Real Edges in Retail Forex (2025-2026): Evidence-Based Report

Treat this as a friend's honest read — not marketing. The headline: most retail "AI forex" is dressed-up curve-fitting on top of edges that are real in institutional form but mostly evaporate once a retail account pays the spread.

1. Documented Edges (and What's Decayed)

1.1 Carry Trade — real, but compressed and crash-prone

The most heavily documented FX anomaly. Lustig, Roussanov & Verdelhan (2011, NBER) showed currency excess returns are highly predictable and counter-cyclical — high-rate currencies earn ~4.8% p.a. more than low-rate currencies after transaction costs (NBER w14082).

Sharpe estimates vary widely by methodology and sample:

Naive equal-weighted carry: Sharpe ~0.76 in calm periods
Forward-discount-adjusted carry: Sharpe ~0.99
Volatility-adjusted: Sharpe ~0.84
Including crisis years (2008, 2011, 2015-16, COVID 2020, Aug 2024 yen unwind): drops to 0.3-0.5
"Hedged against unpriced risks" academic version: up to 1.29 (Quantpedia FX Carry)

The honest read: Carry is real but its Sharpe is a fiction in the first moments — the return distribution is famously left-skewed ("picking up nickels in front of a steamroller"). Drawdowns of 15-20% inside 4-6 weeks happen on a roughly decade cadence. The Aug 2024 yen carry unwind is the most recent reminder (mas-markets).

Retail viability: Marginal. The G10 carry spread compressed dramatically 2008-2022 (near-zero rates everywhere), expanded 2022-2024, and is compressing again as the ECB cuts. Retail brokers' overnight swap rates are systematically worse than the actual interbank carry — you're already paying 20-40% of the theoretical edge to the broker.

1.2 Time-Series Momentum (Trend) — strongest published FX edge

Moskowitz, Ooi & Pedersen (2012) "Time Series Momentum" is the canonical reference: across 58 instruments including 9 currency pairs, 1965-2009, Sharpe ratios above 1.20 for diversified TSMOM portfolios (Moskowitz et al). 12-month lookback, monthly rebalance.

The catch: that 1.2+ Sharpe is diversified across 58 markets. Single-pair FX TSMOM is dramatically lower — somewhere in the 0.3-0.6 range, and that's before retail spreads.

1.3 Cross-Sectional Currency Momentum — published, partly decayed

Menkhoff, Sarno, Schmeling & Schrimpf (2012) found a 10% p.a. cross-sectional spread in a long-short portfolio of 40+ currencies, 1976-2010 (Quantpedia Currency Momentum).

Post-publication decay is real and measurable. Equity momentum returned ~10% in the 1990s and ~2% today (Research Affiliates). FX momentum has followed a similar arc — visible in Lustig-Verdelhan replications showing "decreasing strategy performance" post-2011 publication.

Retail blocker: Cross-sectional needs 20+ currencies, including emerging-market crosses, which retail brokers charge brutal spreads on (USDZAR, USDTRY can be 10-50 pips at retail vs. 0.5-2 institutional).

1.4 FX Value (PPP-based) — long-horizon only

Real exchange rate deviations from PPP mean-revert over 3-5 year horizons (Quantpedia FX Value). The factor exists, but holding-period requirement makes it nearly useless for a retail trader who'd compound 1095 daily spread/swap costs against it.

The combined "FX Carry + Momentum + Value" diversified portfolio is where AQR/Two Sigma/Bridgewater earn their money — Sharpe close to 2 reported on ~100 liquid futures/forwards/swaps over 25 years (Quantpedia 200-year study). Retail cannot replicate this without access to NDFs, emerging-market forwards, and prime-broker spreads.

1.5 Mean Reversion — works on cointegrated crosses, marginal elsewhere

Documented only weakly as a standalone edge in liquid majors. Where it works: cointegrated crosses (AUD/NZD, EUR/CHF pre-2015 floor removal, USDCAD/AUDUSD pairs trading). A published exponentially-sized FX-futures mean-reversion strategy: Sharpe ~0.35 (ScienceDirect Serban).

One vendor backtest claims 11% annual / 0.8 Sharpe / 11% DD across EURUSD, GBPUSD, USDCAD, USDJPY combined mean-reversion + momentum, but this is unaudited.

1.6 News/Event Trading — real volatility, brutally hard to extract

NFP, FOMC, CPI, ECB cause measurable 50-100+ pip spikes. But: "spike 80 pips one direction after NFP, only to completely reverse within 15 minutes" is a common pattern (QuantifiedStrategies NFP). At release, spreads widen 5-10x and slippage on retail accounts is severe. Most published "NFP strategies" don't survive realistic execution assumptions.

The only retail-viable version is pre-positioning ahead of statistically biased outcomes (e.g., FOMC drift, post-NFP day-2 fade), and even those are crowded.

1.7 Breakouts (London/Asian session) — mixed, mostly overfit

The London Breakout literature: "Backtesting on EUR/USD often resulted in losses... be skeptical of backtests showing 5-10% monthly returns — likely overfitted" (QuantifiedStrategies London). One trader's GBPUSD 5-min version showed >50% win rate at 1.5 R:R, but this is single-instance, not a published edge.

Honest assessment: This category is where ~90% of retail "strategies" live, and where overfitting is most rampant. Treat any claimed breakout edge as guilty until proven via genuine walk-forward + transaction-cost-realistic backtest.

1.8 Central Bank Divergence — macro-level, not signal-level

The 2024-2025 macro environment (Fed steady, ECB/BoE cutting, BoJ hiking) created real FX trends (Fusion Markets). EUR/USD downtrend, USD/JPY volatility, etc. But "divergence" is a story, not a systematic signal — by the time you can name it, the move is mostly priced. Macro discretionary funds extract this; systematic implementations are essentially carry + momentum with extra steps.

2. AI/ML in FX: Hype vs. Reality

2.1 LSTM/Transformer price prediction — published "wins" rarely survive

The literature is full of papers claiming high R² and low RMSE: e.g., "feature-augmented multivariate LSTM, R²=0.94, RMSE=0.000127" (ScienceDirect 2025). These metrics are misleading. A random walk with drift achieves R² > 0.99 on prices — what matters is return prediction, not price prediction, and almost no paper is honest about this distinction.

The 2024 controlled comparison study (Comparative Analysis LSTM/GRU/Transformer): under walk-forward out-of-sample protocol, "ARIMA and Random Forest remain strong baselines; deep learning shows asset-dependent performance, LSTM occasionally competitive, Transformer competitive but not uniformly dominant." Translation: no clear edge from deep learning on FX prices.

2.2 Reinforcement learning — backtest-impressive, sim-to-real catastrophic

Published RL papers regularly claim "improved overall returns from -25.25% to 14.86%" on EUR/USD (RL FX 2024). The critical caveat from the literature itself: "Many studies rely on simplified simulators that under-model execution friction, financing costs, and liquidation mechanics, introducing a significant sim-to-real gap that often yields strategies economically brittle outside controlled simulations."

The smoking gun for retail RL skepticism: JP Morgan launched "DNA" (Deep Neural Network for Algo Execution) in 2019 and shut it down around Oct 2023 citing "issues with data interpretation and the complexity involved" (FX Markets). If JPM's quant team with their tick data and prime broker spreads couldn't make deep learning work for execution (a much easier problem than alpha), retail RL alpha claims should be treated with maximum skepticism.

2.3 Sentiment NLP — modest edge, mostly at institutional latency

Real result: GPT-4 outperforms traditional NLP on FOMC hawkish/dovish classification — but the F1 score is 0.57 (GPT Fedspeak). That's barely above coin-flip on multi-class. The 11% improvement on short-term trend prediction vs. traditional sentiment benchmarks (RavenPack) is real but small — and that benchmark is already in the noise floor of execution costs at retail spreads.

2.4 LLM trading agents — actively debunked in 2025

The most important paper from this cycle: "The Alpha Illusion: Reported Alpha from LLM Trading Agents Should Not Be Treated as Deployment Evidence" (arxiv 2605.16895). Direct quote: "Many LLM trading papers report high Sharpe ratios over short windows, but the financial econometrics literature has established that short-window Sharpe estimates carry substantial uncertainty and repeated search inflates the false-positive rate."

LiveTradeBench (Nov 2025) found LLMs struggle to distinguish "attention-grabbing but non-decisive news" from genuinely influential events, leading to overreaction (arxiv 2511.03628).

2.5 What hedge funds actually do

Renaissance, Two Sigma, DE Shaw use ML — but mostly for:

Execution (TWAP/VWAP/POV optimization, market-impact modeling) — not for alpha
Feature engineering at scale across thousands of cross-asset signals
Alternative data fusion (shipping, satellite, payment, web scrape)
Portfolio construction (Markowitz on ML-predicted covariance)

Renaissance's structural edge is "everything unified under one model with all resources behind one arrow" (INN). A solo retail trader running an LSTM on EUR/USD 15-min bars is not competing with this. They're competing with the same retail signal services that lose 86% of accounts.

3. Where Retail Loses Systematically

The lose-rate numbers are not folklore — they're disclosed under ESMA MiFID II. Selected hard numbers:

74-89% of retail CFD accounts lose money (ESMA, mandatory broker disclosure since 2018)
86% of all retail forex traders lost money in 2024 (broker-aggregated)
CFTC: ~75-80% US retail forex loses over time; 2 of 3 lose each quarter
Only ~1% sustain profit for 5+ consecutive years
Average retail FX account lifespan: ~4 months
Average loss range: €1,600 - €29,000 per client
72% of retail FX traders have no prior experience in other financial markets
Sources: ESMA disclosures via traderslog, zipdo, compareforexbrokers

Why retail loses (mechanical, not psychological):

Spread + commission — A 1-pip spread on EUR/USD is ~0.7 bps. A retail scalper trading 10x/day pays 700 bps/year in friction alone. That's the entire expected return of carry-trade-on-leverage.
Slippage — "Most underreported cost in retail trading. Switching from 1.0 pip to 0.7 pip spread saves 0.3 pips, but 0.5 pips of execution slippage erases the saving" (OANDA).
Overnight swap markups — Brokers skim 20-40% of true interbank carry.
Leverage-induced ruin — Retail brokers offer 30:1 (EU) to 500:1 (offshore). Even a Sharpe-1.0 strategy at 30x leverage has a meaningful probability of ruin over 12 months.
B-book broker conflict — Most retail brokers internalize flow; your losses are their P&L. This isn't conspiracy — it's their disclosed business model.

Signal services

The 20-Telegram-providers analysis found virtually none had Myfxbook/FxBlue verified track records (Medium analysis). The base rate for "AI forex signal" services being legitimate is, generously, single-digit percent. Most are affiliate funnels to B-book brokers — they get paid on your losses regardless of signal quality.

4. What's Actually Changed 2023-2026

New edges that emerged:

Sub-second Fed communication parsing — LLMs can classify FOMC statement tone faster than a human can read it. Real but institutional-only (you need direct-feed news, co-located execution). Bloomberg/RavenPack/Quiver Quant sell this; retail latency to a CFD broker kills any edge.
Aug 2024 yen carry unwind — created the largest single-week FX dislocation since 2008 (mas-markets). Carry-monitoring with position-unwind detection became live alpha. Decaying fast.
CB divergence regime 2022-2025 — genuine macro trends restored time-series momentum performance after the 2010s flatline. Already crowded again.

Edges that died or compressed:

Pure technical breakouts on majors — algorithmically saturated. EBS estimates 30-35% of platform volume is HFT; any visible breakout level has been arbitraged before retail can see it.
Simple carry on G10 — compressed by ZIRP era (2010-2022), re-opened briefly, compressing again.
Naive cross-sectional momentum — measurably decayed post-Menkhoff et al 2012 publication.

What hasn't changed:

The 70-90% retail lose rate has been remarkably stable across 15+ years of ESMA, CFTC, ASIC data. The mix of edges that work shifts; the structural disadvantage of leveraged retail accounts does not.

5. Honest Bottom Line for an AI-Assisted Retail FX Build

What's worth building:

Diversified TSMOM/carry overlay on a portfolio of 8-15 majors + EM crosses you have decent spreads on. Target Sharpe 0.4-0.6 net of costs, not 2.0. Anything higher in backtest = you've overfit.
LLM as research assistant — parsing central bank statements, screening news, building features. Not as the trader itself. The Alpha Illusion paper is now the citation to defend this position.
ML for execution and risk (vol forecasting, position sizing, regime detection) — this is where modest ML wins are real and don't require alpha to be correct. JPM's pivot away from deep-learning-for-execution is a warning even here; stick to simpler ML (gradient boosting, regime HMMs) before going deep.

What's almost certainly bullshit if pitched at you:

"AI forex signal" Telegram/Discord services. Base rate of legitimacy: ~0-5%.
LSTM/Transformer "predicts EUR/USD with 94% accuracy." Always a confused R² on prices, not directional accuracy on returns net of costs.
Any backtest without realistic spread/slippage, walk-forward, and out-of-sample purged CV (Lopez de Prado CPCV) (Surmount).
"Quantum AI" anything in this space — pure marketing.

The realistic upper bound for a sophisticated retail AI-assisted FX system, honestly executed, is probably a Sharpe of 0.5-0.8 net of costs on a multi-strategy portfolio, with annual drawdowns of 15-25%. That's a real outcome — better than the 86% lose rate — but it is not "AI alpha." It is disciplined factor exposure with ML on the periphery.

Edges in FX: which ones survive the evidence