Machine Learning for Smarter Businesses
While the world heads to globalization, companies and services working with people (B2C) follow the way of personalization. People are different, they have different problems, needs, and social standing. Can we afford treating in the same manner people of the same age or financial standing? Can we offer them the same services, show the same ads, make same decisions on loans granting? It’s not the best way if we want more clients to make the choice in our favor.
With Big Data in hand, we actually know a lot about our clients: from behavior patterns to numerical and categorical indicators, for example, credit and deposit accounts statistics of a client, browsing history, number of transactions, history of banking products usage, and more. This data enables banks to increase efficiency in selling credit or other banking products. While now bank sales persons need to call 100 clients to get 2-3 sales, after proper clients selection they can call 100 pre-selected clients and get 70+ sold offers. The results are self-evident: better service, decreased workload, improved financial performance of the institution.
Analysis of your data can help you with other tasks, forecasts, and plans. Churn prediction, customer segmentation/target customer prediction, credit scoring, the list is long. For example, churn prediction means that you can determine if your regular client will stay with you in the near future based on their history and history of other clients that left. When you know that some clients are going to leave you, you can take measures to prevent that.
The key to all such tasks is classification, and to solve them, you need to define object parameters. It’s great if data changes with time, then you can isolate seasonal indicators and trends. In addition, medians, quantiles, means, values of a certain indicator for a specific time period are of great use. After receiving a range of such parameters and having used preprocessing and normalization to them according to the model needs, you can proceed to selection, optimization, and tuning of the model. Model selection depends on many factors, such as training speed, data volume required for training, forecasting speed, model size in memory, etc. Depending on incoming data, up to 70-85% accuracy can be reached for the mentioned tasks today. It also worth giving attention to validation of the obtained model to avoid model overfitting for a specific dataset and bend every effort to make the model handle equally well both test data and real work data at the production server. Using these technologies, you can get accurate results faster and on larger data volumes. This means smarter decisions and more efficient resource utilization.
Let’s consider business benefits from machine learning solutions in the banking context. Specifically, a model for user ad targeting created to increase efficiency in selling a credit service to the existing bank clients. When standard cold calls across unfiltered lists are used, the mean probability that a client buys a credit product is 2-5%. In this case, a machine learning solution can help select clients that are likely to be interested in the credit service. Such solution analyzes available data of existing clients and determine the features related to credit demand based on historic data.
The input data we used covered 500,000 bank clients described by 300 parameters each (e.g., client type, gender, education, employer name, residence district, age, if they took on credits before, credit and debit history for the last 6 months, if the client uses the services of other banks, spendings by categories over the month, history of mobile banking usage. In the sample dataset, only 2 percent of 500 thousand clients took on credits before. It means that class balance for forecasting is violated.
We began with performing the initial data analysis using input parameters to understand if we can build a model based on the data, if data normalization or conversion is required, which model can be used. Then we created a baseline model to determine its initial accuracy and suitability of the selected model type. Further, we increased model accuracy using validation data by generating new features and model regulation to avoid overfitting. As a result, we got the model with 74% accuracy, that is we could correctly forecast if the client needs a credit in 74 cases out of 100 (in contrast to 2-5% with unfiltered lists). The most significant for forecasting attributes turned out to be: the residence place, credit limit, presence of open credits, proportion of expenses and income over the last several months, expenses trend, age, and other parameters reflecting credit cards usage.
Thus, a forecasting model like the one we discussed can increase the efficiency of sales calls from 2-5% to 74%. This means that your clients will not be annoyed by non-relevant contacts, and your personnel will not spend time profitably instead of making fruitless calls. Of course, this is just one of countless examples of how machine learning and artificial intelligence can make your business smarter. Meaningful insights and predictions derived from your data provide a solid foundation for business decisions and planning.