Deep learning methods for constructing sparse mean-reverting portfolios and cointegration of financial time series
PI: Dr. Zhiwen Zhang, HKU
Convergence trading was made popular by the hedge fund Long-Term Capital Management (LTCM), which benefits from the phenomenon that the price of a portfolio fluctuates around a certain level. Since deviations from this level are temporary, investors can build appropriate trading strategies accordingly. Ideally, convergence trading is market-neutral and investors will always make profits if this statistical arbitrage happens. In practice, however, the expected convergence may not happen, or it may diverge before converging. The near-collapse of LTCM in 1998 demonstrates this serious risk. Moreover, investors prefer a sparse portfolio since sparsity means fewer transaction costs. Many methods have been developed to study convergence trading, including its existence and convergence speed. However, many methods become expensive for large-scale problems.
In this project, we develop novel deep learning methods to address these issues. Specifically, we will construct sparse mean-reverting portfolios and study the cointegration of financial time series. We will evaluate the performance of our trading strategies using real financial data in the Hong Kong stock market. Due to its powerful approximation ability, we expect that deep learning methods can efficiently solve these two problems. Thus, the proposed project is promising in generating broader impacts in the financial engineering community.