A Statistical-Finance Benchmark for Same-Day Directional Stock Prediction: Walk-Forward Evidence from SPY

Battle Count: 221

The paper provides a rigorous statistical benchmark for same-day directional prediction using simple OHLC features, which is directly relevant to quantitative trading signal generation. The threshold-conditioned accuracy analysis (72.7% at 1% predicted move) and regime-specific results are practically informative. However, the paper explicitly avoids trading-system claims, lacks execution modeling (slippage, market impact, transaction costs beyond the illustrative backtest), and the high-accuracy regions have very small sample sizes. The illustrative backtest (Table 7) shows 239% return vs 34.63% B&H at zero cost, but this is explicitly labeled as execution-illustrative only. The finding that Logistic Regression outperforms tree models for direction is practically useful. Overall, it is a solid statistical foundation but not a deployable trading strategy.

The methodology is straightforward: standard walk-forward validation with off-the-shelf ML models (XGBoost, Random Forest, LightGBM, Logistic Regression) on 5-6 features. No deep learning architectures, no complex feature engineering, no custom optimization. The main complexity lies in the careful experimental design (leakage controls, threshold-conditioned reporting, statistical tests) rather than in model implementation. A competent ML engineer could reproduce the core benchmark in a few days.

The paper provides an OSF archive (https://osf.io/6thqk) with experiment scripts for the main SPY benchmark and the auxiliary stock-screen summary. Walk-forward validation protocol is clearly specified with fixed model specifications. Data source (Yahoo Finance SPY OHLC) is publicly available. However, the auxiliary 541-equity screen details are summarized rather than fully reproduced, and the exact XGBoost configuration used in the stock pipeline is described but not exhaustively parameterized in the main text. The paper explicitly separates stock-only results from cryptocurrency pipelines in the wider repository.