The present work integrates backpropagation networks into a hierarchical system and develops a simple probabilistic approach to time price forecasts. A standard neural net is critiqued by a second, simpler network, whose ouput is used as the predictor. It is argued that network ouput should not be treated as a single-valued answer, but rather as a variable that determines an empirical likelihood function. The technique has been tried on a portfolio of 34 stocks, trained over a several-year period and tested on approximately a year's worth of daily data. Two conclusions are drawn: (1) Actual returns correlate better with the discrete "critic" predictions than with the "evaluator" net's price estimates, and (2) the set of probability distributions p(r|s) contains significantly more information than a random walk distribution based on overall stock volatility.
展开▼
