| Uygun et al.(2025) DOI: 10.1007/s00521-025-11586-8 |
FX & Crypto |
100 |
170 |
35.70% |
27.49% |
0.76 |
104 weeks (Ending 2023) |
Not given |
54.95% |
MTGNN (Multivariate Graph Neural Network with temporal convolutions) |
| Wangchailert et al. (2025). DOI:10.37936/ecti-cit.2025191.256994 |
8 pairs EUR/USD |
Not given |
+$10,999.00 |
$916 per year |
Not given |
Not given |
2010–2022 (13 years) |
28,525 |
51.3% |
Candle pattern based trading |
| Papatsimpas et. al (2025) DOI:10.1007/s10898-025-01505-5 |
EUR/USD |
Not given |
Not given |
+2198.00 pips (Profit) |
-95.90 pips |
N/A |
July-Oct 2022 (64 days) |
36 |
86.11% |
Hybrid EMD + Stacked LSTM + PSO aggregation optimization |
| López-Herrera, et al. (2025) DOI: 10.1007/s44163-025-00424-4 |
USD vs EUR, CNY, JPY, AUD, CHF, MXN, ZAR, and TRY |
Not given |
Not given |
31.1% (TRY/USD Max) |
Near-Zero |
3.38 (Median Bootstrap) |
2018–2023 |
35 (Avg per period) |
71.1% |
Logistic Regression with Mean Absolute Directional Loss (MADL) optimization |
| Iswara et al. (2025)DOI: 10.1109/ISCT66099.2025.11297372 |
XAU/USD |
Not given |
(Net Profit: +$1.53) |
Not given |
$0.41 |
Not given |
4 to 5 hours |
26 |
57.69% |
LSTM + Inductive Conformal Prediction (ICP) uncertainty filter |
|
Zhang & Khushi (2020), “GA‑MSSR: Genetic Algorithm Maximizing Sharpe and Sterling Ratio…”
|
EUR/USD |
100,000 |
187,900 |
+13.4 |
‑19.7 |
1.68 |
2010‑01‑01 → 2019‑12‑31 |
2,305 |
61.0 |
SS Ratio‑optimised GA |
| Arabha, Sarani & Rashidi-Khazaee (2024). arXiv:2411.01456 |
EUR/USD |
Not given |
Not given |
42.22% (DS2 OOS, 7 months) |
Not given |
0.47 (DS2 OOS) |
DS1: Jan–Jul 2017; DS2: Jan–Jul 2023 |
Not given |
Not given |
PPO + Auxiliary Task (PPO+AXT); LSTM Actor-Critic; DRL with OHLC + auto-encoder features; qualitative cost mention only; no slippage stated; short OOS windows (7 months each) |
| Pillai, Ajith & Sumesh K J (2026). arXiv:2601.19504. |
S&P 500 (100 stocks) |
$100,000 |
$235,492.83 |
53.46% CAGR |
15.60% |
1.68 |
Jan 2023 – Jan 2025 (2 years OOS) |
Not given |
61.5% |
Hybrid XGBoost + FinBERT sentiment + EMA/MACD/RSI regime filter; Backtrader simulation; 70/30 train/test split; no transaction costs stated; long-only daily |
| Nyo et al. (2026). arXiv:2603.15848. |
S&P 500 (equities) |
$100,000 |
Not given |
Not given (189.10% total return on validation 2018–2024) |
23.50% (validation) |
2.04 (validation); 2.07 (test) |
Dev: 2000–2017; Validation: 2018–2024; held-out test set |
Not given |
38.50% (validation) |
Enhanced momentum + FinBERT sentiment; EMA50/200, ATR14 trailing stop, top-10 cross-sectional momentum; strict 3-way train/val/test split; no transaction costs stated; student project paper – treat with caution |
| Saly-Kaufmann et. al. (2026). arXiv:2603.01820 |
Multi-asset futures & FX (bonds, commodities, energy, equity indices, FX – 60+ instruments) |
Not given |
Not given |
26.32% CAGR (VLSTM, best model) |
22.90% (VLSTM) |
2.40 (VLSTM, 2010–2025) |
2010–2025 (15 years, rolling OOS) |
Not given (turnover: ~967 ann.) |
58.8% hit rate (VLSTM) |
Large-scale benchmark of 16 DL architectures; Sharpe-ratio optimisation objective; rolling OOS; gross returns with breakeven cost analysis per asset; statistically significant vs. passive (HAC t=8.81); Oxford-Man Institute |
| Azevedo et. al. (2024). SSRN:4702406 |
US equities (stock anomaly long-short) |
Not given |
Not given |
Not given |
Not given |
0.84 net (LSTM, 1 hidden layer) |
Not given (post-2000 implied) |
Not given |
Not given |
LSTM anomaly-based long-short; net Sharpe 0.84 after all frictions; 57% cumulative performance reduction from full cost modelling; transaction costs, post-publication decay, post-decimalization all modelled |
| Buchanan & Benhamou (2026). arXiv:2603.14453 |
Top 30 S&P 500 stocks |
Not given |
Not given |
Not given |
Not given |
1.10 (OOS 2019–2025; baseline 0.85) |
In-sample: 2005–2018; OOS: 2019–2025 |
Not given |
Not given |
E-TRENDS LSTM with Sharpe-ratio training loss; 70/15/15 train/val/test split; 2–5 bps round-trip transaction costs modelled; OOS Sharpe +0.25 vs. baseline; no total return or MDD stated |
| Cohen, Aiche & Eichel (2025). DOI: 10.3390/e27060550 |
NASDAQ-100 stocks |
Not given |
Not given |
24.99% avg annual return (best: technical-quarterly framework) |
Not given |
1.2967 (technical-quarterly, best model) |
Jan 2020 – Jan 2025 (5 years, rolling-window OOS) |
Not given |
Not given |
ChatGPT-4o semantic intelligence + ML (technical/fundamental/entropy frameworks); top-10 equal-weighted portfolio; cumulative return 573.37% (best); rolling-window OOS retraining; no transaction costs or MDD stated; peer-reviewed MDPI Entropy |
| Ghatak et al. (2025). arXiv:2509.16707 |
814 US equities (walk-forward production system) |
Not given |
Not given |
26.38% cumulative (Jun 2021 – Jun 2025) |
3.04% |
2.54 |
Jun 2021 – Jun 2025 (4-year production walk-forward) |
8,859 |
56.6% |
Deep learning (feed-forward + recurrent), six-quarter rolling calibration; production system (not pure academic backtest); trading commissions/slippage not stated; very low MDD 3.04% notable; Zanista AI industry paper |
| Holzer et al. (2024). arXiv:2501.10709 |
Stock task: DJIA 30 stocks; Crypto task: BTC LOB data |
Not given |
Not given |
63.37% cumulative (stock task, PPO best, Jan 2021 – Dec 2023) |
Not given |
1.55 (stock task, PPO); 0.28 (crypto task, ensemble) |
Stock: Jan 2021 – Dec 2023; Crypto: Apr 7–19 2021 (very short LOB window) |
Not given |
Not given (win/loss ratio 1.62 crypto task) |
PPO, SAC, DDPG ensemble; rolling 30-day training / 5-day test windows; Sortino (stock) 2.44; costs mentioned but no numeric value; crypto window 13 days only; ACM ICAIF competition paper |
| Hajdini, Nuhiu & Leka (2025). PeerJ Computer Science. DOI: 10.7717/peerj-cs.3630 |
US equities: AAPL, MSFT, AMZN, BAC, NVDA |
$1,000,000 per security |
Not given |
53.87% avg annualised return (LLM-MAS-DRL framework) |
12.54% avg |
1.702 avg (LLM-MAS-DRL) |
Jul 2024 – Jun 2025 (OOS, 12 months) |
25 avg per security |
71.30% avg |
Three-layer LLM multi-agent + DRL framework (market analysis, risk management, execution); 0.1% transaction cost per trade modelled; Sortino, Calmar, CVaR also reported; comparison vs. Buy&Hold and PPO/A3C baselines; peer-reviewed PeerJ Computer Science |
| Fan et al. (2025). DOI: 10.1145/3766918.3766922 |
Ethereum (ETH) – multi-factor quantitative model |
Not given |
Not given |
97% annualised return (main result, threshold ±1.0) |
22% |
2.5 (threshold ±1.0); 2.2 at ±0.5; 2.3 at ±1.5 |
Q4 2024 (Oct–Jan 2025) bull test; Q1 2025 (Jan–Apr 2025) bear test; sensitivity range Oct 2024–Apr 2025 |
Not given (1.7 trades/week at ±1.0) |
59% |
ML-driven multi-factor model (RSI, MACD + on-chain gas/active address metrics); Information Ratio 1.2; avg holding period 4.5 days; IC mean 0.12; no transaction costs stated; note: 6-month backtest window is short; bull and bear sub-period tests reported separately |
| Huang et al. (2025). arXiv:2502.17493 |
US stocks (daily rebalancing, public stock data) |
Not given |
Not given |
61.73% p.a. (2019–2024 test period) |
Not given |
1.18 (2019–2024) |
Test 1: 2019–2024 (1,340 days); Test 2: 2005–2010 (1,360 days) |
Not given (daily rebalancing) |
Not given |
Return-weighted loss function for deep learning stock selection; 37.61% p.a. on 2005–2010 OOS test (Sharpe 0.97); two disjoint test windows including bear-market 2005–2010; no transaction costs or MDD stated; arXiv preprint 2025 |
| Nguyen (2026). arXiv:2602.11708 |
Crypto perpetual swaps (150+ pairs, Binance Futures; top 20 by market cap) |
Not given |
Not given |
40.5% annualised (70/30 long-short) |
12.70% |
2.41 |
Jan 2022 – Dec 2024 (OOS, 36 months; in-sample Jan–Dec 2021) |
~142 trades/month portfolio-wide |
54.2% (bull regime); overall not stated |
AdaptiveTrend: 6-hour momentum + dynamic trailing stop + Sharpe-based asset selection; 4 bps taker fee + slippage + funding modelled; Sharpe retained >2.0 at 8 bps; strict OOS separation; bootstrap significance tests |
| Zarattini, Pagani & Barbon (2025). SSRN:5209907 |
Crypto (Bitcoin + top-20 liquid altcoin rotation) |
Not given |
Not given |
+10.8% annualised alpha vs. BTC |
Not given |
>1.5 |
2015 onwards (exact end date not stated) |
Not given |
Not given |
Ensemble Donchian channel trend models (multi-lookback) + volatility position sizing; rotational portfolio; net-of-fees returns; transaction cost impact assessed; limited metric disclosure – backtest end date not stated |
| Jay & Berlanga (2024). SSRN:4987237 |
Cryptocurrency (BTC / crypto market) |
Not given |
Not given |
Not given (annualised return reported per model) |
Reported (model-specific) |
Reported per model (DQN best profit; LSTM best consistency; RF best drawdown control) |
Not explicitly stated |
Not given |
Not given |
Benchmarks DQN, LSTM, RF agents on crypto TA indicators; full suite: Total Return, Ann. Return, Volatility, Sharpe, Sortino, MDD, Calmar; no transaction costs stated; no single extractable top-line; exact period not published |
| Sattarov & Choi (2024). DOI: 10.1038/s41598-024-51408-w |
Bitcoin (BTC-USD) |
Not given |
Not given |
29.93% annualised |
Not given |
2.74 |
Training: Oct 2014 – Oct 2018; Test: last 30 days (~720 hrs) within Oct 2018 – Mar 2019 |
Not given |
Not given |
M-DQN (Multi-level Deep Q-Network) + Twitter sentiment; 3-module architecture; 1.5% round-trip transaction fee explicitly modelled; caution: test window only 30 days; Nature Scientific Reports (peer-reviewed) |
| Nguyen et al. (2025). DOI: 10.1080/23322039.2025.2594873 |
Bitcoin (BTC-USD) |
$1,000,000 |
$125,903,751.60 |
Not given (total return: +12,490.38% over Jan 2022 – May 2025) |
Not given |
Not given |
Train: Jan 2012 – Dec 2021; Test (OOS): Jan 2022 – May 2025 |
1,240 |
Not given |
DQN meta-strategy selector (chooses among RSI, SMA Crossover, Bollinger Bands, Momentum-20d, VWAP Reversion); ⚠ No Sharpe or MDD stated; no transaction costs stated; 12,490% return is unvalidated by risk metrics – credibility caution; peer-reviewed journal |
| Ni, Zhang & Fu (2025). arXiv:2412.18202. |
Cryptocurrency (BTC, ETH and major crypto assets) |
Not given |
Not given |
Not given (outperforms buy-and-hold benchmark) |
Not given |
Not given |
Not stated (dataset ends 2024) |
Not given |
Not given |
Denoising autoencoder + CNN + GAN pipeline for crypto trading signal generation; outperforms buy-and-hold; Sharpe/return/drawdown numerics not stated in abstract – insufficient for main leaderboard row; listed for transparency |
| Al-Waked & Al-Zoubi (2025). DOI: 10.14569/IJACSA.2025.0161181 |
Gold (XAU/USD) |
Not given |
Not given (cumulative: 80.21%) |
27.10% CAGR (PPO + Kalman filtering) |
0.48% |
12.10 (PPO + Kalman) ⚠ |
Jan 2017 – Jan 2025 (8 years, hourly data; N=47,304) |
Not given |
Not given |
PPO + Kalman filter for noise-resilient DRL trading; Kalman reduces microstructure noise; raw PPO baseline: Sharpe 0.45, CAGR 3.46%; ⚠ Sharpe 12.10 is unusually high – likely driven by very low realised volatility in filtered series and no transaction costs stated; no train/test OOS split stated; IJACSA peer-reviewed |
| Fatouros et al. (2024). arXiv:2412.19245 |
US common stocks (news-based long-short portfolio) |
Not given |
Not given |
Not given |
Not given |
3.05 (OPT long-short); 2.11 (BERT); 2.07 (FinBERT) |
Aug 2021 – Jul 2023 (24 months OOS) |
Not given |
Not given |
Sentiment trading with LLMs (OPT, BERT, FinBERT, Loughran-McDonald); daily long-short strategy on next-day stock returns using news sentiment; Loughran-McDonald baseline Sharpe 1.23; no transaction costs or MDD stated; Sharpe only metric extractable; arXiv preprint 2024 |