| Abstract [eng] |
This paper develops a systematic algorithmic portfolio management framework, which consists of 4 core parts: information extraction, asset preselection, portfolio optimization and online regulation. As a case study for the proposed framework, we use Markov Chains model to extract information from the assets return series and forecast the future wealth distribution attainable by each asset. We utilize these forecasts to reduce dimensionality of the portfolio optimization problem via asset preselection using two algorithms: Data Envelopment Analysis and simple ranking algorithm of our own design. The portfolio weights are optimized by maximizing 3 semi-parametric functions of the forecasted portfolio wealth distribution, prevalent in the related literature, and a proposed extension of Omega Ratio to the Markovian, which demonstrates good performance in the empirical tests. Optimization is performed using two heuristic algorithms -- Genetic Algorithm and a problem-specific simplex exploration algorithm, the latter of which consistently demonstrated better performance. Moreover, we propose and validate a simple regularization method in order to control the portfolio weight redistribution in time performed by the considered model. Within the developed framework, we were able to construct multiple end-to-end portfolio management algorithms that outperformed the benchmark S\&P500 Index both under favorable market conditions and during the 2022 recession. |
