NeuroBayes - a modern analysis technique and its way as prognosis tool into business

Michael FEINDT
U Karlsruhe (KIT)

NeuroBayes is an algorithm based on neural networks and Bayesian
statistics which can learn complex multivariate dependencies from
historic or simulated data and use this as a basis for predictions of
future events. In particular it is optimized to distinguish
statistically relevant dynamics from stochastic noise. Next to
classification tasks also complete probability densities for real-valued
quantities can be predicted. On the basis of this optimal decisions can
be made for arbitrarily given cost functions.

After an introduction into the technology a number of different example
applications from experimental particle physics are presented, ranging
from particle identification to optimization of resonance signals, the
separation of quark- and antiquark-jets, optimization of energy, angle
and lifetime measurements as well as spin-parity-analyses.

The company Phi-T GmbH, founded and run by particle physicists, performs
NeuroBayes analyses for private economy. A number of very successful
projects are described. For insurance companies the contract
cancellation probability, the accident risk and the cost distribution in
case of an accident are predicted for individual customers. Other
examples are short-time predictions and risk prognoses of financial time
series for banks, fraud detection, turnover prognoses for individual
shops of chain stores or individual articles of a mail order trading
company.