Fraud Detection - Blocker Fraud Company

There is a lack of public available datasets on financial services and specially in the emerging mobile money transactions domain. Financial datasets are important to many researchers and in particular to us performing research in the domain of fraud detection. Part of the problem is the intrinsically private nature of financial transactions, that leads to no publicly available datasets.

We present a synthetic dataset generated using the simulator called PaySim as an approach to such a problem. PaySim uses aggregated data from the private dataset to generate a synthetic dataset that resembles the normal operation of transactions and injects malicious behaviour to later evaluate the performance of fraud detection method

A fraudulent transfer, is an attempt to avoid debt by transferring money to another person or company. It is generally a civil, not a criminal matter, meaning that one cannot go to jail for it, but in some jurisdictions there is potential for criminal prosecution. It is generally treated as a civil cause of action that arises in debtor/creditor relations, particularly with reference to insolvent debtors. The cause of action is typically brought by creditors or by bankruptcy trustees.

Business Problem

For the development of the project, we created a fictitious company called "Blocker Fraud Company", which is a company specialized in detecting fraud in financial transactions made through mobile devices. The company has a service called “Blocker Fraud” in which it guarantees the blocking of fraudulent transactions.

The company's business model is of the Service type with the monetization made by the performance of the service provided, that is, the user pays a fixed fee on the success in detecting fraud in the client's transactions.

However, the Blocker Fraud Company is expanding in Brazil and to acquire customers more quickly, it has adopted a very aggressive strategy. The strategy works as follows:

1 - The company receives 25% of each transaction value truly detected as fraud.

2 - The company receives 5% of each transaction value detected as fraud, however the transaction is legitimate.

3 - The company gives back 100% of the value for the customer in each transaction detected as legitimate, however the transaction is actually a fraud.

What do we need to show?

• What is the model's Precision and Accuracy?
• How Reliable is the model in classifying transactions as legitimate or fraudulent?
• What is the Expected Billing by the Company if we classify 100% of transactions with the model?

Dataset: https://www.kaggle.com/ntnu-testimon/paysim1/notebooks?sortBy=hotness&group=everyone&pageSize=20&datasetId=1069&language=Python

Approach

1.0. Initial Data Analysis.

2.0. Feature Engineering

3.0. Filtering the features

4.0. Exploratory Data Analysis

5.0. Data preparation

6.0. Feature selection

7.0. Balanced Data

8.0. Machine Learning Modelling

9.0. Hyperparameter Fine Tunning

10.0. Business Performance and Financial Return

1.0. DATA DESCRIPTION

1.1. Data Dimensions

Let's understand how big our dataset is. This will be important because a robust machine learning model needs a considerable amount of data to train our algorithm.

Although we have a lot of data, we need to know if our computational capacity is sufficient. If a problem appears during the project, it will be necessary to use a GPU.

1.2. Data Types

Apparently our dataset does not have inconsistent data types.

1.3. Descriptive Statistical

1.3.1. Numerical Attributes

• It's important to note that we don't have null values.
• is_flagged_fraud --> Most of the data in the "is_flagged_fraud" column is zero.
• oldbalance_org --> If we look at the median, which represents the intermediate value, we have the value of 14,691.00 However, the range is equal to 59,585,040.00, which is very high. This variable will need to be checked soon.
• newbalance_org, oldbalance_dest, newbalance_dest, amount --> Respects the same behavior as the oldbalance_org variable

1.3.2. Categorical Attributes

• Only 0.8% of our data is considered fraudulent.
• High presence of outliers who transferred to our type variable.
• Cash_out is the most common form of payment, followed by payment.

2.0. MINDMAP HYPOTHESIS

From the mind map above, we created some features to validate the hypotheses.

3.0. DATA FILTERING

In the descriptive analysis section, we observed the presence of many outliers in our financial variables. So, let's take a look to the distribution of the boxplot, since we do not have more information about the business.

4.0. EXPLORATORY DATA ANALYSIS

4.1. Univariate Analysis

4.1.1. Response Variable

• 99.2% of our data is not considered fraudulent, against 0.8% fraudulent.
• Fraud has a very oscillatory tendency in relation to days.

4.1.2. Numerical Attributes

• Financial variables with high concentration of values in a small range.
• In the last 10 days of the month, there were far fewer transactions.
• The first two weeks had a higher number of transactions.
• There are more transactions on weekdays than on weekends. This is predictable due to the number of days.

4.1.3. Categorical Attributes

• The most frequent types of transactions are, in descending order: CASH_OUT, PAYMENT, CASH_IN.
• There are twice as many transactions for Costumers as compared to Merchant.
• All transactions are made from Costumers. That is, there are no transactions executed by Merchant

4.2. Bivariate Analysis

HYPOTHESIS VALIDATION

H1. High value transactions are more likely to be fraudulent

True - There are more fraudulent transactions with higher amounts.

• The more the amount increases, the more cases of fraudulent than legitimate transactions.
• The correlation between the variable amount and is_fraud is positive and low.
• From the graphs above, we can conclude that the hypothesis is true.

H2. There are more cases of transfer fraud

True - There are more cases in transactions that were made through transfer.

• There are cases of fraud only through transfer and cash out.
• Transactions with higher amounts are usually transfer and cash_out.
• The sum of the values of fraudulent transactions is much larger than legitimate ones.

H3.There is more chance of having fraud when the final balance is zero.

• As we can see above, fraudulent transactions usually happen when oldbalance_org is greater than zero.
• 80% of the values that are not fraudulent are within the range 0 to 250,000.
• We can say that the hypothesis is true.

4.3. Multivariate Analysis

Very correlated variables are not good for the models, as we would increase the dimensionality without placing relevant variables. That is, if we have two highly correlated variables, we remove one from it.

In this case, we have a high correlation between the variables:

• error_balance_orig x Erro_balance_dest
• newbalance_dest x oldbalance_dest
• error_balance_dest x amount
• error_balance_orig x amount

5.0. MACHINE LEARNING MODELLING

5.1. Machine Learning Performance

Random Forest is the machine learning model with the best performance.

Real Performance - Cross-Validation

10.0. Conclusions Business Performance and Financial Returns

A successful model has been created for fraud detection using advanced machine learning models but how does it apply to the business in terms of financial metrics.

First of all, let's remember what is the business model:

• 1 - The company receives 25% of each transaction value truly detected as fraud.
• 2 - The company receives 5% of each transaction value detected as fraud, however the transaction is legitimate.
• 3 - The company gives back 100% of the value for the customer in each transaction detected as legitimate, however the transaction is actually a fraud.

The Current model's revenue is in negative as there is more fraud than the flagged fraud, it shows that the current method can't recognize fraud efficiently.

However, with our model and calculating these on a sample of 241971 Customers

• 1. Blocker Fraud Company will Recieve $110,914,765.84 due to transactions truly detected as fraud.
• 2. Blocker Fraud Comapany will receive $0.00 due to transactions detected as fraud, but actually it was legitimate.
• 3. Blocker Fraud Company will lose $898,095.27 due to transactions detected as legitimate, but actually it was fraudulent.

Therefore, Blocker Fraud Company can project a revenue of $110,024,670 with this model, positioning itself for confident expansion into global markets and a stronger foothold in combating fraud worldwide.