Date-Level Financial Temporal Anchoring and Anchor-Centric Priors for News-Enhanced Overnight Stock Return Prediction

Abstract

This research explores how news contributes to predicting stock returns overnight, specifically when the timestamps have limited precision. Data from public news datasets are typically only dated or normalized so they can produce a false signal when they have been returned to the market at an exact release time. To mitigate this problem, the conservative Date-Level Financial Temporal Anchors protocol is proposed to anchor news articles to tradable, pre-open prediction anchors using the U.S. market time zone, the trading calendar, and parsed timestamps, without the intent to recover precise release times to the minute. This anchored representation is then used to provide bounded article-level weights and statistics from event type, sentiment confidence and negative asymmetry, attention shock, and sentiment consensus or dispersion to develop an Anchor-Centric mechanism. Experiments were run with price data for nine U.S. stocks from 2018 to 2023 comparing to the FNSPID financial news dataset using anchored news features under similar target definition, splits, and configuration improved the majority of correlation, ranking and ranking direction, and portfolio-type diagnostic metrics. In the main LSTM case, Anchor-Centric improved Pearson correlation, Daily IC, Rank IC, AUC, and LS Top-3 Sharpe relative to the anchored-news version. Delay-control experiments found that artificial anchor delays and news signal decay were negatively correlated. The results build a strong case that temporal misalignment negatively impacts predictive signals based on financial news, and that conservative finite anchoring with prior distribution based on financial variables improves the alignment of financial news to overnight return labels given limited timestamp precision.