💰 quantitative finance

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

This paper demonstrates that neither gradient boosting nor LSTM models can achieve statistically significant predictive accuracy over a 51.8% base rate when forecasting intraday MNQ futures direction using four years of five-minute OHLCV data, suggesting that such single-instrument datasets are insufficient for reliable sequential financial machine learning.

Original authors: Mathias Mesfin

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mathias Mesfin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for a specific city, but you only have four years of daily weather logs. You want to know: "If it rains in the first hour of the day, will it rain all day?"

This paper is a report from a researcher named Mathias Mesfin who tried to answer that question, but instead of weather, he looked at stock market futures (specifically the MNQ, which tracks the Nasdaq 100). He wanted to see if modern, high-tech computer brains (Machine Learning) could find hidden patterns in the market's "heartbeat" that humans miss.

Here is the story of what he did, how he did it, and what he found, explained simply.

The Big Idea: The "Kronos" vs. The "Local Shop"

Recently, a massive AI model called Kronos was released. It was trained on millions of stock charts from many different markets. It learned that the order in which price bars appear (the sequence) actually holds clues about what happens next. It's like a master chef who has tasted millions of dishes and knows exactly how a specific sequence of ingredients will taste.

Mathias asked a simple question: "If we take that same idea, but apply it to just one market (MNQ) over just four years, will it still work?"

He set up a race between two types of computer learners:

The "Scattergun" (Gradient Boosting): This model looks at a list of facts (like "yesterday was hot," "the gap was wide") but treats them as a pile of unrelated items. It doesn't care about the order they happened in.
The "Storyteller" (LSTM): This model is designed to read a story. It looks at the 12 five-minute bars of the morning and tries to understand the sequence of events, just like reading a sentence word-by-word.

The Experiment: A Strict Test

Mathias didn't just let the computers guess. He set up a very strict test called Walk-Forward Validation.

The Rule: The computer learns from 2022, then tries to predict 2023. Then it learns from 2022–2023 and predicts 2024. Finally, it learns from 2022–2024 and predicts 2025.
The Goal: Predict if the market will close higher than it opened at 10:30 AM by at least 10 points.
The Baseline: If you just flipped a coin, you'd be right about 51.8% of the time (because the market goes up slightly more often than down). To win, the computer had to beat this coin flip.

The Results: The Computers Got Lost

The results were honest and disappointing for the "high-tech" hope, but very informative for science:

The "Scattergun" (Gradient Boosting): It tried to find patterns in the daily data. Sometimes it got lucky (54% accuracy in one year), but in other years, it got it wrong more often than a coin flip. When Mathias checked why it was making decisions, the reasons changed every year. One year it cared about "overnight gaps," the next year it cared about "volume." It was like a student who memorized the answers to last year's test but failed to learn the actual math.
The "Storyteller" (LSTM): This was the big test. Did the sequence of the morning bars matter? The answer was no. The model essentially gave up. It predicted a 50/50 chance for every single day, regardless of what the morning looked like. Its "confidence" was flat. It learned nothing.

The Verdict: None of the models could predict the market better than a coin flip. The "Storyteller" didn't find a story; it just found noise.

The "Why": The Library Analogy

Why did the high-tech AI fail? The paper offers a great explanation using a Library Analogy:

The Kronos Model (The Winner): Imagine a library with 9 million books. If you want to learn a rare language, you can read millions of examples and find the rules.
Mathias's Dataset (The Loser): Imagine you only have 944 pages (one day per year for four years).
- If a specific pattern happens only 5% of the time, it appears only 47 times in your entire library.
- Trying to teach a complex AI to find a pattern that only appears 47 times is like trying to teach a child to recognize a specific type of cloud by showing them a picture of that cloud once every two weeks for four years. There simply isn't enough "data fuel" for the engine to run.

The AI didn't fail because it was "dumb." It failed because it was starving for data. It needed millions of examples to learn the "grammar" of the market, but it was only given a few hundred.

The Conclusion: An Honest "No"

The most important part of this paper isn't that the models failed; it's that the researcher admitted they failed and explained exactly why.

The Claim: You cannot use advanced AI to predict the daily direction of a single stock market using only four years of 5-minute data. The data is too small.
The Lesson: To make these models work, you need either:
1. Much more history (decades, not years).
2. Much faster data (tick-by-tick data, which multiplies the number of data points by 50x).
3. More markets (training on Nasdaq, S&P, and Dow Jones all at once to give the AI more examples).

In short: The "Kronos" super-AI works because it has read the entire encyclopedia of the market. Mathias tried to teach a similar AI using just a single page of a dictionary, and unsurprisingly, it couldn't learn the language. This paper is a valuable map showing researchers exactly where the "data cliff" is, so they don't waste time trying to jump over it with insufficient data.

Technical Summary: Sequential Structure in Intraday Futures Data

Problem Statement
This research addresses a critical gap in quantitative finance: determining whether sequential bar-level structure in intraday futures data contains exploitable predictive information when processed at a scale significantly smaller than that of recent financial foundation models. Motivated by the Kronos architecture (Shi et al., 2025), which demonstrated that transformer-based models can learn predictive patterns from millions of candlestick bars across multiple instruments, this study investigates whether similar sequential dependencies exist within a constrained, single-instrument dataset. Specifically, it asks if a four-year history of Micro E-Mini Nasdaq 100 (MNQ) five-minute bars contains sufficient structural signal for directional forecasting, or if the "sequential" advantage is purely a function of massive data volume.

Methodology
The study employs a rigorous walk-forward validation framework across three out-of-sample periods (2023, 2024, 2025) using 944 trading days of MNQ Regular Trading Hours (RTH) data (December 2021–September 2025). The target variable is binary: whether the session close exceeds the 10:30 AM bar open by more than 10 points (a threshold representing approximately one baseline ATR).

Four model configurations were evaluated:

Gradient Boosting (Daily): A HistGradientBoostingClassifier trained on 29 engineered daily features (returns, gaps, volatility, session structure) tokenized into decile bins to mimic foundation model quantization.
Gradient Boosting (Intraday): The same classifier trained on 30 intraday features derived from the first 60 minutes of the session, including a 12-bar return sequence treated as independent features.
Gradient Boosting (Volatility-Adjusted): An intraday model using an ATR-normalized target to isolate directional signal from volatility regimes.
LSTM (Sequential): A minimal Long Short-Term Memory network (single layer, 16 units, 0.3 dropout) designed to process the 12-bar return sequence as a genuine time series, preserving temporal ordering that gradient boosting discards.

All models utilized expanding training windows and were subjected to permutation testing (200 iterations) on the final fold (2025) to establish statistical significance against a null distribution of shuffled labels.

Key Results
The study found no evidence of statistically significant out-of-sample predictive power across any configuration:

Accuracy: Combined out-of-sample accuracies ranged from 50.00% (GB-Daily) to 50.89% (GB-VolAdj) for gradient boosting variants, and 50.59% for the LSTM. These figures are indistinguishable from the 51.80% base rate and a naive coin-flip baseline.
Statistical Significance: Permutation tests yielded p-values of 0.135 (best GB-Intraday), 0.390 (GB-VolAdj), and 0.515 (LSTM). All failed to meet the $p < 0.05$ threshold, indicating the results are consistent with random noise.
Feature Stability: Feature importance analysis revealed systematic instability across folds. Dominant features shifted entirely between training windows (e.g., from gap features to specific bar returns), confirming that models were fitting spurious patterns rather than capturing structural signals.
LSTM Behavior: The LSTM model converged to a flat, uncalibrated classifier, predicting the base rate probability (~0.50) for all inputs regardless of the sequential input. This suggests the model learned nothing rather than overfitting to noise.

Key Contributions
The primary contribution of this paper is an honest empirical lower bound on the dataset scale required for sequential machine learning in intraday forecasting. By implementing a tractable approximation of the Kronos conceptual framework on a real-world, constrained dataset, the study demonstrates that:

Scale Dependency: Sequential architectures (LSTMs) and foundation model-inspired tokenization strategies fail to extract predictive structure from single-instrument data at the scale of ~944 days.
Noise vs. Signal: The lack of consistent feature importance and the convergence to random prediction indicate that the "sequential structure" observed in large-scale foundation models is not present—or is not learnable—at this specific resolution and sample size.
Methodological Documentation: The paper provides a complete, reproducible pipeline for testing sequential hypotheses in finance, including strict walk-forward validation and permutation testing, serving as a counter-example to over-optimistic claims about small-data ML applications.

Significance and Claims
The paper modestly claims that its negative result is a necessary empirical correction to the field. It posits that the transition from "failure" (as seen here) to "success" (as seen in Kronos) likely requires a dataset size orders of magnitude larger—potentially between 500,000 and 5,000,000 bars. The study concludes that for researchers limited to standard retail-accessible historical data (single instrument, 5-minute bars), sequential ML architectures currently offer no advantage over random guessing. The paper suggests that meaningful progress in this domain requires either a shift to tick-level data, multi-instrument training, or transfer learning from pre-trained foundation models, rather than simply applying these architectures to small, isolated datasets.