A portfolio of 200 heterogeneous technical trading rules is tested for their directional predictabilities on the DJIAI from 1988 to 1999. We also explore several nonparametric techniques designed for brain research and detected possibly other forms of dependencies more significant than the traditional linear autocorrelation for the time series. The overall conditional mean directional predictability is 46%. 36 percent of the rules have more than 50% directional predictability, and the top 20 percent rules has a 73% directional predictability, whereas the bottom 80 percent has a directional predictability of 40%. Buy signals consistently generate higher predictability than sell signals but do not commensurate with their respective risk levels. The relationship between two sub-periods is not stable, while the difference between the conditional mean directional predictability of buy only and sell only signals is highly significance. The belief that most successful rules have a directional predictability of 25% to 50% coincides with the mode of distribution. We observe counter intuitive relationship between volatility and directional predictability. The results of directional predictability in a downtrend concur with the argument that buy-and-hold strategy is not a suitable benchmark. Attempts are made to tackle the issues of small sample bias, data snooping, size of test window, bootstrap or t-test, and homogeneity. Issues are discussed on empirical testing for their real-world applications, statistical and non-statistical interpretations; also, randomness test; physical or biological science approach.