Relevance
7/10
Highly relevant as a negative result paper that establishes empirical lower bounds on dataset requirements for ML-based intraday forecasting. Directly informs quantitative traders about the futility of applying sequential ML architectures to small single-instrument datasets. The walk-forward validation methodology and permutation testing framework are directly applicable to any quant research pipeline. The paper is part of a three-paper series systematically testing classical signals, behavioral classifiers, and ML approaches on MNQ. The finding that 944 days of 5-min data is insufficient for sequential ML is actionable guidance for strategy development.
Implementation Complexity
4/10
The implementations are relatively straightforward: HistGradientBoostingClassifier from scikit-learn and a minimal single-layer LSTM in Keras/JAX. Feature engineering involves standard rolling statistics, gap calculations, and decile tokenization. Walk-forward validation with expanding windows is standard practice. The main complexity lies in ensuring strict no-lookahead construction and the permutation testing framework. No custom architectures or novel algorithms are introduced.
Reproducibility
4/5
The paper provides detailed hyperparameters (max_leaf_nodes=15, min_samples_leaf=50, learning_rate=0.05, max_iter=200, l2_regularization=1.0 for GB; 16-unit LSTM with dropout 0.3, Adam lr=0.001, batch_size=32, max 50 epochs, patience=5). Data source (MNQ continuous front-month, 5-min RTH bars, Dec 2021–Sep 2025) is specified. Feature engineering pipeline is fully described. However, no code repository is linked, and the dataset requires purchase from a futures data provider. The AI disclosure statement notes AI tools were used for editorial and code debugging support.