Financial Fraud Detection

Analyzing a dataset of bank transactions and capturing as many fraudulent transactions as possible while minimizing false positives.

Dataset Description

With over 6,000,000 transactions totaling up to 1,144 billion and consist of 11 columns:

• Step: Represents a unit of time in hours. This is essentially the timestamp of each transaction (e.g., hour 1, hour 2, ..., hour 534, etc.).
• Type: Specifies the type of transaction.
• Amount: The amount of money transferred in the transaction.
• NameOrig: The identifier for the origin account from which the transaction was initiated.
• OldBalanceOrg: The balance of the origin account before the transaction took place.
• NewBalanceOrg: The balance of the origin account after the transaction has been processed.
• NameDest: The identifier for the destination account to which the transaction was directed.
• OldbalanceDest: The balance of the destination account before receiving the transaction.
• NewbalanceDest: The balance of the destination account after receiving the transaction.
• IsFlaggedFraud: A simplistic model flag that indicates a transaction as fraudulent if the amount transferred is greater than 200,000 (note: the currency is unspecified).
• IsFraud: Indicates whether a transaction was actually fraudulent, defined as a malicious attempt to transfer funds out of a victim’s account without their consent.

The types of transactions

The total amount

total_transaction_amount = df['amount'].sum()
total = total_transaction_amount / 1e9

The total amount of transactions made is: 1144.39 billion.

Fraud occurred in 'cash_out' and 'transfer'

# fraudulent transactions
fraudulent_transactions = df[df['isFraud'] == 1]

# the distribution of fraudulent transactions by transaction type
fraudulent_transaction_distribution = fraudulent_transactions['type'].value_counts()

Distribution of Fraudulent Transactions by Transaction Type: type
CASH_OUT    4116
TRANSFER    4097

Fraudulent Transactions

Fraud Detection Analysis

• Created a random sample of 10% of the dataset due to the large size of the dataset.
• Dropped columns: 'nameDest', 'nameOrig','isFlaggedFraud'
• 'isFlaggedFraud' only flags a transaction as fraudulent if it is greater than 200,000.
• 'nameDest' and 'nameOrig' may have a large number of unique categorical values which can be computationally expensive.
• converted 'step' to 'HourOfDay' assuming 'step' starts from hour 1

Hourly Transactions

Box Plot of Hourly Transactions

Model Performance

• Ran a Gradient Boosting Classifier to received an accuracy of 0.9995725044949895

• Ran a RandomizedSearchCV as the dataset was too large for a GridSearchCV

• Using it to find the best parameters: 'n_estimators','learning_rate, and 'max_depth'

param_dist = {
    'n_estimators': randint(50, 300),  # number of trees between 50 and 500
    'learning_rate': uniform(0.01, 0.1),  # prevents overfitting
    'max_depth': randint(3, 10)  # max depth of each tree between 3 and 10 
}

random_search = RandomizedSearchCV(
    estimator=gb_class, 
    param_distributions=param_dist, 
    n_iter=10, 
    cv=5, 
    random_state=42, 
    scoring='accuracy', 
    n_jobs=-1
)

Received a best score of 0.9995389733717996 & the best parameters:

best_params = {
    'n_estimators': 199,
    'learning_rate': 0.02428668179219408,
    'max_depth': 5
}

f1 Score

Optimized f1_score 

0.8169761273209549