EURUSD Forecasting — 30+ Models (Quantum · GNN · Diffusion · GA)
Most comprehensive EURUSD benchmark: 30+ models including Quantum ML (QSVM/QNN/QAE/VQC), Genetic Algorithms (7 variants + Neural Chromosomes), GNN, Neural SDE, Diffusion DDPM, Informer, PatchTST, TFT. Delta-target methodology. NSGA-2 multi-objective optimization.
4,211 EURUSD 1D candles (2010–2026), 39 features
Delta-target + 30+ models spanning classical TS, quantum ML, GNNs, diffusion, and 8 GA variants
The most comprehensive Forex forecasting benchmark, covering every modern ML paradigm from classical time series to quantum computing.
Delta-Target Methodology Predict Δclose (daily price change) not absolute price — eliminates non-stationarity and prevents trend leakage.
pred_close(t+1) = close(t) + pred_Δclose
Δclose ∈ [−0.05, +0.05] for EURUSD daily. Near-stationary. RobustScaler per-feature.
Dataset 4,211 EURUSD 1D candles (2010-01-01 → 2026-03-06), 39 engineered features. Top correlator: upper_shadow (+0.618 with Δclose).
30+ Models Across 8 Paradigms
1. Classical Statistical ARIMA(2,1,2), SARIMA(1,1,1)(1,1,1)5, GARCH(1,1)
2. Financial/Stochastic Black-Scholes GBM Monte Carlo, Heston Stochastic Volatility Model
3. Machine Learning ROCKET+Ridge, XGBoost+CatBoost (Optuna-tuned), Stacking (Inv-RMSE weighted)
4. Deep Sequence BiLSTM+Attention, BiGRU+Attention, CNN 1D-Conv, Informer, Autoformer, PatchTST, TFT (Temporal Fusion Transformer), TST Transformer
5. Graph Neural Networks GNN with Feature-Correlation Graph (edges = Pearson correlation between features)
6. Generative / Probabilistic Neural SDE (stochastic differential equation), Diffusion Model DDPM (denoising)
7. Quantum ML (4 architectures)
| Model | Approach |
|---|---|
| QNN VQC | Variational Quantum Circuit |
| QSVM | Quantum Kernel SVM |
| QBM / VQNN | Ising-VQC Boltzmann Machine |
| QAE | Quantum Amplitude Estimation |
8. Evolutionary / Genetic Algorithms (8 variants)
| GA Variant | Key Feature |
|---|---|
| Basic GA | Tournament + 1-point + uniform |
| Classic GA | Roulette + uniform + Gaussian |
| Advanced GA | BLX-α + Polynomial + Niching + Context |
| Lamarckian GA | GA + gradient fine-tune (Lamarckian inheritance) |
| Neural Chromo MLP | NN as chromosome (predict Δclose) |
| Neural Chromo Transformer | Transformer chromosome |
| Neural Chromo Context | preds + market features + attention overlay |
| NSGA-2 | Multi-objective (RMSE vs model complexity) Pareto front |
Grand Leaderboard (top performers) Stacking ensemble (Inv-RMSE weighted) consistently tops the leaderboard. Quantum models approach classical ML performance on this dataset size. Neural Chromosomes show promise as a genetic representation.
Honest Assessment EURUSD daily is near-efficient — even optimal models achieve directional accuracy of 52–56%. The value of this project is the framework, not the Sharpe ratio.