This paper presents a visual analytics workflow for Exploratory Model Analysis (EMA) — discovering and selecting predictive models on a dataset when neither the modeling task nor the model type is predetermined. Data Exploration. The analyst examines dataset attributes through interactive, cross-linked visualizations appropriate to the data types present (histograms for tabular data, node-link diagrams for graphs, line charts for time-series), identifying feature distributions, inter-attribute relationships, and potentially predictive variables. From this exploration, the analyst forms initial modelling interests — which variables appear predictive and what types of models might be suitable. Problem Exploration. The system automatically enumerates valid modelling problems by pairing each variable as a potential prediction target with compatible model types (classification, regression, clustering, forecasting, etc.) and evaluation metrics. The analyst browses these candidate problems in context of their data understanding. Iterative Model Discovery. The analyst specifies a modelling problem by selecting a target variable, model type, evaluation metric, and predictor features — choosing from system-generated candidates or defining a custom specification, informed by their analytical direction from data exploration and diagnostic reasoning from any prior modelling attempts. An automated system trains a diverse set of candidate models, which are visualized through type-appropriate prediction displays (confusion matrices for classification, residual bar charts for regression) cross-linked with data exploration views. The analyst identifies comparative performance patterns across models and assesses their suitability for deployment. If models are unsatisfactory — exhibiting poor performance or systematic bias in certain subgroups — the analyst diagnoses deficiencies and returns to specify a different modelling problem. This cycle continues until suitable models are found. Model Selection. When satisfactory models have been identified, the analyst selects preferred models based on prediction quality, error distribution, and deployment requirements, and exports them for use on unseen data.