Daily equity return forecasting exhibits low signal-to-noise ratios, heavy tails, volatility clustering, and frequent regime changes. We propose a compact, leak-safe forecasting pipeline that combines (i) guard-banded rolling MODWT/MRA feature construction (LA8, \(J=3\) ) computed strictly from past-only windows (rolling window \(W_{\textrm{mra}}=256\) , guard \(g_{\textrm{mra}}=49\) , periodic boundary), (ii) volatility-normalized targets with past-only scaling, and (iii) time-decayed residual stacking with validation-only simplex-constrained weights and a post-hoc affine calibration. On SPY daily log-returns over 2018–2024 (with auxiliary series QQQ, IWM, GLD, TLT), the reference ensemble (TD+Cal) attains a test \(\textrm{RMSE}=0.007979\) and \(\textrm{MAE}=0.005856\) . Its out-of-sample \(R^2_{\textrm{OS}}=0.481\) is defined relative to the NaiveLast benchmark \(n_t=y_{t-1}\) and is therefore baseline-dependent. Diebold–Mariano and Clark–West tests with Newey–West HAC (main lag=5, lag sensitivity 0/5/10, Holm-adjusted across contrasts) provide strong evidence against simple benchmarks (NaiveLast and a rolling ARIMA benchmark), while differences relative to strong regularized linear and tree-based learners are small in magnitude and not statistically distinguishable after multiple-testing adjustment. Residual diagnostics (Ljung–Box and ARCH–LM on test errors) are reported as sanity checks and do not suggest pronounced remaining dependence at standard lags.
