Typically, the data analyst approaches a problem directly with an (inflexible) procedure designed specifically for that purpose. For example, the everyday statistical problems of classification (i.e., assigning class membership with a categorical target variable), and prediction of a continuous target variable (e.g., sale or profit) are solved by the “old” standard binary or polynomial logistic regression (LR) models, and the ordinary least-squares regression (OLS) model, respectively. This is in stark contrast to the newer machine learning “algorithmic” methods, which are nominally statistical models, or more aptly non-statistical models, in that no effort is made to represent how the data were generated. There are nonparametric, assumption-free “flexible” procedures that let the data define the form of the model itself. The working assumption that today’s (big) data fit the OLS and LR models – which were formulated within the small-data setting of the day over 200 years ago, and 50 years ago, respectively – is not tenable. A flexible, any-size data model that is self-defining clearly offers a potential for building a reliable, highly predictive model, which was unimaginable two centuries ago, even a half century ago. The most notable algorithmic methods for the everyday statistical problems are the decision trees (sets of if-then rules), such as CART, CHAID, and C5.0.

To paraphrase, "Even the best-laid models go oft astray, and have poor performance." When a statistical model performance is poor: Try something new, an algorithmic method, which promises a better solution. Big data (and small data too!) bring the prospects of data complexity, namely, data points that are difficult to deal with, referred to as a "data mass," which renders initially poor model performance. Algorithmic methods are flexible, which implies they are resistant, yet are not immune to the hardest of data mass. To take on a data mass, do the following: 1) Apply the algorithmic method. 2) Retain the initial results (initial model), albeit the performance is wanting. 3) Reuse the initial model (as a new variable) by appending it into the original data. 4) Reapply the algorithmic method. 6) Repeat the previous tasks, as often as necessary, until model performance is worthy of being accepted. The purpose of this article is to present the algorithmic GenIQ Model©, a flexible, any-size data method that lets the data alone defines the model. I use the GenIQ Model in a real case study to show what to do when a model performance is poor.

Log of odds of GROUP_(=1) = - 3.8768 + 2.2470*SEX_M + 0.0894*AGE_AT_ONSET + 0.0639*ONSET_TIME