Advanced Analytics Group has released Auguri 2.1, an integrated Windows data exploration, analysis, and forecasting tool with emphasis on nonlinear dynamical methods. Its purpose is to provide the tools for the manipulation and analysis of data through the process of predictive data mining.

Auguri reads data from most sources in ANSI or binary format, as well as its own format. They are also working on the ability to comply with the Predictive Model Markup Language (PMML).It supports drag and drop, as well as copy and paste from outside sources such as spreadsheets. It maps data to worksheets, where a worksheet may consist of one or more sheets containing data, models, solutions, and reports.

Along with common statistical analysis methods such as ANOVA, test of means, and variances, Auguri also provides extensive nonlinear methods, such as generalized fractal dimensions, Poincare surface of sections, maximal Lyapunov exponent, false nearest neighbors, space-time separation plots, averaged mutual information, and phase portraits in up to four-dimensions, as well as others.

Auguri also provides tools for the analysis of signals and series in the time and frequency domains, such as power spectrum estimation, Fourier transforms, auto- and cross-covariance and correlation functions, time-evolving statistics, and simultaneous
solutions to linear equations.

In addition, Auguri includes several methods for the generation of random numbers according to a chosen distribution, for sampling data from existing populations, and for generating surrogate data, where statistical and nonlinearity tests may be additionally carried.

With Auguri, you begin by inspecting and visualizing the data. The information is plotted to get a general idea on its format and dependency; examining the plot for trends, cycles, and missing values. After that, you typically continue to prepare the data by removing observed cycles and trends, replacing missing values and outliers, and centering and, optionally, normalizing the data. A spectral analysis may be desirable to detect noise, and remove it via filtering.

With this step done, you can continue with data reduction. Here the data is searched for dependencies (correlations), discarding irrelevant information before proceeding to create a model that explains the system. At this point, data may be split into in-sample and out-of-sample sections. The in-sample part serves to find the described dependencies and create the model, the out-of-sample one, for testing the predictive power of the proposed model. If, after testing, a model is deemed unsatisfactory, we may be tempted to go back and propose a different model, repeating these steps until we are satisfied we have achieved the best performing model.