This work stands apart: there is almost nothing nearby, with the nearest neighbor at a distance of 0.526. Surrounding areas include topics related to large language models, genes, explainable AI, and self-supervised learning, but there are no direct overlaps. Traditionally, time series prediction relied on linear models or nearest-neighbor methods in reconstructed state spaces. However, the curse of dimensionality quickly devalues any attempt to use more than a few predictors. Existing selection strategies, such as ranking by mutual information or direct selection with conditional mutual information, either include redundant variables or fail to capture synergistic effects, where only the joint action of several variables provides predictive power.
The authors build a three-step scheme. First, a modified PC algorithm (named after Peter Spirtes and Clark Glymour) identifies the causal parents of the target variable in the time series graph. The algorithm iteratively tests conditional independencies using a nearest-neighbor mutual information estimator with the maximum norm, which avoids the curse of dimensionality. Then, for all subsets of the found causal predictors, the multivariate mutual information with the target variable is estimated, and the subset with the maximum value is selected. Finally, the forecast is made either by nearest neighbors or by fitting a linear model on the selected predictors. For choosing the number of predictors, a heuristic criterion based on the bias of the mutual information estimate is proposed and compared with cross-validation.
On synthetic nonlinear stochastic delay processes, the optimal scheme achieves a median relative prediction error of zero, and in 90% of realizations the error is below 0.1. This is significantly better than schemes based on mutual information or direct selection, which often pick non-causal variables and miss synergistic combinations. The computational complexity of the optimal scheme grows exponentially with the number of causal predictors, but in practice it is lower than that of non-causal direct selection if causal predictors can be reduced to 8–10. For the El Niño index (Nino3.4), a linear model on optimal predictors reduces forecast error by 0.05–0.1 compared to the nearest-neighbor method for horizons up to 5 months.
Unlike neighboring works that focus either on model explainability or representation learning, this paper offers a rigorous theoretical foundation for variable selection in prediction. It shows that causality is not just a philosophical category but a practical tool that turns exponential search into a feasible task. The method combines the advantages of a model-free approach for discovering relevant variables with those of a model-based approach for efficiently using these variables, paving the way for more accurate forecasts in climatology, economics, and other fields.
The topic could evolve toward adapting the scheme for forecasting rare events, such as the onset of El Niño, rather than just average values. Also promising is the investigation of criteria for selecting the number of predictors that better account for the bias of mutual information estimates. Extensions to non-stationary processes and data where causal relationships change over time are possible.
The method could improve forecasts in climatology, econophysics, neuroscience, and other areas where data come as multivariate time series. It also offers a general framework for statistical inference when mechanisms are poorly understood.
The immediate next step is adapting the scheme for forecasting extreme events, as indicated in the paper. It is also worth testing the method on a broader class of real-world data and comparing it with modern machine learning methods.
This work directly addresses the unresolved problem of variable selection in multivariate time series, where the curse of dimensionality and overfitting remain major obstacles. It also touches on the fundamental question of how to distinguish causal relationships from correlations in observational data.
The PC algorithm underlying causal selection was proposed in 1991 and named after the initials of its creators: Peter Spirtes and Clark Glymour. Since then, it has become a cornerstone of causal inference, and here it is applied for the first time to accelerate optimal prediction.
The idea of selecting causal drivers echoes the 'psychohistory' from Isaac Asimov's Foundation series, where mathematicians predict the future of society by identifying key driving forces. Instead of a galactic empire, here it is a climate index, but the principle is the same: find the few variables that truly govern the system.