WaveESN–RegimeMLP: GA-Tuned Reservoirs and Regime-Aware Multiscale Forecasting

Financial series change their behavior over time and contain noise at many scales, which weakens standard linear forecasts. We present WaveESN–RegimeMLP, a modular pipeline that combines (i) wavelet-based multiscale features, (ii) a reservoir network with an elastic-net readout, (iii) a hidden-state model that detects market regimes from residuals, and (iv) genetic search for hyperparameters. Evaluated on daily prices of AAPL, DIA, SPY, and JPM from 1 January 2021 to 1 January 2022, the approach reduces one-day-ahead RMSE by 10–37% and z-normalized RMSE by 10–32% relative to a standard reservoir baseline. Gains are largest during volatile periods with frequent regime switches. The design is transparent, computationally light, and readily extendable to more assets and horizons.