A u t o R e l a t i o n   A n a l y s i s ™
 Aura®
The expert system in time series forecasting

You can find more detail about this project and download free demo on my special site dedicated to this product.

The problem

Every day many people working on quite different tasks - managers in finance departments, traders on a stock market, economists at government agencies - encounter a sort of tricky puzzle. They have to discover, as quickly as possible, the likely tendencies in large domains of poorly organized data. In many cases, the stochastic nature of the underlying datasets suggests that the statistical forecast can give the reasonable solution of the problem.

Known solutions

While there are many computer programs aimed to resolve this problem, they often fail to meet expectations of a novice user at least in the following ways:

• Tightly specialized applications can effectively forecast only a narrow set of typical situations. The nuisance is, situations tend to change swiftly.
• Diverse statistical analyzers require at least a moderate knowledge of the algorithms they expose. Such knowledge costs time and money.
• Powerful and flexible neuronets appear quite dumb on a short data series. In fact, additional data happen to be expensive, if available at all.

Concept

To consolidate and mutually reinforce the above approaches we suggest the universal shell built up as an expert system and aimed to combine the most popular forecasting algorithms in the automated competitive environment. Given a multidimensional time series, the system automatically hypothesizes on a set of all available solutions seeking to minimize the aggregate difference between backforecasts and the actual values of the supplied series. The most effective hypotheses then integrate into the final model which, in turn, is used to predict future responses.

Interface

The friendly user interface allows for visual definition of datasets as well as imports from most popular spreadsheet formats. The user needn't to worry about preparing every time series to be of equal length with others. Moreover, the time step of each series in the model might differ drastically. All these complications are automatically detected and handled. Once the dataset was defined, the user is on one mouse click distance from the forecast, which is ready available both in numerical form and as a chart with variance estimates. All backforecasts are also supplied to illustrate the quality of the final model. More experienced users will be surprised to learn all the subtleties in the exhaustive report on the model found. Furthermore, the experts are provided with a full set of tools for manual tuning of automatically generated model.

Specification

With its breakthrough data mining technology Aura® Forecast Engine is able to deal effectively with data sets of practically any length and dimensions. The exact characteristics are listed in the following table:

Algorithms

At present, the project supports the following:

1. Input control procedures for automatic conversion of raw data into a data series.
2. Monitoring and smart filtering of extreme values within data series.
3. Mutual synchronization and aggregation of datasets.
4. Multidimensional analysis of distributed lags.
5. Simple linear regression (for instant prediction of a very short time series).
6. A number of statistical tests, which detect seasonal components and estimate their characteristic times.
7. Non-seasonal ARIMA models for individual forecasting of medium and long time series.
8. Adaptive selection of seasonal ARIMA models.
9. Multiple linear regression analysis with confidence intervals for the future responses.
10. Selection of the regression model using a forward stepwise algorithm.
11. Leaps and bounds algorithm for determining a number of best regression subsets from a full regression model.
12. Special set of multivariate tests, which assist combining the above methods into a most effective model.

While most of the above algorithms are widely known and available, the real power of our solution proceeds from the ability to automatically merge these computational methods into a flexible and effective forecast model.

Staff for developers

To meet the requests of software developers the product is available as a structured set of object modules ready to link with Visual C++. A simple interface object exposed as a class of over 30 simple high-level functions gives you the real feel and power of data mining and statistical forecasting in almost infinite dimensions.

Customers

In the short time (two months after the first beta-release) the product has won the sympathies of the Russian government agencies. In particular, the Ministry of Economy of Russian Federation has successfully integrated this software into its networking computational environment. As a part of the Intranet, the product supports the transparent access and automated analysis of corporate databases of the Ministry.