Generic Multi-Expert Discovery

Any field of science, business, and technology dealing with complex systems has its own set of challenging problems that are critical for real-life applications and still lacking acceptable universal solution. This is especially relevant for practical decision support systems where accurate and timely forecasting or pattern identification are crucial. Explosive growth of data in many scientific and business fields dealing with complex systems opens new perspectives for quantitative and computational solutions in many challenging applications.

However, even with apparent abundance of data, many typical problems of complex system modeling such as “curse” of dimensionality and non-stationarity prevent discovery of accurate and stable data-driven solutions for many important applications.

Simplified analytical and computational models based on existing knowledge and heuristics of the considered field (domain) require much less data for calibration. However, such models often have limited accuracy due to underlying simplified assumptions and can be applied only in limited range of dynamical regimes. Limitations of such simplified models can be compensated by model combination via different ensemble learning techniques.

Our generic boosting-like model combination and optimization framework for the discovery of robust multi-expert models and strategies combines the best existing machine learning and statistical approaches. In addition, it includes our proprietary regularization techniques, multi-objective functions and other critical algorithmic and empirical know-how. More importantly, our collection of base models and strategies from considered fields (domains) is constantly expanding. We are continuously screening novel results reported by domain experts in conference and journal publications and other sources. Any interesting new approach proposed by others or by our own research efforts is tested for complementary value to other models and strategies.

In contrast to questionable search of the best single model pursued by many researchers and practitioners, the main goal of our system is optimal combination of complementary models (experts). Such models, empirical rules, practitioners’ heuristics and other results often remain unnoticed or underutilized if their global (overall) performance is inferior to the best state-of-the-art solutions. However, our system detects and incorporates such models and other knowledge based on their ability of improving performance of the overall model ensemble. This approach extracts maximum benefit from all kinds of existing quantifiable knowledge. Our generic data-driven and knowledge-driven framework has proven to be effective in several challenging financial, business intelligence and biomedical applications.

Several of cutting-edge solutions discovered by our framework for financial, biomedical, and sport applications are available for evaluation and usage by our clients on 24/7 basis. For a short overview please go to the appropriate section (biomedical, finance, etc.). Please contact us with your questions and to receive your free password. We could also consider developing new nonstandard quantitative solution tailored for your needs.