Borrageiro Firoozye Barucca 2022
“Reinforcement Learning for Systematic FX Trading” by Gabriel Borrageiro, Nick Firoozye and Paolo Barucca (University College London), published in IEEE Access, vol. 10, pp. 5024-5036, 2022 (arXiv:2110.04745). It applies a direct/recurrent reinforcement learning agent — explicitly in the lineage of Moody and Saffell 2001 — to the major spot FX currency pairs, with online inductive transfer learning from a radial-basis-function / Gaussian-mixture feature network and a quadratic utility through which the agent learns to target a risk position directly.
It is the most credible modern, cost-realistic test of the foundational Recurrent Reinforcement Learning Trading claims. Over a seven-year out-of-sample window, with transaction and funding costs accurately accounted for and trades deliberately forced at the most expensive 5pm EST daily close, the agent achieves an annualised portfolio information ratio of only 0.52 and a 9.3% compound return net of costs. That is positive but modest — roughly a quarter of the ~2.3 annualised Sharpe ratio Moody & Saffell reported for their six-month 1996 USD/GBP study. The paper therefore functions in this vault as an honest replication-class data point: direct RRL on FX can be slightly profitable net of realistic costs over a long horizon, but the dramatic risk-adjusted returns of the 1998-2001 papers do not reproduce. Graded moderate — out-of-sample, realistic costs, multi-year, but a thin edge and a single study.
Borrageiro Firoozye Barucca 2022 [contradicts] Moody and Saffell 2001 Borrageiro Firoozye Barucca 2022 [tests_strategy] Recurrent Reinforcement Learning Trading
Connections
- Recurrent Reinforcement Learning Trading — tests_strategy / reports_profitability / includes_costs, 2022, source: https://arxiv.org/abs/2110.04745
- Moody and Saffell 2001 — reports_underperformance / replication_missing (far weaker net result than the foundational claims), source: https://arxiv.org/abs/2110.04745
- Reinforcement Learning Trading Policy — tests_strategy, source: https://arxiv.org/abs/2110.04745
- Transaction Costs and Slippage — includes_costs, source: https://arxiv.org/abs/2110.04745
- Out-of-Sample Backtesting — compares_benchmark, source: https://arxiv.org/abs/2110.04745