Understanding Model Complexity for temporal tabular and multi-variate time series, case study with Numerai data science tournament

In this paper, we explore the use of different feature engineering and dimensionality reduction methods in multi-variate time-series modelling. Using a feature-target cross correlation time series dataset created from Numerai tournament, we demonstrate under over-parameterised regime, both the performance and predictions from different feature engineering methods converge to the same equilibrium, which can be characterised by the reproducing kernel Hilbert space. We suggest a new Ensemble method, which combines different random non-linear transforms followed by ridge regression for modelling high dimensional time-series. Compared to some commonly used deep learning models for sequence modelling, such as LSTM and transformers, our method is more robust (lower model variance over different random seeds and less sensitive to the choice of architecture) and more efficient. An additional advantage of our method is model simplicity as there is no need to use sophisticated deep learning frameworks such as PyTorch. The learned feature rankings are then applied to the temporal tabular prediction problem in the Numerai tournament, and the predictive power of feature rankings obtained from our method is better than the baseline prediction model based on moving averages