Paper: arXiv 2611.23126

Authors: Chariton Chalvatzis

Abstract

More historical data can improve estimation while introducing relationships that transfer poorly to current market conditions. We therefore examine how training quantity and historical selection interact with forecast horizon in intraday equity prediction. We deploy twelve configurations which combine four sampling and validation policies with 6, 12 and 24 month training quantities under a fixed cross-sectional Transformer and feature set. A monthly walk-forward experiment produces 264 refits and eight cumulative return forecasts from 5 to 120 minutes. Evaluation covers January 2025 through 20 October 2026 and uses cross-sectional Spearman information coefficients against future five-factor residual returns. Six-month recency sampling leads individual models at 5-15 minutes, while multi recency configurations lead at 20-30 minutes and uniform configurations at 60-120 minutes. Increasing training quantity does not uniformly improve accuracy while an equal-weight ensemble of all twelve models exceeds every individual mean at 15, 20 and 25 minutes, but not at the remaining horizons. These findings support evaluating historical selection jointly with forecast horizon without identifying a universal policy or a causal benefit from recency.

Complexity vs Empirical Score

  • Math Complexity: 5.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper demonstrates strong empirical rigor through extensive walk-forward testing and detailed evaluation metrics. Its novelty lies in systematically exploring the interaction of training history and forecast horizon in intraday equity prediction using Transformers. While the mathematical complexity of the Transformer itself is high, the paper focuses more on the empirical setup than deep theoretical derivations.

Research Flowchart

  flowchart TD
    A["Research Goal: Which Past Matters?"] --> B{"Methodology: Evaluate Training History & Multi-Forecast Horizon"};
    B --> C{"Inputs: Transformer Model, Fixed Feature Set, 4 Sampling Policies, 3 Training Quantities (6, 12, 24 months)"};
    C --> D["Computational Process: Monthly Walk-Forward Experiment (264 Refits), 8 Cumulative Return Forecasts (5-120 mins)"];
    D --> E{"Evaluation: Jan 2025 - Oct 2026, Cross-sectional Spearman Information Coefficients vs. 5-Factor Residual Returns"};
    E --> F["Findings: Recency sampling leads at short horizons (5-15 min), Multi-recency at mid (20-30 min), Uniform at long (60-120 min)"];
    F --> G["Outcome: No universal policy, joint evaluation of historical selection & forecast horizon supported, Ensemble outperforms individuals at some horizons"];

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a common‑model empirical framework that evaluates training quantity, historical‑selection policy and intraday forecast horizon jointly using a fixed cross‑sectional Transformer architecture and feature set.

Source sentence
Our contribution is a common-model empirical framework for evaluating historical-data selection jointly with intraday forecast horizon.

Assumes 5

  • The forecasting architecture, feature set and cross‑sectional Transformer model are held constant across all configurations. model
    Source sentence
    We investigate this question using a fixed cross-sectional Transformer and feature set. A fixed cross-sectional Transformer is refitted monthly under twelve historical-selection configurations.
  • Universe selection is limited to up to 1,000 common US equities with prior‑session price ≥ $5, ADV ≥ $10 M and valid volume in the preceding 90 sessions. data
    Source sentence
    At the first trading session of each month, select up to 1,000 common stocks ranked by trailing 90-trading-session average dollar volume (ADV). Require a prior-session unadjusted close of at least $5, ADV of at least $10 million and valid positive dollar volume in all 90 prior sessions.
  • Residual returns are computed using a five‑factor model and are not used as input features for prediction. data
    Source sentence
    Evaluation covers January 2025 through 20 October 2026 and uses cross-sectional Spearman information coefficients against future five-factor residual returns. Realized factor returns are part of the label construction and not input features for prediction.
  • Information coefficients (Spearman rank IC) are the primary measure of predictive accuracy. statistical
    Source sentence
    Evaluation covers January 2025 through 20 October 2026 and uses cross-sectional Spearman information coefficients against future five-factor residual returns. An equal-weight forecast ensemble supplies a common combination benchmark, while calibration, dependence and persistence diagnostics help interpret what the accuracy results imply.
Show 1 more
  • Training and validation periods are defined by contiguous calendar months and two‑month validation blocks as described in the policy definitions. data
    Source sentence
    The historical pool consists of 𝐻consecutive calendar months ending immediately before a two-month validation block. A policy selects 𝑇distinct months from this pool, and the available observations within those selected months form its selection-training sample while unselected months are excluded from that fit.

Data

  • Massive (formerly Polygon) US‑equity SIP one‑minute and daily aggregate flat files · Massive market-prices commercial
    US equities (up to 1,000 stocks per month)
    Source sentence
    Market data come from Massive (formerly Polygon), using its US-equity SIP one-minute and daily aggregate flat files.
  • Five‑factor model residual returns fundamentals unknown
    same stock universe as price data
    Source sentence
    Evaluation covers January 2025 through 20 October 2026 and uses cross-sectional Spearman information coefficients against future five-factor residual returns. Realized factor returns are part of the label construction and not input features for prediction.
  • Split and cash‑dividend reference data for price adjustment fundamentals unknown
    Source sentence
    The research pipeline constructs adjusted prices using split and cash-dividend reference data.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • An equal‑weight ensemble of all twelve models exceeds every individual model’s mean rank IC at 15, 20 and 25‑minute horizons but not at the other horizons. predictive-accuracy
    Source sentence
    Increasing training quantity does not uniformly improve accuracy while an equal-weight ensemble of all twelve models exceeds every individual mean at 15, 20 and 25 minutes, but not at the remaining horizons. 2 The ensemble exceeds every individual mean at 15, 20 and 25 minutes but not at the other horizons.
  • Increasing training quantity does not uniformly improve accuracy; the benefit of larger samples is conditional on the sampling family and horizon. predictive-accuracy
    Source sentence
    Increasing training quantity does not uniformly improve accuracy while an equal-weight ensemble of all twelve models exceeds every individual mean at 15, 20 and 25 minutes, but not at the remaining horizons. Shared architecture and inputs do not produce identical forecasts. not uniformly improve accuracy and the progression toward larger samples at longer horizons is conditional on the sampling family.
  • The leading individual model changes across horizons: R6 dominates 5–15 min, M12 at 20 min, M24 at 25–30 min, U24 at 60 min, and U12 at 120 min. performance
    Source sentence
    Across these configurations, the leading individual model changes from R6 at 5–15 minutes to M12 at 20 minutes, M24 at 25–30 minutes, U24 at 60 minutes and U12 at 120 minutes.
  • Six‑month recency sampling leads individual models at 5–15 min, multi‑recency configurations lead at 20–30 min, and uniform configurations lead at 60–120 min. predictive-accuracy
    Source sentence
    Six-month recency sampling leads individual models at 5-15 minutes, while multi recency configurations lead at 20-30 minutes and uniform configurations at 60-120 minutes.
  • Predictive rankings vary substantially across calendar periods, e.g., 5‑minute Rank IC falls from 3.72 % in H1 2025 to 0.28 % in H2 2025 and rises to 1.01 % in 2026. statistical
    Source sentence
    Performance also varies substantially across calendar periods: ensemble 5-minute Rank IC falls from 3.72% in H1 2025 to 0.28% in H2 before reaching 1.01% in 2026 through 20 October. Substantial calendar-period variation further shows why pooled rankings alone are insufficient for selecting a forecasting policy.

Checked by others

Qualified by

  • Global Neural Networks and The Data Scaling Effect in Financial Time Series Forecasting (2023)
    On the question: Does increasing the size and heterogeneity of training data improve forecasting accuracy?
    Compare the findings

    This paper: Increasing training quantity does not uniformly improve accuracy; the benefit of larger samples is conditional on the sampling family and horizon.

    That paper: Forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous.

    A finds accuracy improves markedly with larger, more heterogeneous data; B finds larger samples help only conditionally, depending on sampling family and forecast horizon.

Code

No repository by the authors found in the paper.

Extracted 2026-11-18 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (13)

  1. Selection of estimation window in the presence of breaks (2007) — link
  2. Optimal Forecasts in the Presence of Structural Breaks (2013) — link
  3. Rolling window selection for out-of-sample forecasting with time-varying parameters (2017) — link
  4. Forecast Combination Across Estimation Windows (2011) — link
  5. DDG-DA : Data Distribution Generation for Predictable Concept Drift Adaptation (2022) — link
  6. Scaling Laws for Neural Language Models (2020) — link
  7. DoubleAdapt : A Meta-learning Approach to Incremental Learning for Stock Trend Forecasting — → 2306.09862
  8. Filter, Augment, Forecast: Online Data Selection for Robust Time Series Forecasting (2026) — link
and 5 more
  1. Data Selection for Language Models via Importance Resampling (2023) — link
  2. DoReMi : Optimizing Data Mixtures Speeds Up Language Model Pretraining (2023) — link
  3. D4 : Improving LLM Pretraining via Document De-Duplication and Diversification (2023) — link
  4. Attention Is All You Need (2017) — link
  5. Language Models are Unsupervised Multitask Learners (2019) — link

1 of 13 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.