Parasite trading model using neuro-fuzzy system

Abstract

The popular use of program trading in current financial market, had gained a lot of attention by the public. However, the rule-based approached often determined by experience traders, does not fully utilized the capability of computerized trading. Not only so, many of the neuro-fuzzy architecture implemented for financial application is focusing on Time-Series Price Prediction.

Predicting the impact of anticipatory action had been proposed and 3 days before event day had being found to be the entering day for smart investors. However, the anticipatory model lacks an exit strategy as it assumes that exiting on event day is the best timing. Datasets used are also found to be too short (3 years) for meaningful results. Moreover, the idea lacks practical result on applying to financial domain, similarly to Price Time-Series approach (SeroTSK and GSETSK).

This paper sought to address the issues, in addition improve with the use of 3 additional machine learning techniques, eMFIS, RBFN, and Q-SVM. The dataset is updated with 15 years of data points from BLOOMBERG. New set of predictors are used that obtain higher accuracy than previous reports. Exit strategy are also defined to be Event Day to Event Day + 3 Days period, due to the trading day + 3 days is the settlement period and is the perceived holding period for feedback traders. During the analyzing of test result we also found weakness in predicting in Indices and combined events data set, showing that individual stocks and event types data set is recommended for future trading model.

Directional change of Event and Exit Day for trade was predicted from our parasite trading model, the results were used against Price Time-Series approach (SeroTSK and GSETSK) in a benchmarking, to analyze the practical result on applying our proposed parasite model in financial domain.