Dynamic Portfolio Management with Transaction Costs

We develop a recurrent reinforcement learning (RRL) system that directly induces portfolio management policies from time series of asset prices and indicators, while accounting for transaction costs. The RRL approach learns a direct mapping from indicator series to portfolio weights, bypassing the need to explicitly model the time series of price returns. The resulting policies dynamically optimize the portfolio Sharpe ratio, while incorporating changing conditions and transaction costs. A key problem with many portfolio optimization methods, including Markowitz, is discovering ”corner solutions” with weight concentrated on just a few assets. In a dynamic context, naive portfolio algorithms can exhibit switching behavior, particularly when transaction costs are ignored. In this work, we extend the RRL approach to produce better diversified portfolios and smoother asset allocations over time. The solutions we propose are to include realistic transaction costs and to shrink portfolio weights toward the prior portfolio. The methods are assessed on a global asset allocation problem consisting of the Pacific, North America and Europe MSCI International Equity Indices.