EUR USD Currency Pair

EUR/USD is the exchange rate between the euro and the US dollar and the single most liquid market in global finance. The BIS Triennial Central Bank Survey put total over-the-counter FX turnover at USD 7.5 trillion per day in April 2022 — the US dollar on one side of 88% of all trades and the euro on 30.5%, which makes EUR/USD the largest currency pair by turnover, ahead of USD/JPY and GBP/USD. FX spot alone turns over roughly USD 2.1 trillion per day (28% of all FX turnover). The market is over-the-counter and effectively continuous from the Monday Sydney open to the Friday New York close, with no central exchange and no single closing auction; dealing is geographically concentrated, with five jurisdictions (the UK, the US, Hong Kong, Singapore and Japan) handling about 78% of all FX trading. In this vault EUR/USD appears as the canonical low-cost test ground for Markov-chain, hidden-Markov and reinforcement-learning trading research.

The reason FX dominates as a testbed is structural: depth and low frictions make it the friendliest possible environment for a quantitative strategy. Institutional EUR/USD spreads run well under one pip — a fraction of a basis point — so the cost hurdle that sinks many equity and crypto backtests is at its lowest here. There is no exchange fee on OTC spot; the cost structure is essentially the bid-ask spread plus, for any position held overnight, a funding (rollover/swap) charge set by the euro-vs-dollar interest-rate differential. That last item matters: it is the same interest-rate differential that drives the Currency Carry Trade, and it means an FX strategy’s profit-and-loss has a carry component distinct from price movement. The implication for grading is two-edged — a costed FX backtest faces a low and well-measured cost hurdle, but a strategy that clears it in FX has cleared the easiest version of the test, not the hardest.

EUR USD Currency Pair [relates] Currency Carry Trade EUR USD Currency Pair [part-of] Futures Markets

The discrete-Markov literature uses EUR/USD precisely because of this clean, deep data. Wilinski 2019 — the only peer-reviewed Markov Chain Trading Model paper in the vault that reports a trading profit — ran its heterogeneous rolling-transition-matrix chain on 60,000 one-hour EUR/USD candles ending 28 April 2016 (alongside a daily WIG20 series). EUR/USD’s role there was to demonstrate that the rolling-matrix chain could “adapt across asset classes” and intraday frequencies; but, as that note records, the reported Calmar-criterion profit was the output of a machine-learning parameter search with no disclosed out-of-sample window, no transaction costs and no benchmark — a weak grade. Hourly EUR/USD trading is exactly where excluding the spread is most misleading: round-trip spread and commission on frequent intraday trades can dominate any gross edge, even in the lowest-cost market in the world.

Wilinski 2019 [tests_strategy] Markov Chain Trading Model EUR USD Currency Pair [supports] Markov Chain Trading Model

The reinforcement-learning literature treats FX as a foundational and recurring test market. Moody and Saffell 2001, the foundational direct/recurrent-RL trading paper, used an intra-daily USD/GBP currency trader with bid and ask prices supplying realistic transaction costs — and reported an encouraging but un-replicated ~2.3 annualised Sharpe over a six-month 1996 window. The most rigorous modern descendant, Borrageiro Firoozye Barucca 2022, trades the major spot currency pairs over a seven-year out-of-sample window with carefully modelled transaction and funding costs, deliberately forcing trades at the 17:00 EST close when costs are statistically highest — and reports a far thinner annualised information ratio of 0.52 with a 9.3% compound return. The two-decade gap between the foundational FX claim and its careful replication is itself a central evidence point of this vault: even in the most liquid, lowest-cost market available, a properly costed out-of-sample RL result is modest, not spectacular.

Moody and Saffell 2001 [tests_strategy] Reinforcement Learning Trading Policy Borrageiro Firoozye Barucca 2022 [contradicts] Moody and Saffell 2001

FX also carries a regime structure that is directly relevant to Markov-style modelling, and a cautionary one. Brunnermeier Nagel Pedersen 2008 documents that carry-trade returns are negatively skewed in investment currencies — exchange rates “go up by the stairs and down by the elevator” — because carry positions unwind suddenly when risk appetite and funding liquidity fall, and a rising VIX predicts carry-trade losses. This means FX has exactly the kind of two-state (calm-accumulation vs crash-unwind) behaviour that hidden-Markov and regime-switching models are designed to detect — but it also means a Markov model fitted on a long calm sample will systematically under-weight the rare, severe crash state, the same Non-Stationarity and rare-event problem that recurs across this vault. FX is a favourable place to test a regime model and a dangerous place to trust one’s calm-period transition probabilities.

Brunnermeier Nagel Pedersen 2008 [defines] Currency Carry Trade Currency Carry Trade [causes] Non-Stationarity

The net reading for this vault: EUR/USD is the deepest, lowest-cost, most-studied test market for Markov and RL trading — which makes it the least demanding venue in which to demonstrate a costed edge, not a hostile one. No FX result tracked here rises above a costed-but-academic backtest: Wiliński’s discrete-Markov profit is weak (ML-tuned, cost-free), Moody & Saffell’s foundational RRL claim is dated and un-replicated, and Borrageiro et al.’s rigorous modern FX RL result is a modest 0.52 information ratio with no live track record. FX favourable conditions do not, on the evidence, convert into demonstrated tradeable alpha.

Connections

• Wilinski 2019 — trades_market, 60,000 1h EUR/USD candles ending Apr 2016, source: https://www.sciencedirect.com/science/article/abs/pii/S0957417419303033
• Markov Chain Trading Model — trades_market, rolling-transition-matrix chain run on hourly EUR/USD, source: https://www.sciencedirect.com/science/article/abs/pii/S0957417419303033
• Moody and Saffell 2001 — trades_market, intra-daily USD/GBP RRL trader, 1996, source: https://ieeexplore.ieee.org/document/935097/
• Borrageiro Firoozye Barucca 2022 — trades_market, major spot pairs, 7-year OOS, IR 0.52 net of costs, source: https://arxiv.org/abs/2110.04745
• Reinforcement Learning Trading Policy — trades_market, FX is a recurring RL test market, source: https://arxiv.org/abs/2110.04745
• Currency Carry Trade — relates, FX funding/swap cost and carry-unwind regime dynamics, source: https://www.nber.org/papers/w14473
• Brunnermeier Nagel Pedersen 2008 — detects_regime, negatively skewed carry-unwind crash regime in investment currencies, source: https://www.nber.org/papers/w14473
• Futures Markets — relates, EUR/USD also trades as a CME-listed currency future inside diversified futures portfolios, source: https://arxiv.org/abs/1911.10107
• Non-Stationarity — suffers_overfitting_risk, calm-sample fits under-weight rare crash regime, source: https://www.nber.org/papers/w14473

Sources

• BIS Triennial Central Bank Survey — OTC foreign exchange turnover in April 2022
• Brunnermeier, Nagel & Pedersen, “Carry Trades and Currency Crashes” (NBER Working Paper 14473, 2008)
• Wiliński, “Time series modeling and forecasting based on a Markov chain with changing transition matrices” (Expert Systems with Applications, 2019)
• Borrageiro, Firoozye & Barucca, “Reinforcement Learning for Systematic FX Trading” (IEEE Access, 2022 / arXiv:2110.04745)
• Moody & Saffell, “Learning to Trade via Direct Reinforcement” (IEEE Trans. Neural Networks, 2001)