Predictive modeling is the science of developing mathematical constructs that enable reliable prediction of future events or measurements based on past information. As an area of data mining, predictive modeling deals with extracting information from data and using it to forecast unknown events. This information can be exploited to make better business decisions, ultimately improving overall profitability. For example, predictive modeling can be used in a marketing campaign to identify specific products and specific distribution channels that match consumer characteristics. Models can be built using data from consumers' past purchasing history and past response rates for each channel. Assessment of business patterns, customer loyalty, portfolio performances and pricing variations are some other examples where predictive modeling has been exploited at Inductis to estimate future trends and increase client business growth.

Predictive modeling relies on the detection of the underlying relationships among different variables in order to predict the future values of variables of interest based on models of those relationships. For instance, in credit card fraud detection, variables such as a customer's recent transaction pattern, location and previous credit card usage patterns can predict the likelihood of a current transaction representing fraudulent behavior. A preemptive intervention in such circumstances can lead to a significant cost reduction in loss recovery. Estimating the relationship amongst these variables and their effect on the future is the critical modeling process. Successful models involve the use of advanced statistical and machine learning approaches to capture these relationships. These modeling approaches directly employ all available data to build predictive models. In some applications, auxiliary information (e.g. credit bureau ratings) can supplement available data (e.g. customer transactions) to provide for more accurate predictions. Models can continuously be updated by refining them as more data becomes available. Apart from predicting results these models can also assign a probability value to each result. These probabilities can serve as a numerical measure to score one result against another. Higher scores would typically indicate a targeted behavior. For instance, a high score could indicate a higher likelihood for a customer to terminate their current insurance policy. Identifying such customers and taking appropriate actions can lead to a significant increase in customer retention.

The main challenge in predictive modeling is the design of a correct model that will generalize well to unknown cases. Different models can be compared by evaluating prediction performance on separate held-out data. In some cases, the models can represent simple mathematical or logical relationships between variables and sometimes they can involve complex neural networks. The exact nature of the model depends on the structure of the available data and the targeted application. In many cases, interpreting predictive models is an important issue as this can identify critical variables which can lead to efficient decision rules. Another challenge in predictive modeling is with regard to the amount and the quality of data available. A lot of data from different sources involves precise identification and integration of required variables. Very little data or data with missing information can increase unreliability in prediction. Advanced modeling approaches (data cleansing, missing value imputation, variable reduction, model revalidation) are aimed at addressing these challenges and they in turn provide viable solutions. At Inductis, advanced predictive modeling relies on systematic application of evolving advanced techniques in conjunction with professional judgment based on deep insight into relevant business intelligence and extensive experience with statistical modeling.