Abstract
We adopt Deep Reinforcement Learning algorithms to design trading strategiesfor continuous futures contracts. Both discrete and continuous action spacesare considered and volatility scaling is incorporated to create rewardfunctions which scale trade positions based on market volatility. We test ouralgorithms on the 50 most liquid futures contracts from 2011 to 2019, andinvestigate how performance varies across different asset classes includingcommodities, equity indices, fixed income and FX markets. We compare ouralgorithms against classical time series momentum strategies, and show that ourmethod outperforms such baseline models, delivering positive profits despiteheavy transaction costs. The experiments show that the proposed algorithms canfollow large market trends without changing positions and can also scale down,or hold, through consolidation periods.