Stock prices prediction via tensor decomposition and links forecast

Many complex systems display fluctuations between alternative states in correspondence to tipping points. These critical shifts are usually associated with generic empirical phenomena such as strengthening correlations between entities composing the system. In finance, for instance, market crashes are the consequence of herding behaviors that make the units of the system strongly correlated, lowering their distances. Consequently, determining future distances between stocks can be a valuable starting point for predicting market down-turns. This is the scope of the work. It introduces a multi-way procedure for forecasting stock prices by decomposing a distance tensor. This multidimensional method avoids aggregation processes that could lead to the loss of crucial features of the system. The technique is applied to a basket of stocks composing the S&P500 composite index and to the index itself so as to demonstrate its ability to predict the large market shifts that arise in times of turbulence, such as the ongoing financial crisis.