Navier-Stokes-Inspired Turbulence Features for LSTM Forex Prediction: A Walk-Forward Ablation Study

We study whether fluid-dynamics-inspired features improve multi-class directional forecasting of foreign exchange (forex) rates when used as inputs to a compact LSTM. Building on the turbulence analogy of Ghashghaie et al. (1996) and the financial turbulence index of Kritzman and Li (2010), we engineer four Navier-Stokes-inspired features—turbulence index (Mahalanobis distance), Reynolds-number analog, vorticity proxy, and spectral energy slope—and evaluate them under expanding-window walk-forward validation on five liquid currency pairs (EUR_USD, USD_CAD, EUR_GBP, NZD_USD, AUD_USD; seven windows per pair). Alone, the four features **do not** improve a 26-feature technical-analysis baseline (mean Sharpe −0.217 vs −0.001; EXP-018 reject). Combined with four stationary-wavelet features, however, they produce a large, cross-pair lift: mean Sharpe **12.35**, mean balanced accuracy **63.7%**, zero class collapse (EXP-018 accept). Monte Carlo block bootstrap (10,000 resamples, mean block length 3, seed 42) places a 95% CI of **[10.29, 14.39]** on the portfolio Sharpe point estimate (EXP-021); the block-bootstrap lower bound exceeds a Sharpe = 2.0 reference by a factor of **5.14**. A second generation of four dynamics-oriented NS features (EXP-019) does not beat the first-generation NS+wavelet configuration. We interpret the result as **feature synergy**—turbulence state features become informative only when multi-scale wavelet context is present—not as a claim of live trading edge. Reported Sharpe ratios are **cost-free, H1-bar annualized** strategy returns from argmax class predictions and must not be read as live P&L.