A Novel Implementation of Siamese Type Neural Networks in Predicting Rare Fluctuations in Financial Time Series

Joshua Ward, Treena Basu, Olaf Menzer, Indranil Sengupta

Abstract:

Abstract: Stock trading has tremendous importance not just as a profession but also as an income
source for individuals. Many investment account holders use the appreciation of their portfolio (as
a combination of stocks or indexes) as income for their retirement years, mostly betting on stocks
or indexes with low risk/low volatility. However, every stock-based investment portfolio has an
inherent risk to lose money through negative progression and crash. This study presents a novel
technique to predict such rare negative events in financial time series (e.g., a drop in the S&P 500 by a
certain percent in a designated period of time). We use a time series of approximately seven years
(2517 values) of the S&P 500 index stocks with publicly available features: the high, low and close
price (HLC). We utilize a Siamese type neural network for pattern recognition in images followed
by a bootstrapped image similarity distribution to predict rare events as they pertain to financial
market analysis. Extending on literature about rare event classification and stochastic modeling in
financial analytics, the proposed method uses a sliding window to store the input features as tabular
data (HLC price), creates an image of the time series window, and then uses the feature vector of
a pre-trained convolutional neural network (CNN) to leverage pre-event images and predict rare
events. This research does not just indicate that our proposed method is capable of distinguishing
event images from non-event images, but more importantly, the method is effective even when only
limited and strongly imbalanced data is available.