Skip to content

Sequential Structure in Intraday Futures Data: LSTM vs Gradient Boosting on MNQ

Home / Papers / Sequential Structure in Intraday Futures Data: LSTM vs Gradient Boosting on MNQ

Sequential Structure in Intraday Futures Data: LSTM vs Gradient Boosting on MNQ

Authors: Mathias Mesfin Date: 2026-05-18 Paper ID: openalex:2605.17724

Summary

This paper investigates the viability of using LSTM and gradient boosting models for intraday directional prediction of Micro E-Mini Nasdaq 100 (MNQ) futures. By training on four years of 5-minute OHLCV data using a strict expanding-window validation framework, the study evaluates whether sequential structure can be reliably learned from this scale of financial data. Results show that no configuration performs significantly better than the base rate, with permutation tests failing to reject the null hypothesis of no predictive edge. The study concludes that current sequential ML approaches, including those inspired by foundation models like Kronos, struggle to extract stable signals from isolated, medium-scale financial datasets.

Key Contributions

  • Evaluates LSTM and gradient boosting performance on Micro E-Mini Nasdaq 100 (MNQ) 5-minute OHLCV data across 944 trading days.
  • Demonstrates that none of the tested configurations achieve statistically significant predictive accuracy above the base rate using expanding-window walk-forward validation.
  • Establishes an empirical lower bound on data requirements for sequential machine learning in financial forecasting by showing four years of single-instrument data are insufficient for capturing exploitable intraday structure.

Metadata & Links

url
https://arxiv.org/abs/2605.17724
paper_id
2605.17724
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-21T06:28:08Z