reported                    0
_c39                           1000
dtype: int64

This technique of detecting null values demonstrates an incorrect logical assumption that all missing values are already labelled as null values in the dataframe. Note that for the column 'authorities_contacted' the value 'None' is considered a null value despite importantly signifying that no authority was contacted.

Upon closer inspection of the data frame we can observe that the character '?' is used as a place holder for a null value in all affected columns and represents no further insights within the data set.

place_holder = df[df=='?'].count()
place_holder

months_as_customer               0
age                              0
policy_number                    0
policy_bind_date                 0
policy_state                     0
policy_csl                       0
policy_deductable                0
policy_annual_premium            0
umbrella_limit                   0
insured_zip                      0
insured_sex                      0
insured_education_level          0
insured_occupation               0
insured_hobbies                  0
insured_relationship             0
capital-gains                    0
capital-loss                     0
incident_date                    0
incident_type                    0
collision_type                 178
incident_severity                0
authorities_contacted            0
incident_state                   0
incident_city                    0
incident_location                0
incident_hour_of_the_day         0
number_of_vehicles_involved      0
property_damage                360
bodily_injuries                  0
witnesses                        0
police_report_available        343
total_claim_amount               0
injury_claim                     0
property_claim                   0
vehicle_claim                    0
auto_make                        0
auto_model                       0
auto_year                        0
fraud_reported                   0
_c39                             0
dtype: int64

To visually describe the scale of our missing values, we will first replace all placeholder characters with nulls and use a heatmap to observe our findings for each column of the data frame.

replaced_var = df.replace('?', np.NaN)

def is_null_data_heat(df):
    df.replace('Varies with device', np.nan, inplace=True)
    plt.figure(figsize=(10, 6))
    sns.heatmap(df.isnull(), cmap='viridis', cbar=False)
    plt.title('Heatmap of Missing Data')
    plt.show()

is_null_data_heat(replaced_var)

Section 2: Data Cleaning

Initial aims of data cleaning:

• Removal of duplicates
• Identifying missing values
• Ensuring uniform formatting
• Identifying outliers

The following remove the columns judged as being not of interest to the report.

columns_to_remove = ['umbrella_limit', 'insured_relationship', 'property_damage', '_c39']
df = df.drop(columns=columns_to_remove)

The following will give an overview of the data points within the data frame.

for column in df.columns:
    unique_strings_overview = df[column].value_counts()
    print(f"\nOverview for {column}:")
    print(unique_strings_overview)

Overview for months_as_customer:
months_as_customer
194    8
128    7
254    7
140    7
210    7
      ..
390    1
411    1
453    1
448    1
17     1
Name: count, Length: 391, dtype: int64

Overview for age:
age
43    49
39    48
41    45
34    44
38    42
30    42
31    42
37    41
33    39
40    38
32    38
29    35
46    33
42    32
35    32
36    32
44    32
28    30
26    26
45    26
48    25
47    24
27    24
57    16
25    14
55    14
49    14
53    13
50    13
24    10
54    10
61    10
51     9
60     9
58     8
56     8
23     7
21     6
59     5
62     4
52     4
64     2
63     2
19     1
20     1
22     1
Name: count, dtype: int64

Overview for policy_number:
policy_number
521585    1
687755    1
674485    1
223404    1
991480    1
         ..
563878    1
620855    1
583169    1
337677    1
556080    1
Name: count, Length: 1000, dtype: int64

Overview for policy_bind_date:
policy_bind_date
2006-01-01    3
1992-04-28    3
1992-08-05    3
1991-12-14    2
2004-08-09    2
             ..
2014-06-03    1
1998-12-12    1
1999-02-18    1
1997-10-30    1
1996-11-11    1
Name: count, Length: 951, dtype: int64

Overview for policy_state:
policy_state
OH    352
IL    338
IN    310
Name: count, dtype: int64

Overview for policy_csl:
policy_csl
250/500     351
100/300     349
500/1000    300
Name: count, dtype: int64

Overview for policy_deductable:
policy_deductable
1000    351
500     342
2000    307
Name: count, dtype: int64

Overview for policy_annual_premium:
policy_annual_premium
1558.29    2
1215.36    2
1362.87    2
1073.83    2
1389.13    2
          ..
1085.03    1
1437.33    1
988.29     1
1238.89    1
766.19     1
Name: count, Length: 991, dtype: int64

Overview for insured_zip:
insured_zip
477695    2
469429    2
446895    2
431202    2
456602    2
         ..
476303    1
450339    1
476502    1
600561    1
612260    1
Name: count, Length: 995, dtype: int64

Overview for insured_sex:
insured_sex
FEMALE    537
MALE      463
Name: count, dtype: int64

Overview for insured_education_level:
insured_education_level
JD             161
High School    160
Associate      145
MD             144
Masters        143
PhD            125
College        122
Name: count, dtype: int64

Overview for insured_occupation:
insured_occupation
machine-op-inspct    93
prof-specialty       85
tech-support         78
sales                76
exec-managerial      76
craft-repair         74
transport-moving     72
other-service        71
priv-house-serv      71
armed-forces         69
adm-clerical         65
protective-serv      63
handlers-cleaners    54
farming-fishing      53
Name: count, dtype: int64

Overview for insured_hobbies:
insured_hobbies
reading           64
exercise          57
paintball         57
bungie-jumping    56
movies            55
golf              55
camping           55
kayaking          54
yachting          53
hiking            52
video-games       50
skydiving         49
base-jumping      49
board-games       48
polo              47
chess             46
dancing           43
sleeping          41
cross-fit         35
basketball        34
Name: count, dtype: int64

Overview for capital-gains:
capital-gains
0        508
46300      5
51500      4
68500      4
55600      3
        ... 
36700      1
54900      1
69200      1
48800      1
50300      1
Name: count, Length: 338, dtype: int64

Overview for capital-loss:
capital-loss
 0        475
-31700      5
-53700      5
-50300      5
-45300      4
         ... 
-12100      1
-17000      1
-72900      1
-19700      1
-82100      1
Name: count, Length: 354, dtype: int64

Overview for incident_date:
incident_date
2015-02-02    28
2015-02-17    26
2015-01-07    25
2015-01-10    24
2015-02-04    24
2015-01-24    24
2015-01-19    23
2015-01-08    22
2015-01-13    21
2015-01-30    21
2015-02-12    20
2015-02-22    20
2015-01-31    20
2015-02-06    20
2015-02-21    19
2015-01-01    19
2015-02-23    19
2015-01-12    19
2015-01-14    19
2015-01-21    19
2015-01-03    18
2015-02-14    18
2015-02-01    18
2015-02-28    18
2015-01-20    18
2015-01-18    18
2015-02-25    18
2015-01-06    17
2015-01-09    17
2015-02-08    17
2015-02-24    17
2015-02-26    17
2015-02-13    16
2015-02-15    16
2015-02-16    16
2015-02-05    16
2015-01-16    16
2015-01-17    15
2015-02-18    15
2015-01-28    15
2015-01-15    15
2015-01-22    14
2015-02-20    14
2015-02-27    14
2015-01-23    13
2015-02-03    13
2015-01-27    13
2015-02-09    13
2015-01-04    12
2015-03-01    12
2015-01-26    11
2015-01-29    11
2015-01-02    11
2015-02-19    10
2015-02-11    10
2015-02-10    10
2015-02-07    10
2015-01-25    10
2015-01-11     9
2015-01-05     7
Name: count, dtype: int64

Overview for incident_type:
incident_type
Multi-vehicle Collision     419
Single Vehicle Collision    403
Vehicle Theft                94
Parked Car                   84
Name: count, dtype: int64

Overview for collision_type:
collision_type
Rear Collision     292
Side Collision     276
Front Collision    254
?                  178
Name: count, dtype: int64

Overview for incident_severity:
incident_severity
Minor Damage      354
Total Loss        280
Major Damage      276
Trivial Damage     90
Name: count, dtype: int64

Overview for authorities_contacted:
authorities_contacted
Police       292
Fire         223
Other        198
Ambulance    196
Name: count, dtype: int64

Overview for incident_state:
incident_state
NY    262
SC    248
WV    217
VA    110
NC    110
PA     30
OH     23
Name: count, dtype: int64

Overview for incident_city:
incident_city
Springfield    157
Arlington      152
Columbus       149
Northbend      145
Hillsdale      141
Riverwood      134
Northbrook     122
Name: count, dtype: int64

Overview for incident_location:
incident_location
9935 4th Drive         1
4214 MLK Ridge         1
8548 Cherokee Ridge    1
2352 MLK Drive         1
9734 2nd Ridge         1
                      ..
6770 1st St            1
4119 Texas St          1
4347 2nd Ridge         1
1091 1st Drive         1
1416 Cherokee Ridge    1
Name: count, Length: 1000, dtype: int64

Overview for incident_hour_of_the_day:
incident_hour_of_the_day
17    54
3     53
0     52
23    51
16    49
13    46
10    46
4     46
6     44
9     43
14    43
21    42
18    41
12    40
19    40
7     40
15    39
22    38
8     36
20    34
5     33
2     31
11    30
1     29
Name: count, dtype: int64

Overview for number_of_vehicles_involved:
number_of_vehicles_involved
1    581
3    358
4     31
2     30
Name: count, dtype: int64

Overview for bodily_injuries:
bodily_injuries
0    340
2    332
1    328
Name: count, dtype: int64

Overview for witnesses:
witnesses
1    258
2    250
0    249
3    243
Name: count, dtype: int64

Overview for police_report_available:
police_report_available
?      343
NO     343
YES    314
Name: count, dtype: int64

Overview for total_claim_amount:
total_claim_amount
59400    5
2640     4
70400    4
4320     4
44200    4
        ..
65250    1
87100    1
6240     1
66600    1
67500    1
Name: count, Length: 763, dtype: int64

Overview for injury_claim:
injury_claim
0        25
640       7
480       7
660       5
580       5
         ..
14840     1
6580      1
11820     1
16650     1
7500      1
Name: count, Length: 638, dtype: int64

Overview for property_claim:
property_claim
0        19
860       6
480       5
660       5
10000     5
         ..
3590      1
6480      1
4580      1
4920      1
7500      1
Name: count, Length: 626, dtype: int64

Overview for vehicle_claim:
vehicle_claim
5040     7
3360     6
52080    5
4720     5
3600     5
        ..
43360    1
25130    1
38940    1
47430    1
52500    1
Name: count, Length: 726, dtype: int64

Overview for auto_make:
auto_make
Saab          80
Dodge         80
Suburu        80
Nissan        78
Chevrolet     76
Ford          72
BMW           72
Toyota        70
Audi          69
Accura        68
Volkswagen    68
Jeep          67
Mercedes      65
Honda         55
Name: count, dtype: int64

Overview for auto_model:
auto_model
RAM               43
Wrangler          42
A3                37
Neon              37
MDX               36
Jetta             35
Passat            33
A5                32
Legacy            32
Pathfinder        31
Malibu            30
92x               28
Camry             28
Forrestor         28
F150              27
95                27
E400              27
93                25
Grand Cherokee    25
Escape            24
Tahoe             24
Maxima            24
Ultima            23
X5                23
Highlander        22
Civic             22
Silverado         22
Fusion            21
ML350             20
Impreza           20
Corolla           20
TL                20
CRV               20
C300              18
3 Series          18
X6                16
M5                15
Accord            13
RSX               12
Name: count, dtype: int64

Overview for auto_year:
auto_year
1995    56
1999    55
2005    54
2006    53
2011    53
2007    52
2003    51
2009    50
2010    50
2013    49
2002    49
2015    47
1997    46
2012    46
2008    45
2014    44
2001    42
2000    42
1998    40
2004    39
1996    37
Name: count, dtype: int64

Overview for fraud_reported:
fraud_reported
N    753
Y    247
Name: count, dtype: int64

Once identifiied the character '?' is officially replaced within the cleaned data frame by the null value representation 'NA'.

df.replace('?', pd.NA, inplace=True)

The following is the subsequent heat map.

plt.figure(figsize=(8, 6))
sns.heatmap(df.isnull(), cmap='viridis', cbar=False, yticklabels=False)
plt.title('Missing Values Heatmap')
plt.show()

To check for duplicates the following code is used to sum all duplicated rows in the data frame.

df.duplicated().sum()

0

The next code removes the primary key (which should all be individual values), to check for further duplicates

df.duplicated(subset=df.columns.difference(['policy_number']), keep=False).sum()

0

In order to identifying outliers we first found the interquartile range, upper bound and lower bound. Data points considred outside of this bounded range were treated as outliers.

def detect_outliers_iqr(column):
    q1 = column.quantile(0.25)
    q3 = column.quantile(0.75)
    iqr = q3 - q1
    lower_bound = q1 - 1.5 * iqr
    upper_bound = q3 + 1.5 * iqr
    outliers = (column < lower_bound) | (column > upper_bound)
    return outliers

outliers_column1 = detect_outliers_iqr(df['months_as_customer'])
outliers_column1.sum()

0

outliers_column2 = detect_outliers_iqr(df['age'])
outliers_column2.sum()
df[outliers_column2]

	months_as_customer	age	policy_number	policy_bind_date	policy_state	policy_csl	policy_deductable	policy_annual_premium	insured_zip	insured_sex	...	witnesses	police_report_available	total_claim_amount	injury_claim	property_claim	vehicle_claim	auto_make	auto_model	auto_year	fraud_reported
67	478	64	346002	1990-08-20	OH	250/500	500	835.02	602433	FEMALE	...	1	NO	33930	0	3770	30160	BMW	X6	1998	N
538	478	63	852002	2009-06-29	IL	250/500	1000	1292.30	479724	MALE	...	3	NO	4700	940	470	3290	Dodge	Neon	2007	Y
588	472	64	911429	2012-08-25	IN	250/500	500	989.24	615821	MALE	...	3	NO	70900	14180	7090	49630	Mercedes	ML350	2002	N
947	465	63	531160	2012-01-12	IL	250/500	500	1006.99	458936	FEMALE	...	3	<NA>	43560	4840	4840	33880	Suburu	Legacy	2015	N

4 rows × 36 columns

outliers_column3 = detect_outliers_iqr(df['policy_deductable'])
outliers_column3.sum()

0

outliers_column4 = detect_outliers_iqr(df['policy_annual_premium'])
outliers_column4
df[outliers_column4]

	months_as_customer	age	policy_number	policy_bind_date	policy_state	policy_csl	policy_deductable	policy_annual_premium	insured_zip	insured_sex	...	witnesses	police_report_available	total_claim_amount	injury_claim	property_claim	vehicle_claim	auto_make	auto_model	auto_year	fraud_reported
43	106	30	790442	2003-04-13	OH	250/500	500	538.17	433782	FEMALE	...	2	NO	51120	5680	5680	39760	Mercedes	E400	2005	N
133	24	33	649082	1996-01-19	IL	500/1000	1000	1922.84	431277	FEMALE	...	1	NO	46800	4680	9360	32760	Jeep	Wrangler	2002	N
229	154	37	241562	2010-01-28	IL	250/500	1000	2047.59	439269	FEMALE	...	3	NO	79530	14460	7230	57840	Accura	MDX	2000	N
248	101	27	326322	2007-02-10	IL	250/500	1000	433.33	603269	MALE	...	3	NO	5900	1180	590	4130	Mercedes	E400	2009	N
544	254	40	349658	1994-06-07	IN	100/300	500	1927.87	609317	MALE	...	1	<NA>	43280	0	5410	37870	Honda	Civic	1996	Y
681	32	38	719989	1994-04-07	IL	250/500	2000	566.11	453164	MALE	...	0	NO	2640	440	440	1760	Honda	CRV	2015	N
747	37	31	230223	2008-09-06	IL	500/1000	500	1969.63	615218	FEMALE	...	1	YES	76450	6950	13900	55600	Dodge	RAM	1995	N
763	66	30	984456	2003-06-24	IN	500/1000	500	484.67	608309	FEMALE	...	2	YES	65560	11920	11920	41720	Volkswagen	Passat	2015	Y
796	421	56	728025	1990-02-15	IN	100/300	500	1935.85	470826	MALE	...	3	<NA>	92730	16860	8430	67440	Mercedes	E400	2004	Y

9 rows × 36 columns

As demonstrated above, we identified four outliers in the age column and nine in the policy annual premium column. Upon further inspection we elected to leave these rows in the data frame as they do not contradict the purpose of this investegation and do not drastically affect the data trends.

Section 3 : Data Visualisations

• Profile age by gender (Kernel density estimation)
• Total count of incidents by state(bar)
• Average capital gains asnd losses by auto make (inverse bar)
• Mean number of witnesses by hour of the day (Line Graph)
• Frequency of Incidents by Hour of the Day (Line Graph)
• Ratio of Fraud Reported (Pie chart)
• Total Claim Amount for Fraudulent Claims (Histogram)

plt.figure(figsize=(10, 6))

sns.kdeplot(data=df[df['insured_sex'] == 'MALE']['age'], label='Male', fill=True)
sns.kdeplot(data=df[df['insured_sex'] == 'FEMALE']['age'], label='Female', fill=True)


plt.xlabel('Age Range')
plt.ylabel('Age Density Estimate')
plt.title('Average age of insured female and male clients (KDE)')
plt.legend()

plt.show()

Figure 1 : A KDE to show the age density according to age and gender.

There is a slight trend in the insured client pool implying that insured females are on average younger that the insured male. Most insured clients fall in the age range of 25 to 45 years of age.

state_counts = df['incident_state'].value_counts()


plt.figure(figsize=(10, 6)) 
state_counts.plot(kind='bar', color='blue')
plt.title('Total Count of incidents per State')
plt.xlabel('State')
plt.ylabel('Count of Incidents')
plt.xticks(rotation=45)  
plt.grid(axis='y')

plt.show()

Figure 2: A bar chart showing the counts of incidents per state.

The trends seen here shows a general decrease from left to right showing New York has the highest rate of incidents at 260 and Ohio having the least at 25.

grouped_data = df.groupby('auto_make')[['capital-gains', 'capital-loss']].mean()


fig, ax = plt.subplots(figsize=(10, 5))

grouped_data.plot(kind='barh', ax=ax)
ax.set_xlabel('Total')
ax.set_ylabel('Auto Make')
ax.set_title('Average Capital Gains and Losses by Auto Make')

plt.show()

Figure 3: Inverse bar chart showing the average capital gains and losses by automake

Figure 3 shows how profitable each car brand is to the insurance company. Volkswagon has a higher capital losss than gain demonstrating a poor profit auto make where as Nissan has a greater capital gain than loss.

mean_witnesses = df.groupby('incident_hour_of_the_day')['witnesses'].mean()

fig = plt.figure(figsize=(10, 7))
plt.plot(mean_witnesses.index, mean_witnesses.values, marker='o', linestyle='-')
plt.xlabel('Incident Hour of the Day')
plt.ylabel('Mean Number of Witnesses')
plt.title('Mean Number of Witnesses by Incident Hour of the Day')

plt.xticks(np.arange(0, 24, step=1), labels=np.arange(0, 24, step=1))  # Set x-axis ticks for each hour
plt.grid(True)
plt.show()

Figure 4: Line graph showing the likelihood of finding a witness per hour of the day.

There is a general trend demonstrating more witnesses being present at the scene of the incident during the day than at night as shown by three lowest witness counts at 1am, 3am and 9pm.

mean_incidents = df['incident_hour_of_the_day'].value_counts().sort_index()

fig = plt.figure(figsize=(10, 7))
plt.plot(mean_incidents, marker='o', linestyle='-')
plt.xlabel('Hour of the Day')
plt.ylabel('Frequency of incidents')
plt.title('Frequency of Incidents by Hour of the Day')

plt.xticks(np.arange(0, 24, step=1), labels=np.arange(0, 24, step=1))  
plt.grid(True)
plt.show()

Figure 5: Line graph showing the likelihood of an incident occuring per hour of the day.

There is a peak at 5pm depite a low average of incidents during daylight hours.

grouped_claims = df['fraud_reported'].value_counts()

plt.pie(grouped_claims, labels=grouped_claims.index, autopct='%1.1f%%')
plt.title('Fraud Reported')
plt.show()

Figure 6: A pie chart showing the ratio of fraudulant and unfraudulant claims.

Relative to predicted trends, there is a high number of fraudulant cases.

fraudulent_claims = df[df['fraud_reported'] == 'Y']['total_claim_amount']

plt.hist(fraudulent_claims, bins=20, color='skyblue', edgecolor='black')
plt.xlabel('Total Claim Amount')
plt.ylabel('Frequency of claims')
plt.title('Total Claim Amount for Fraudulent Claims')
plt.grid(True)
plt.show()

Figure 7: Histogram showing the total claim amount where fraud is reported.

The general trend observed is that most fraudulent claims fall at about the 60,000 mark with a secondary notable peak below 10,000.

Section 3a: Report Based Evaluations and Insights

Age and Gender

The age range and gender of insured clients could be of importance when considering the most profitable profile group to market a premium or insurance product towards. As seen in figure 1 the most populous age range currently insured is between the ages of 25 and 45. For this reason, marketing strategies would be most successful if targeted towards individuals within this age bracket, however to expand the current client base separate marketing strategies could be implemented to attract both younger and older clients.

Geographical

Further considerations into commuting habits of insured clients should be made prior to issuing a premium policy. As seen in figure 2, the highest amount of incidents occur in New York, South Carolina and West Virginia, therefore this should increase the premium pricing of clients frequenting these states.

Auto Make

Figure 3 aims to show the most profitable auto makes to the insurance company and an assumption can be made that these are safer cars to insure, representing low risk to the profits of the company. Whilst it is hard to compare bars of capital gains against capital loss, Nissan is demonstrated to be more profitable while Volkswagen has greater loss than gains, losing the company money. In regards to issuing premiums we judge this to be of great importance therefore we have further explored and improved the visual in the referenced dashboard (see section 3b, ).

Witnesses and time of day

In regards to figure 4, it may be assumed that an incident would present fewer witnesses during night and early morning hours. There is however no remarked trend in the data supporting this assumption. Although there is a visually noticable trough at 9pm, there is no supporting evidence to suggest that this is due to being at an advanced hour of the day. Given that the data set is relatively small and therefore presents many limitations when considering trends, an expanded view of incident records may improve predicted trends and aid evaluation of insurance fraud claims. Figure 5 supports a very similar conclusion that hour of the day does not predict the likelihood of incidence occurance. Similarly with expanded records this could help determine premium pricing.

Fraud Reported

Figure 6 demonstrates the incidence of recorded fraud within the data set. The limitation of this is that our data may represent a select group of records where fraud was initially suspected and sent for further investigation. When fraud is confirmed, total claim amount tends to fall at about 60,000 dollars. However it can also be noted that many fraudulant claims are also predictable a the lower end of 10,000 dollars meaning that smaller amount claims should be highly monitored.

Section 3b: Dashboard Based Evaluations and Insights

Vehicle Insurance Claims Fraud

Figure 3 is further improved in this visual, with capital gains and losses being more easily compared. Additionally we consider the total annual premium for both fraud and non fraud cases. Interestingly fraud cases account for about a quarter of total incident claims, representing 308,000(dollars) of the total collected 1.26 million annual premium. The interactive nature of the dashboard allows for further personal exploration into total annual premium per auto make as well as the number of clients per auto make. Furthermore the fraud slider, present on each page of the dashboard allows for a drill down approach, investigating data accounted for only by fraud cases. The car make Accura despite having only 13 insured clients, the second lowest insured vehicle, accounts for the highest capital loss towards the company.

Insured Client Profiles

Upon inspection it seems that there are more incidents for insured clients in the age range of 30-40, however in conjunction with figure 1 we cannot make the assumption that 40 year olds have more incidents on average than any other age group. This is because the data set we used only contains insured clients who have had incidents and as shown in figure 1 this age range accounts for the largest number of clients. Looking more closely at individul profiles we can consider the likelihood that a certain occupation ofr education level may impact a fraud claim or higher incident rate. Namely our primary key influencer of a major damage collision, is if the profile indicates the preferred hobby 'sleeping' closely followed by 'movies'. For non fraud cases also exhibiting major damage, the primary hobbies are 'chess' and 'cross-fit'. The key influencers segment on the bottom left of the dashboard also allows for inspection of cases consituting: 'minor damage', 'total losss' and 'trivial damage' by hobby.

Geographical mapping of Incidents

The dashboard allows for further geographical exploration of the data namely allowing for a comparitive investigation of the state where the policy was issued and the state of incidence. This is of particular importance when noting that the incidence of a collision occuring in the state where the policy was taken is extremly low. In fact of the three states where a policy was issued: Ohio, Illinois and Indiana, the only state also representing state of incident was Ohio. Furthermore of the 1000 incidents recorded only 23 were in the state of Ohio. The dashboard also allows the user to track a client by their policy number such that the previously describe profiles can be further monitored in regards to their driving activities. Type of collision and witnesses at the scene are also data points that can be explored.

Section 5: Conclusion

Since we have previously discussed the limitations of the data in regards to sample size and insured client profiles, we will reiterate this fact when considered an absence of definite trends and that all the claims represent 100% incidence rate combined. Financial the important point to consider is that the company has paid out 53M (dollars) in insurance claims. Of this figure 15M (dollars) represents the total amount paid in the instance of fraudulent claims while the remaining 38M (dollars) represents non fraudulent pay outs. Fraud pay outs therefore account for abut 28.3% of total insurance pay outs. Two auto makes were considered to be significantly important in this regard; Accura (all registered cars accounting for the highest capital loss), ford (accounting for largest total claims amount in fraudulent instances).

Given that data set only contains accidents, this justifies why total claim amount is significantly larger than total premium annual policy. If these dashboards were used in conjunction with real insurance data it would be possible to predict the profit for the insurance company by considering total premium annual policy over total claim amount.

Our system aims to help with identification of fraudulent cases, so far there are many fraudulent cases that are apparently unjustified. In the example of policy number 594783 for example, the total months as a customer is recorded as being 119, as such it raises the question as to why a client would pay for insurance for 10 years before filing a fraudulent claim. In addition, 1 witness is provided in the entered record.

In conclusion while no definite trends are observed, a typical fraudulent profile can be outlined:

Gender : male or female Occupation: Executive managerial Education Level: JD Hobby : Sleeping (major damage cases), Cross-fit and chess (total and minor loss cases), golf (trivial damage cases). Auto Make : Accura (All fraud), Mercedes/Ford (highest client base in regards to fraudulent cases)