LSTM vs Gradient Boosting for MNQ Futures
A comparative study of deep learning and ensemble methods for intraday directional prediction of Nasdaq futures using OHLCV bars. The paper uses multi-year walk-forward validation to evaluate sequential predictive structures in high-liquid instruments.
Strategy Decoder Editorial · · 4 min read
Key takeaways
- Intraday prediction for MNQ futures using 5-minute OHLCV data was not statistically significant with either LSTMs or Gradient Boosting.
- Models achieved out-of-sample accuracies ranging from 50.00% to 50.89%, barely above the 51.8% baseline.
- Permutation tests confirmed no statistically significant predictive edge, with p-values far from common significance thresholds.
- The study suggests that four years of single-instrument 5-minute OHLCV data are insufficient for robust sequential ML-based intraday forecasting.
- Feature importance instability points to noise fitting rather than the capture of stable, structural market signals.
Algorithmic and quantitative traders constantly seek new methods to eke out an edge in financial markets. Machine learning, particularly deep learning and ensemble techniques, holds immense promise for identifying complex patterns in high-frequency data. A recent study by Mathias Mesfin investigated the efficacy of two prominent machine learning paradigms, Long Short-Term Memory (LSTM) networks and Gradient Boosting, for intraday directional prediction in Micro E-Mini Nasdaq 100 (MNQ) futures.
The Quest for Intraday Predictive Power
The research was motivated by advances in foundation models leveraging financial candlestick data, such as the Kronos architecture. The central question was whether discernible and exploitable sequential predictive structures exist within five-minute OHLCV (Open, High, Low, Close, Volume) bar sequences for a single instrument. Intraday futures markets, with their high liquidity and rapid price movements, represent a challenging yet potentially rewarding environment for such models.
Mesfin's study analyzed a substantial dataset comprising 944 trading days of MNQ futures movements from 2021 to 2025. The objective was to predict whether the session's closing price would exceed its 10:30 AM open by more than ten points, providing a clear binary classification target. This specific target reflects a common objective in short-term trading: identifying periods of sustained upside momentum.
Methodology: A Rigorous Approach
To ensure the robustness of the findings, the author employed a strict expanding-window walk-forward validation scheme. This method is critical for evaluating trading strategies as it mimics real-world conditions by preventing future data from influencing past predictions, thereby guarding against look-ahead bias. Four distinct model configurations were assessed across three out-of-sample periods, allowing for a comprehensive performance comparison.
Two primary classes of models were pitted against each other: Gradient Boosting variants and LSTMs. Gradient Boosting, an ensemble learning technique, builds strong predictors from a series of weaker ones. LSTMs, a type of recurrent neural network, are particularly well-suited for processing sequential data and capturing long-term dependencies, making them a natural choice for time-series forecasting. The expectation was that the sequential nature of intraday OHLCV bars might be effectively modeled by LSTMs.
Disappointing Results: No Edge Found
Contrary to expectations, the findings presented by Mathias Mesfin indicate a significant challenge in extracting statistically significant predictive signals from this specific dataset. None of the tested model configurations, whether Gradient Boosting or LSTM, achieved out-of-sample accuracy substantially above the established 51.8% base rate. The combined out-of-sample accuracies for Gradient Boosting variants ranged from 50.00% to 50.89%, while the LSTM model performed at 50.59%.
To rigorously test the significance of these results, permutation tests were conducted. These tests evaluate whether the observed predictive performance is genuinely due to a learned signal or merely random chance. The p-values obtained were 0.135 for the best Gradient Boosting model and a high 0.515 for the LSTM. These values are well above conventional statistical significance thresholds (e.g., 0.05), leading to the conclusion that neither model demonstrated a statistically significant predictive edge.
Intriguingly, the study also revealed instability in feature importance across different walk-forward folds. This suggests that the models were more likely fitting noise within the data rather than identifying stable, structural market signals that could be reliably generalized to new, unseen periods. This phenomenon is a classic sign of overfitting or a lack of persistent signal in the underlying data.
Data Sufficiency: A Critical Implication
A key contribution of Mesfin's work, as reported in the study, is an empirical lower bound on the data scale requirements for sequential financial machine learning. The results suggest that using four years of single-instrument five-minute OHLCV data was insufficient to uncover reliable sequential ML-based intraday forecasting signals. This implies that for similar prediction tasks on individual assets, researchers and practitioners might need to explore significantly larger datasets, higher-resolution data, or perhaps incorporate a broader array of features including market-wide information or alternative data points.
Why it matters for algo traders
For algorithmic and quantitative traders, Mesfin's study provides crucial insights. Firstly, it serves as a reality check: developing profitable intraday strategies using common machine learning techniques and standard OHLCV data for a single instrument like MNQ futures is extremely difficult, even with advanced models like LSTMs. The modest out-of-sample accuracies and lack of statistical significance highlight the efficiency of modern markets. Traders should be wary of backtests that do not employ strict walk-forward validation and robust statistical testing, as these can easily overstate true predictive power. The finding on feature importance instability underscores the risk of models simply memorizing historical noise rather than learning generalizable patterns. This research implies that successful intraday alpha generation may require more sophisticated data inputs, multi-asset approaches, or the capture of more transient, complex patterns than those discernible from basic OHLCV streams on a single future over this timeframe. It reinforces the notion that substantial data and careful validation are paramount in quant research, and that a 'bigger hammer' (more complex model) doesn't always guarantee better results without sufficient, clean data.
Frequently asked questions
What was the purpose of the study?
The study aimed to compare Gradient Boosting and LSTM models for predicting intraday directional movements in Micro E-Mini Nasdaq 100 (MNQ) futures using five-minute OHLCV bar sequences from 2021-2025.
Did the models find a statistically significant predictive edge?
No, neither the Gradient Boosting nor the LSTM models achieved statistically significant out-of-sample accuracy above the 51.8% baseline, as confirmed by permutation tests with high p-values.
What does the study suggest about data requirements for financial machine learning?
The research indicates that four years of single-instrument five-minute OHLCV data may be insufficient for reliable sequential machine learning-based intraday forecasting, implying a need for more extensive or diverse datasets.
Tags: deep learning, futures trading, time-series, backtesting
Based on reporting by arXiv q-fin.TR.