Paper: arXiv 2512.12727

Authors: Dinggao Liu, Robert Ślepaczuk, Zhenpeng Tang

Abstract

Accurately forecasting daily exchange rate returns represents a longstanding challenge in international finance, as the exchange rate returns are driven by a multitude of correlated market factors and exhibit high-frequency fluctuations. This paper proposes EXFormer, a novel Transformer-based architecture specifically designed for forecasting the daily exchange rate returns. We introduce a multi-scale trend-aware self-attention mechanism that employs parallel convolutional branches with differing receptive fields to align observations on the basis of local slopes, preserving long-range dependencies while remaining sensitive to regime shifts. A dynamic variable selector assigns time-varying importance weights to 28 exogenous covariates related to exchange rate returns, providing pre-hoc interpretability. An embedded squeeze-and-excitation block recalibrates channel responses to emphasize informative features and depress noise in the forecasting. Using the daily data for EUR/USD, USD/JPY, and GBP/USD, we conduct out-of-sample evaluations across five different sliding windows. EXFormer consistently outperforms the random walk and other baselines, improving directional accuracy by a statistically significant margin of up to 8.5–22.8%. In nearly one year of trading backtests, the model converts these gains into cumulative returns of 18%, 25%, and 18% for the three pairs, with Sharpe ratios exceeding 1.8. When conservative transaction costs and slippage are accounted for, EXFormer retains cumulative returns of 7%, 19%, and 9%, while other baselines achieve negative. The robustness checks further confirm the model’s superiority under high-volatility and bear-market regimes. EXFormer furnishes both economically valuable forecasts and transparent, time-varying insights into the drivers of exchange rate dynamics for international investors, corporations, and central bank practitioners.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced deep learning concepts like Transformers, multi-scale trend-aware attention, and squeeze-and-excitation blocks, requiring dense mathematical formulation, while providing rigorous empirical validation with multiple currency pairs, sliding-window backtests, transaction cost adjustments, and statistical significance testing.

Research Flowchart

  flowchart TD
  Start["Research Goal<br>Daily FX Returns Forecasting<br>EUR/USD, USD/JPY, GBP/USD"] --> Inputs["Data & Inputs<br>28 Exogenous Covariates<br>Daily Historical Prices"]
  Inputs --> Method["Core Methodology<br>EXFormer Architecture"]
  Method --> A["Multi-Scale Trend-Aware<br>Self-Attention (Parallel Convs)"]
  Method --> B["Dynamic Variable Selector<br>Time-Varying Weights"]
  Method --> C["Squeeze-and-Excitation Block<br>Feature Recalibration"]
  A & B & C --> Process["Learning & Out-of-Sample Forecasting<br>Sliding Windows & Backtesting"]
  Process --> Results["Key Findings & Outcomes"]
  Results --> R1["Outperforms Baselines & Random Walk<br>Dir. Acc. Gain: 8.5–22.8%"]
  Results --> R2["Backtest Returns (No Costs):<br>18%, 25%, 18%"]
  Results --> R3["Returns w/ Costs (>0):<br>7%, 19%, 9% vs Baseline Negative"]
  Results --> R4["Robust in High Volatility<br>Transparent Driver Insights"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes EXFormer, a novel Transformer-based architecture for forecasting daily exchange rate returns.

Source sentence
EXFormer, a novel Transformer-based architecture specifically designed for forecasting the daily exchange rate returns.

Assumes 5

  • Daily exchange rate returns are stationary. statistical
    Source sentence
    are stationary by construction and confirmed by the ADF tests above
  • No strong cointegrating relation among the three spot exchange rates. statistical
    Source sentence
    The results indicate that no strong cointegrating relation is present among the three series.
  • Transaction costs are a fixed 5 bps plus 2 bps slippage per trade. execution
    Source sentence
    We model frictions as a fixed per-trade deduction equal to the sum of a 5 bps transaction cost and a 2 bps slippage cost
  • A 15‑day sliding window provides the best trade‑off between adaptability and noise resilience. model
    Source sentence
    Only the 15-day configuration yields both immediate and sustained trading profits, balancing adaptability to fresh information with resilience against noise.
Show 1 more
  • Heteroskedasticity and autocorrelation consistent (HAC) standard errors are appropriate for forecast evaluation. statistical
    Source sentence
    using heteroskedasticity and autocorrelation consistent (HAC)

Data

Backtest sample: EUR/USD, USD/JPY, GBP/USD · fx · daily

  • Daily EUR/USD, USD/JPY, GBP/USD exchange rate returns market-prices unknown
    Three G10 currency pairs
    Source sentence
    Using the daily data for EUR/USD, USD/JPY, and GBP/USD, we conduct out-of-sample evaluations across five different sliding windows.

Recognised sources: Wind

Costs

Modeled Costs are part of the reported headline results.

Fixed per‑trade deduction of 5 bps transaction cost plus 2 bps slippage, applied whenever the position changes sign or moves between {-1,0,1}.

Source sentence
We model frictions as a fixed per-trade deduction equal to the sum of a 5 bps transaction cost and a 2 bps slippage cost, applied whenever the position changes sign or moves between {-1,0,1}.

Finds 5

Reported performance (EXFormer): Sharpe 2.09 (benchmark -0.05) · total return 25.8% · max drawdown 7.1%

  • EXFormer improves directional accuracy over the random walk by 8.5–22.8% in out‑of‑sample tests. predictive-accuracy
    Source sentence
    EXFormer consistently outperforms the random walk and other baselines, improving directional accuracy by a statistically significant margin of up to 8.5–22.8%.
  • Cumulative returns of 18%, 25% and 18% for EUR/USD, USD/JPY and GBP/USD respectively in near‑one‑year backtests, with Sharpe ratios exceeding 1.8. performance
    Source sentence
    In nearly one year of trading backtests, the model converts these gains into cumulative returns of 18%, 25%, and 18% for the three pairs, with Sharpe ratios exceeding 1.8.
  • After accounting for conservative transaction costs and slippage, EXFormer still yields positive cumulative returns of 7%, 19% and 9%, while baselines turn negative. performance
    Source sentence
    When conservative transaction costs and slippage are accounted for, EXFormer retains cumulative returns of 7%, 19%, and 9%, while other baselines achieve negative.
  • Statistical superiority confirmed by Clark–West test and dynamic regression tests for both MSFE and directional accuracy. statistical
    Source sentence
    and directional accuracy, with superiority confirmed by the Clark–West test and dynamic regression tests.
  • The dynamic variable selector provides time‑varying importance weights for 28 exogenous covariates, offering pre‑hoc interpretability. other
    Source sentence
    A dynamic variable selector assigns time‑varying importance weights to 28 exogenous covariates related to exchange rate returns, providing pre‑hoc interpretability.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 44 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (83)

  1. Exchange rate forecasting, order flow and macroeconomic information (2010) — link
  2. Exchange rate forecasting: the errors we’ve really made (2003) — link
  3. Exchange rate predictability (2013) — link
  4. Deep learning-based exchange rate prediction during the COVID-19 pandemic (2025) — link
  5. 30 years of exchange rate analysis and forecasting: A bibliometric review (2024) — link
  6. Empirical exchange rate models of the seventies: Do they fit out of sample? (1983) — link
  7. Deep learning with long short-term memory networks for financial market predictions (2018) — link
  8. Empirical asset pricing via machine learning (2020) — link
and 75 more
  1. Approximately normal tests for equal predictive accuracy in nested models (2007) — link
  2. Testing the value of directional forecasts in the presence of serial correlation (2014) — link
  3. Exchange rates and monetary policy uncertainty (2017) — link
  4. Structural breaks and GARCH models of exchange rate volatility: Re-examination and extension (2024) — link
  5. Unveiling the potential: Exploring the predictability of complex exchange rate trends (2024) — link
  6. A fundamental connection: Exchange rates and macroeconomic expectations (2024) — link
  7. Forecasting USD/RMB exchange rate using the ICEEMDAN-CNN-LSTM model (2025) — link
  8. Forecasting the daily exchange rate of the UK pound sterling against the US dollar (2025) — link
  9. Understanding exchange rate shocks during COVID-19 (2022) — link
  10. Imperfect exchange rate expectations (2025) — link
  11. Exchange rate return predictability in times of geopolitical risk (2022) — link
  12. A bi-level ensemble learning approach to complex time series forecasting: taking exchange rates as an example (2023) — link
  13. Forecasting daily and monthly exchange rates with machine learning techniques (2015) — link
  14. A new ensemble deep learning approach for exchange rates forecasting and trading (2020) — link
  15. Explaining exchange rate forecasts with macroeconomic fundamentals using interpretive machine learning — → 2303.16149
  16. Combined ensemble multi-class SVM and fuzzy NSGA-II for trend forecasting and trading in Forex markets (2021) — link
  17. Long-Term Exchange Rate Probability Density Forecasting Using Gaussian Kernel and Quantile Random Forest (2020) — link
  18. A deep coupled LSTM approach for USD/CNY exchange rate forecasting (2020) — link
  19. Forecasting Currency Exchange Rate through a Hybrid Time Series Analysis (2025) — link
  20. Neural networks performance in exchange rate prediction (2016) — link
  21. Forex market forecasting using machine learning: Systematic Literature Review and meta-analysis (2023) — link
  22. Forecasting of exchange rate time series based on event-aware transformer mode (2025) — link
  23. Fx-spot predictions with state-of-the-art transformer and time embeddings (2024) — link
  24. Improving exchange rate forecasting via a new deep multimodal fusion model (2022) — link
  25. Forecasting gold price with the XGBoost algorithm and SHAP interaction values (2024) — link
  26. LEST: Large language models and spatio-temporal data analysis for enhanced Sino-US exchange rate forecasting (2024) — link
  27. Narrative asset pricing: Interpretable systematic risk factors from news text (2023) — link
  28. From Econometrics to Machine Learning: Transforming Empirical Asset Pricing (2025) — link
  29. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021) — link
  30. Timexer: Empowering transformers for time series forecasting with exogenous variables (2024) — link
  31. Volatility forecasting and assessing risk of financial markets using multi-transformer neural network based architecture (2024) — link
  32. A combined Independent Component Analysis–Neural Network model for forecasting exchange rate variation (2019) — link
  33. A new multiscale decomposition ensemble approach for forecasting exchange rates (2019) — link
  34. A clustering-based nonlinear ensemble approach for exchange rates forecasting (2018) — link
  35. On the sources of uncertainty in exchange rate predictability (2018) — link
  36. Return and volatility linkages among International crude oil price, gold price, exchange rate and stock markets: Evidence from Mexico (2019) — link
  37. Fintech advancements for financial resilience: Analysing exchange rates and digital currencies during oil and financial risk (2024) — link
  38. Shorting the dollar when global stock markets roar: The equity hedging channel of exchange rate determination (2024) — link
  39. Nominal exchange rates and monetary fundamentals: evidence from a small post-Bretton Woods panel (2001) — link
  40. A Twofold Model for Exchange Rate Forecasting: Combining Fundamentals and Market Dynamics (2024) — link
  41. Forecasting directional movement of Forex data using LSTM with technical and macroeconomic indicators (2021) — link
  42. The uncovered interest parity puzzle, exchange rate forecasting, and Taylor rules (2019) — link
  43. Empirical exchange rate models of the nineties: Are any fit to survive? (2005) — link
  44. Exchange rate reconnect (2022) — link
  45. Exchange rate prediction redux: New models, new data, new currencies (2019) — link
  46. The shocks matter: improving our estimates of exchange rate pass-through (2018) — link
  47. Global portfolio rebalancing and exchange rates (2022) — link
  48. Attention is all you need (2017) — link
  49. Predicting the direction, maximum, minimum and closing prices of daily Bitcoin exchange rate using machine learning techniques (2019) — link
  50. A GARCH-temporal fusion transformer model for the volatility prediction of exchange traded funds (2025)
  51. Evolutionary support vector machine for RMB exchange rate forecasting (2019) — link
  52. Google Trends and the forecasting performance of exchange rate models (2018) — link
  53. Wavelet Transformer : An Effective Method on Multiple Periodic Decomposition for Time Series Forecasting (2025) — link
  54. A Kalman filter-based hybridization model of statistical and intelligent approaches for exchange rate forecasting (2021) — link
  55. A hybrid approach for studying the lead-lag relationships between China’s onshore and offshore exchange rates considering the impact of extreme events (2018) — link
  56. The relationship between oil prices and exchange rates: Revisiting theory and evidence (2020) — link
  57. A new LASSO-BiLSTM-based ensemble learning approach for exchange rate forecasting (2024) — link
  58. Interpretable EU ETS Phase 4 prices forecasting based on deep generative data augmentation approach (2024) — link
  59. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting (2019) — link
  60. Learning dynamics and heterogeneity of spatial-temporal graph data for traffic forecasting (2021) — link
  61. DeformTime : Capturing Variable Dependencies with Deformable Attention for Time Series Forecasting (2025) — link
  62. A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix (1987) — link
  63. Mean absolute directional loss as a new loss function for machine learning problems in algorithmic investment strategies — → 2309.10546
  64. Alternative Loss Function in Evaluation of Transformer Models — → 2507.16548
  65. Liquidity and exchange rates: An empirical investigation (2023) — link
  66. Recession prediction using yield curve and stock market liquidity deviation measures (2015) — link
  67. International liquidity and exchange rate dynamics (2015) — link
  68. Deviations from covered interest rate parity (2018) — link
  69. The share of systematic variation in bilateral exchange rates (2018) — link
  70. LSTM-ARIMA as a hybrid approach in algorithmic investment strategies — → 2406.18206
  71. Chronos: Learning the language of time series (2024) — link
  72. AirFormer: Predicting Nationwide Air Quality in China with Transformers (2023) — link
  73. DSTAGNN : Dynamic Spatial-Temporal Aware Graph Neural Network for Traffic Flow Forecasting (2022) — link
  74. Enhancing forex market forecasting with feature-augmented multivariate LSTM models using real-time data (2025) — link
  75. Forecasting the exchange rate with multiple linear regression and heavy ordered weighted average operators (2022) — link

4 of 83 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.