Multi-Task CNN–BiLSTM–Attention Networks for Multivariate FX Rate Forecasting: A Statistically Rigorous Evaluation Against Naive Persistence

Abstract

Deep learning models are widely proposed for foreign exchange (FX) forecasting, but many reported results derive directional accuracy indirectly from the sign of a regression output and are evaluated without correction for testing many currencies simultaneously, which inflates apparent predictability. This paper proposes a multi-task hybrid Convolutional Neural Network–Bidirectional Long Short-Term Memory–Attention (CNN–BiLSTM–Attention) architecture that jointly predicts next-step return magnitude (regression) and direction (direct binary classification) for 22 USD-denominated currency pairs, conditioned on three exogenous macro-financial series (gold, crude oil, and inflation) and causal, backward-only engineered momentum/volatility features. The model is evaluated with a strictly chronological train/validation/test split, scalers fit only on training data, a corrected one-sample Diebold–Mariano test against a naive-persistence baseline, and Benjamini–Hochberg false-discovery-rate correction across the 22 per-currency hypothesis tests. On a held-out test set, the model attains a mean directional accuracy of 50.93% (range 45.99%–57.57%) and level-space magnitude error statistically indistinguishable from naive persistence for every currency (22/22, Diebold–Mariano p > 0.05). Two currencies are nominally significant for direction before correction (MXN_USD, DKK_USD) but neither survives Benjamini–Hochberg correction (0/22 significant at FDR = 0.05). We report this as a rigorous negative result consistent with weak-form efficiency in liquid daily FX markets, and argue that the paper's principal contribution is a leakage-free, multi-task, multiple-comparison-corrected evaluation methodology rather than a claim of exploitable predictability.