Solution offer
for Bank and Finance


Artificial Intelligence is a tool with a consolidated basis and of absolute interest for banking processes. The expected impact in 2023 on the US banking system is a save of $ 447 bln divided into:

44% Front office
48% Middle office
8% Back office

Expertise A.I. Area Banking

Front Office

Conversational interfaces

Model portfolio processing

Robo-Advisory (automated consulting )

Middle Office

Know your customer


Fraud detection

Back Office

Automatic Processing

Churn Prediction

An accurate prediction of the probability of abandon of a customer (churn prediction) has an impact on various aspects of the business, including

  • Customer oriented proactive market
  • Sales forecast
  • Price models oriented towards customer retention

Although deterministic models to solve this need have existed for long, more modern technologies allow to create much more accurate models, even on high-dimensional and even time-dependent databases.

The first step in approaching the solution lies in identifying the set of data that describe the customer compared to his level of loyalty and his purchasing habits.

Thanks to the potential of artificial intelligence, it is possible to obtain data from various sources: Customer demographics

  • Payments
  • Buying habits
  • Call center

To define the customer as accurately as possible, an abandon indicator is also defined:

  • False, if the client is active
  • True, if the client abandoned on a basis of criteria to be shared with business.

The obtained dataset is submitted to the machine learning algorithm, which learns the common characteristics of loyal and leaving customers, highlighting the relationships between the data.

The set goal can be solved through classification algorithms: the model is trained to recognize a leaving customer compared to a loyal customer.

The dataset is divided in

  • Training: data used to train the network
  • Validation: data used to validate and improve the training
  • Test: data used to control the model performances

During the prediction phase, a customer profile is submitted to the trained model and this is able to provide
The classification of a customer, that is, whether he is a candidate to leave or not
The probability of belonging to the assigned class
The confidence interval with which this probability is assigned.

In summary, it is possible to predict whether a customer will abandon or not by considering this as a classification problem and building a descriptive dataset of the customer’s behavior starting from the sources present in the company. The answer is given in terms of the customer’s class (churn / no churn) and associated probability.

It is also possible to give an estimate of which variables in the dataset have greater weight in one or the other class, obtaining information on where it is most advantageous to act to prevent abandon.

Product oversight Governance

The provisions of B.d.I. of 5/12/2018 on governance and control mechanisms for retail banking products introduce the protection of consumers of banking services, through a set of rules governing the life cycle of the product.

During the entire life cycle of the product, internal procedures must ensure that intermediaries take into due consideration the interests, objectives and characteristics of customers, the typical risks of the products that can cause prejudice for customers, and possible conflicts of interest.

Therefore, the identification of the positive and negative target market remains essential, the classes of customers for which a product can be considered suitable or unsuitable.

EBA specifies that the same guidelines “do not regulate (…) the suitability of products with respect to individual consumers”, thus establishing that it does not appear possible to apply rules of conduct comparable to the assessment of suitability typical of some services to the marketing of banking products of investment, regulated by MiFID.

There are no specific indications or prescriptions at the regulatory level regarding the criteria to be used to define the reference market, positive and negative.

The information on the risk and complexity of the individual product is contained in two synthetic indicators

  • Risk grade
  • Complexity grade

valorised through algorithms associated with each product starting from elementary indicators.

Information on risk inclination can be obtained:

  • direttamente da questionari che la Banca propone alla clientela
  • derived from personal, demographic and behavioral information of customers

The first step in approaching the solution is identifying the set of data that describe the customer and the products, describing on the one hand risk propensity and the degree of financial literacy and on the other the degree of risk and complexity.

The obtained datasets are submitted to clustering, an unsupervised method that allows the dataset to be separated into homogeneous subsets.

In this way the initial pairs (Customer, Product) are remapped into families (Customer Cluster, Product Cluster) that define the target markets
Finally, the target markets are submitted to machine learning methods that identify any outliers, discovering the numerical consistency of off-target customers.

It should be noted that this approach can also be used in a proactive key, referring a new customer to a cluster of customers and proposing the products related to the most appropriate cluster of products.

Risk Based Financial Advisory

The financial services industry is undergoing a significant transformation in the way financial management and advisory services are requested and delivered. This evolution is due to a variety of factors: new regulatory parameters, customer demographic changes and, above all, technological advances. These changes are seen at a time when the need and demand for financial product advice has never been higher.

The advice provided through a robo-advisor is customized on a profiling performed on the customer.

This profiling is used to define the characteristics of the investor, in terms of risk propensity or aversion, objectives and / or financial constraints.

Once the customer has been “framed”, an allocation of investments that is appropriate to his profile is proposed to him.

The first step in approaching the solution is identifying the dataset that describe the customer and the products.
Information is collected and organized:


  • personal data
  • riskiness
  • trends


  • risk profile
  • financial habits
  • KYC

Given the capabilities of artificial intelligence, it is possible to build even very complex and highly dimensional representative models.

Algorithms are built to group financial instruments on one side and customers on the other, and then cross the groupings and determine the investment.
Neural networks associate model portfolios with customers.

This tool acts as a valid support in developing investment strategies and in ensuring compliance of the latter with legislation.

Fraud Detection

As part of financial controls, it is possible to implement an A.I. able to identify:
Whether a trading day for a stock is normal or abnormal, based on its past history.

If a customer has performed abnormal movements, considering its past history.

These problems are attributable to the class of so-called anomaly detection problems, that is, related to the discovery of anomalous situations submerged in a background of data considered normal.

To obtain a good accuracy of the algorithm it is necessary to have a dataset rich in information on the past history of securities and customers.

For securities, all the information related to individual market days is certainly necessary, such as average price, volatility and quantity traded.

While for customers it is necessary to have all possible information related to personal data, demographic data, portfolio data and the history of its movements.

The days of stock trading or the movements of a client are thought as points in motion in an appropriate space. As the days pass, the point will draw a trajectory in this space, which will mostly remain confined to an area, occasionally exiting it in conjunction with events that disturb this motion.

The zone in which the point spends most of its time represents the normality of the conditions described by the coordinates, while the occasional points outside this zone represent anomalous situations.

Unsupervised networks are trained to search for points of anomaly regardless of parameters and thresholds, only by referring to the past history of the stock or customer.

Request an online demo

One of our consultants will assist you with the explanation.

By filling out the form above and sending a request for information, I declare that I have read the privacy policy of ATG Artificial Intelligence SRL and accept the terms.