To detect abnormal price jumps of ﬁnancial markets, some indicators based on volatility have been used such as the bipower variation and the BPV ratio. However, these indicators only focus on a single individual stock and do not consider the relationships among all individual stocks composing a complex ﬁnancial system. For this reason, we applied an autoencoder to learn the relationships among all stocks, and we considered a stock price that the autoencoder cannot restore as an abnormal price. Moreover, we identiﬁed that the price movement immediately following an abnormal price is clearly biased, and we conﬁrmed the validity of our trading strategy based on this anomaly by performing some statistical signiﬁcance tests.
2016 Research Institute of Signal Processing, Japan