A0901
Title: Functional motif discovery in stock market prices
Authors: Marzia Cremona - Universite Laval (Canada) [presenting]
Lyubov Doroshenko - Université Laval (Canada)
Federico Severino - Universite Laval (Canada)
Abstract: Financial asset prices display recurrent patterns over time. However, such time series are usually noisy and volatile, making the identification of repetitive patterns particularly difficult. These motifs are rarely exploited for price prediction, even though some of them, such as the surge of a financial bubble, occur periodically and feature similar shapes. Asset prices are embedded in a functional data analysis framework by extending and using probabilistic K-means with local alignment to discover functional motifs in stock price time series. The information of the discovered motifs is then exploited to perform the price forecasts with a novel motif-based (MB) algorithm introduced. After illustrating the technique on simulations of mixed causal-noncausal autoregressive processes, it is applied to the prices of S\&P 500 top components and motif-based forecasting is performed. Finally, its performance is compared to traditional forecasting models.
