Temporal Validation and Incremental News-Text Analysis of a Lightweight Cross-Sectional Equity Ranking Model

Model Construction, Matched Comparisons, and Cost Sensitivity for M8 N0

Abstract

Cross-sectional equity prediction requires explicit information-timing rules, consistent treatment of differences in samples across trading dates, and careful separation of model selection from final evaluation. This study examines whether structured market features and news-event metadata support same-day open-to-close return ranking when inputs are frozen before the market opens. It also assesses the incremental contribution of news text in a matched comparison. M8 N0 combines 23 numerical features, their missingness indicators, and two log-transformed event counts into a 48-dimensional input. A multilayer perceptron with 5,345 trainable parameters produces cross-sectional ranking scores. Training uses within-date rank-normalized targets, a joint SmoothL1 and pairwise ranking objective, and optimization batches organized around complete trading dates. Three expanding-window folds use preprocessing statistics fitted only on their respective training samples, temporal gaps, and controls for duplicate events across partitions.

Across three development-validation folds with seed 42, the model covers 298,769 symbol-day observations and 1,350 trading dates. Mean daily Spearman IC is 0.02088, with positive IC on 55.48% of dates. Fold-level mean IC values are 0.01633, 0.02948, and 0.01668. In the matched comparison on fold 0, the numerical baseline, additive text-residual model, and text branch with relation-weighted pooling achieve IC values of 0.01633, 0.01587, and 0.00396, respectively. The zero-text control matches the numerical baseline. These findings document an observable historical ranking relationship in the structured inputs within the completed configurations and sample scope, while the tested text addition does not improve mean IC on this fold. Post-hoc moving-block bootstrap analyses characterize uncertainty in the saved daily sequences under temporal dependence, and yearly summaries describe temporal variation. An exploratory portfolio analysis demonstrates that ranking metrics and returns after costs require separate interpretation: an equal-weight long-short portfolio selecting 20% of each tail has approximately -3.11% arithmetic annualized return under a simplified cost of 2 basis points per one-way execution. Reload and release-consistency checks support reproducible inference. Because validation data were used for checkpoint selection, the results are developmental rather than an untouched out-of-sample test. The study provides an auditable research workflow connecting information timing, a numerical baseline, incremental-text controls, and released artifacts, establishing an explicit starting point for independent temporal evaluation and multi-seed research.