Predictive Modelling


Predictive modelling refers to a collection of methods whereby specific data can be analyzed and interpreted to derive from it predictions for future events, trends or consumer behavior. These predictions are to be regarded as probabilities rather than actual statements about the future which will absolutely happen. The results of forecasting models are predictions about the respective subject area based on statistics, which depend in their probability on the size of the database. The more input data is available, the more accurate the output data will be. There is no guarantee for the actual occurrence of the prediction, however.

General information[edit]

Predictive modelling is applied in a variety of areas and disciplines, including insurance and finance, telecommunications, science, e-commerce, customer relationship management, and business intelligence. The forecasts can be used in terms of economic applications as a basis for budget planning and the assessment of opportunities and risks. The best-known application of predictive modelling is the calculation of risks in life insurance. In a scientific context, it is about the confirmation or refutation of theories that describe the behavior of specific objects dependent on the topic. The demographic development of societies is an example of this.[1]

How it works[edit]

Three phases can be distinguished in modeling.[2] In order to increase the probabilities of certain statements, these phases are usually repeated several times and are therefore partly iterative.

  • Training: First, data is collected and analyzed. This can be done quantitatively or qualitatively. Factors that are particularly relevant get identified taking into account the specific application. In e-commerce, for example, purchase histories are analyzed to find out which customer segments have purchased certain products. Relevant factors may be age, gender, social status, search requests, and interests. Based on this data, customers are divided into segments and a first prognosis generated.
  • Control: The forecast prognosis is checked against other new data. This data acts as a control factor in order to be able to assess the accuracy and reliability of the model. The model may be modified under certain circumstances. Cross-selling and product recommendations in e-commerce, for example, are scaled to their share of the profits. In other words, it is checked whether the product proposals actually result in an increase of sales.
  • Prognosis: As soon as a more accurate and reliable model is developed, training and control data can be input to create forecasts. If new data is available, further controls are carried out. Users of an e-commerce platform get specific proposals, such as “users who bought product X also bought product Y.” These proposals are the result of different stages of the predictive modelling. It is likely that customer XY will be interested in product Y, because they belong to a specific customer segment and have been associated with a certain purchase history.

Types of Predictive Models[edit]

In general, many different statistical models and multivariate regression analysis can be used to make predictions. In the context of applications in IT, it usually concerns data mining and machine learning. On the one hand relevant input data is found in large databases and on the other hand forecasting models should be self-learning, so that they automatically respond to new data.

Practical relevance[edit]

Forecasts are always problematic in certain respects, since they are a form of inductive arguments. General application or statements will be concluded from empirical observations and individual cases. While it is reasonable to consider such conclusions to be true if the premise (the data) are true, it does not mean that the conclusions are logically stringent or can claim universality. Moreover, large amounts of data may enable more accurate predictions, however, more data does not always mean a better insight into the behavior of customers. Experts often recommend smart instead of big data in this context.

These limitations of predictive modelling become more apparent when it comes to human behavior. To predict the behavior of customers, the forecast model should know all the variables that affect this behavior.[3] But people are influenced by hardly countable factors from their environment. Each model is therefore limited in certain respects. Nevertheless, some e-commerce platforms such as Netflix or Amazon managed to generate a high increase in sales with predictive modelling.

Relevance to search engine optimization[edit]

Using predictive modelling in the fields of web analytics, online marketing, search engine optimization, and social media, not only opportunities and risks can be assessed, but also expected sales can be calculated. Therefore, forecasts in these contexts can be used for planning and budgeting. Advertising companies can decide which marketing channels should be considered in view of their potential. This also applies to actions for online customer loyalty, customer relationship management, email and newsletter marketing, search engine advertising, and affiliate networks. The possible applications for Predictive Modelling are extremely diverse.

References[edit]

  1. Predictive Modeling gartner.com. Accessed on 04/27/2015
  2. Wondering What Lies Ahead? The Power of Predictive Modeling forteconsultancy.wordpress.com. Accessed on 04/27/2015
  3. Predictive Analytics with Data Mining: How It Works predictionimpact.com. Accessed on 04/27/2015

Web Links[edit]