Deep Learning For Portfolio Optimization With Delta Controlled

Abstract

The cryptocurrency market is considered high-risk compared to traditional investment channels such as stocks or bonds. The whole market trend is primarily affected by only a few top-of-the-market capitalization cryptocurrencies like Bitcoin or Ethereum. This significantly affects the ability to diversify a portfolio of cryptocurrencies. In 2020, Binance introduced a new set of derivative assets called Binance Leveraged Tokens that allow traders to perform hedging. While there have been a few attempts to develop Machine Learning algorithms for portfolio optimization or hedging, no work takes advantage of this new class of assets in a data-driven manner to make profits. In this thesis, the author proposes a deep learning based algorithm to take advantage of the above-mentioned investment product. The proposed algorithm constructs a portfolio that contains a pair of Binance Leveraged Token. A deep neural network generates the allocation weights for each asset in the portfolio, and the asset allocation is re-adjusted at a certain timeframe. The neural network is optimized to maximize the Sharpe ratio of the portfolio. In addition, we propose additional loss terms that regulate the network's bias towards a specific asset. These additional loss terms help the network to learn an allocation strategy that is close to a delta-neural strategy, which is a strategy that preserves the value of the portfolio over time. In order to validate the proposed method, the author conducted an extensive set of experiments using the data from Binance spanning 20 months. The experiment results demonstrate that the proposed loss terms can enable the neural networks to make profits in different market situations.