Singular Value Decomposition as a Method for Analyses and Forecasts of Financial Data

Sammanfattning: This paper examines the sufficiency of a trading method based on singular value decomposition (SVD) of past stock prices. The SVD method is frequently used as a tool to reduce data noise, compress big-data, and analyse data components. Hence, the method is well suited to form a ground for a predictive tool of price developments. From the predicted pattern, a strategy was formed by construction of a portfolio of two business sectors concluded to be negatively correlated in one of the price movement components. An algorithm was programmed to receive buy- and sell signals when the difference in the price gradient exceeded a fixed value. The active portfolio was bench-marked against a buy-and-hold strategy for a portfolio consisting of the same stocks and weights. A set of 93 stocks from the NYSE were selected and divided into groups to rep- resent a variety of business sectors in the market. The active strategy showed, in the simulation period 2014-2018, to at best have an average annual excess return to the passive portfolio of 13, 66%. The strategy performance was improved with stronger negatively correlated sector pairs in price components of sizeable significance. It should be emphasised that the paper does not account for trading costs and market risks. However, the general conclusion is assumed not to be affected by the absence of incorporated trading costs. In addition, the strategy assumes an investor would follow the strategy over the entire simulation period, and not interfere with the trading algorithm, hence only be affected by the results on the last trading day.