Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2611.09286

The End-of-Day Reversal After 2019: A Pre-Registered Test on S&P 500 Stocks Excluding Heavily Traded Names

Alex Vidovich·Zivana Zerjal
View original source ↗Read source document ↗Export metadata ↓

Abstract

Baltussen, Da and Soebhag (2025) document an end-of-day reversal in U.S. stocks through December 2019: intraday losers outperform intraday winners over the last half hour of trading. We test it on later data, on point-in-time S&P 500 members excluding about 80 heavily traded stocks we had examined in earlier, unpublished work. Under a private pre-registration, the 2016-2019 data had to reproduce the published sign before any statistic linking the sort to the outcome was computed on later primary-sample data. Measured to the last trade before the closing auction, the dollar-volume-weighted reversal is 3.13 bps per day (t = 7.5) in 2016-2019 and 3.05 bps (t = 3.0) in 2023-2026, with a one-sided 95% bound on the decline of 61%. The result depends on the window: the reversal is indistinguishable from zero in 2020-2022, 2023 carries much of the later estimate, and from 2020 onward it is 1.69 bps (t = 2.3), with a bound of 91%. A proxy sample, added after the run, gives 2.81 and 3.86 bps. On the full index the reversal falls from 3.40 to 0.74 bps (t of the change -1.83, short of the registered bar of -3); descriptive comparisons concentrate this weakness in the very largest stocks. No pre-registered mechanism test passes.

Explore related subjects

q-fin.STq-fin.TR

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Alex Vidovich, Zivana Zerjal. 2026-11-06. The End-of-Day Reversal After 2019: A Pre-Registered Test on S&P 500 Stocks Excluding Heavily Traded Names. https://arxiv.org/abs/2611.09286

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Who Are a Firm's Economic Peers? Mapping a Latent Information Field with Language Models

Identifying a firm's economic peers matters for benchmarking, risk measurement, and understanding return dependence, yet a diversified firm may have no single natural peer. We use language-model representations of corporate news to identify the combination of firms that jointly represents each firm's information environment. Each firm is represented by a distribution of article embeddings, and target-anchored Wasserstein barycentric reconstruction produces a directed map of economic-peer weights. We use this map as the interaction structure in a spatial factor model and ask whether return dependence follows peer relations measured from news without using returns. We construct peer weights from 2018-2022 news for 100 large-cap US traded firms and evaluate their financial relevance in subsequent returns. Relative to a matched model with the Fama-French five factors and momentum, the information field firms' unexpected returns move together after common factor exposures are accounted for. Joint reconstruction contributes substantially more than pairwise Wasserstein proximity or equal weighting of the selected peers, while centroid reconstruction retains most of the gain. Explicit news co-mentions provide complementary information. Language models thus make the interaction structure in spatial asset pricing an object of measurement: economically interpretable peer relations can be constructed from text before their financial relevance is assessed in returns.

q-fin.ST↗
arXiv2026

Portfolio Risk Bounds without Cross-Asset Return Covariances: Distributional Fields from Language-Model Representations

Portfolio risk assessment ordinarily relies on reliable estimates of cross-asset return covariances, which are difficult to obtain in short, high-dimensional panels. We show that firm-level distribution-valued characteristics can instead provide one-sided certificates of portfolio risk. Under maintained links from characteristics to systematic exposures and from exposures to returns, multi-firm Wasserstein-2 dispersion yields a sharp upper bound on systematic portfolio variance and a corresponding bound for standardized returns. A weighted pairwise relaxation produces an objective that is convex under a checkable condition and requires marginal volatility scales but no cross-asset return covariances. With zero firm-specific slack, the common-map scale changes the certified variance reduction but not the normalized allocation, which depends only on observed information geometry. In a 100-firm panel from 2018-2022, only 0.09%-0.46% of portfolios across four prespecified capped reference populations have in-sample variance no greater than an allocation constructed from Qwen3-Embedding-8B news representations; the corresponding share for equal risk weighting is 28.3%-34.7%. The lower in-sample variance ranking relative to equal risk also appears across the reported frozen language-model representations. The framework therefore converts distribution-valued firm information into a coherent risk bound and an implementable allocation rule constructed without cross-asset return covariances.

q-fin.ST↗
arXiv2026

Neural Realized Measures: Forecast-Supervised Intraday Aggregation for Volatility Forecasting

High-frequency volatility forecasting typically separates measurement from prediction: intraday observations are first compressed into a daily realized measure using a fixed aggregation rule, and the resulting statistic is then passed to a forecasting model. This conventional pipeline therefore treats measure construction as forecast-agnostic preprocessing. We ask whether the aggregation stage itself can instead be learned from the downstream forecasting objective while preserving the statistical structure of established realized measures. We propose Neural Realized Measures (NRM), a primitive-preserving, forecast-supervised aggregation framework that retains the original measure-specific intraday contributions but replaces uniform aggregation with positive, mean-normalized context-dependent weights. Uniform weights exactly recover the corresponding raw realized measure (RRM), so the classical construction is nested as a special case. Because NRM remains an explicit weighted sum of established primitive contributions, it also retains direct bar-level attribution. Across SF, HAR, and MIDAS forecasting specifications, NRM is evaluated against corresponding RRM specifications under controlled comparisons. In Heston simulations, NRM reduces RMSE for both contemporaneous and next-day latent integrated variance across all evaluated primitive--regime combinations. On six cryptocurrencies using five-minute data, NRM improves QLIKE in all 54 paired out-of-sample comparisons, with mean gains of 7.23\%, 8.15\%, and 9.22\% under SF, HAR, and MIDAS, respectively. Validation-selected NRM specifications also achieve the lowest mean QLIKE across the six assets among the evaluated approaches. Ablation and SHAP analyses reveal systematic time-of-day and nonlinear local-context effects in the learned allocation.

q-fin.ST↗