Synthetic Data in Finance: Top 4 Applications in 2024

Artificial intelligence has a diverse set of applications in financial services from process automation to chatbots and fraud detection. The estimates show that aggregate potential cost savings for banks from AI applications would be $447 billion by 2023.

However, some of these applications have their limitations because financial data is one of the most sensitive and personally identifiable data types. To illustrate, 87% of Americans consider credit card data as moderately or extremely private. The same figure is 68% for health and genetic data and 62% for location data.

Financial institutions can leverage synthetic data, or data that is generated artificially based on real data, to overcome privacy (and other) challenges and provide innovative products and services to their customers.

You can see synthetic data use cases / benefits in financial institutions below:

1. Enables data sharing, collaboration and innovation

Regulations such as GDPR and CCPA may prevent sharing of financial data both within a company and between institutions. This can hinder valuable collaborations between financial institutions and fintech partners or between teams within an institution. Providing access permits to third parties can either take months of bureaucratic procedures or it may not even be possible. This makes it hard for a financial institution to evaluate potential partners before developing new products.

Anonymizing sensitive data with traditional data masking techniques before sharing can be susceptible to linkage attacks. These attacks aim to re-identify individuals from an anonymized dataset often by combining it with other publicly available datasets. According to a commonly cited 2000 study, 87% of the US population can be uniquely identified by combining their gender, birth date, and zip code.

Synthetic data can eliminate the risks of sharing. Instead of the original dataset, financial institutions can share synthetic data that preserves the important characteristics of the original dataset. Synthetic data generation techniques can be applied to a wide range of data types, from tabular to time series and artificial images. 

2. Enables rare event (e.g. fraud) prediction

Detection of fraudulent transactions is one of the major applications of machine learning in finance. However, a bank transaction dataset that contains fraudulent activities is typically imbalanced: fraudulent activities constitute a small percentage of all activities. It makes challenging for an ML model to learn from this type of dataset to detect new occasions of fraud because small data size can lead to inaccurate results.

Undersampling and oversampling are two techniques for handling imbalanced datasets. Undersampling involves removing non-fraud observations to balance the dataset. It requires the dataset to be large because removing observations can create bias. 

Oversampling, on the other hand, is generating new artificial instances of fraudulent activity that resemble real fraud. The ML model can then be trained on the balanced dataset to achieve more accurate results. Synthetic data generation techniques can be used to create artificial instances of fraud to obtain a balanced dataset.

3. Enables simulations

Sometimes financial institutions may want to test strategies under extreme conditions such as market crashes or app failures. Rather than having an imbalanced dataset of such events, they can lack the data arising from these conditions. Synthetic data can be used to fill these gaps and can help organizations to develop strategies against these kinds of events.

4. Improves supervised deep learning model accuracies

Most of the machine learning models and especially deep learning models are data-hungry. Even if a financial institution does not lack data to train an ML model, the accuracy of ML models considerably depends on data size. Synthetic data can be used to increase data size.

In addition to increased data size, labeled data is another advantage of synthetic data for model accuracy. This is especially relevant for supervised learning applications since these types of models learn from labeled data. Data labeling is a labor-intensive process and manual labeling is prone to errors, which can cause model inaccuracies. Synthetic data comes with correct labels for observations, eliminates the necessity of data labeling efforts, and gives way to more accurate ML models.

If you want to learn more about synthetic data, check our other articles on the topic:

• Synthetic data: Pros, Cons, Case studies & AI Applications
• Synthetic Data Generation: Techniques, Best Practices & Tools
• 26 Synthetic Data Statistics: Benefits, Vendors, Market Size

If you are looking for synthetic data generation software, check our data-driven, sortable/filterable list of vendors.

If you still have questions about synthetic data, do not hesitate to contact us: