Predicting short-horizon equity returns remains one of the most contested problems in financial machine learning, where evaluation rigor is as consequential as model choice. Existing benchmarks frequently suffer from leakage introduced by global preprocessing and noncausal cross-validation, artificially inflating reported performance beyond what genuine predictive structure warrants. This study introduces a leakage-controlled, expanding-window walk-forward benchmark to compare next-day signed return and next-day volatility proxy forecasting across matched model families on US mega-cap equities. Using daily OHLCV data spanning 2010–2026, the benchmark spans 41 strictly causal folds with fold-isolated preprocessing. The benchmark assesses the following model groups under identical conditions: naive baselines, Ridge regression, Random Forest, XGBoost, LightGBM, LSTM, a causal temporal cross-sectional attention Transformer (CSTA), and a dynamic graph attention network (T-GAT). Results reveal a pronounced target asymmetry: signed-return forecasting remains statistically indistinguishable from naive baselines across all architectures, whereas the volatility proxy is markedly more predictable, with Random Forest achieving a 7.19% reduction in RMSE relative to rolling baselines. SHAP-based explainability analyses demonstrate that recoverable short-horizon signal concentrates in recent volatility, range, and liquidity features, favoring risk surveillance over directional alpha generation.