On the optimal prediction of extreme events in heavy-tailed time series with applications to solar flare forecasting

V. K. Verma, Stilian Stoev, Yang ChenarXiv 2024 · DOI · arXiv

TL;DR

This paper establishes a rigorous theoretical framework for the "optimal" prediction of rare, high-impact events in time series with heavy-tailed distributions (where extreme outliers are more common than in a normal distribution). By applying this theory to solar flare forecasting, the authors demonstrate that while heavy tails actually make extreme events more predictable than Gaussian models do, there are fundamental mathematical limits to how much we can foresee using historical data alone.

What Problem Does This Paper Address?

Most time-series forecasting focuses on minimizing the Mean Squared Error (MSE), which prioritizes accuracy for "average" observations. However, in fields like finance or heliophysics, the events we care about most—market crashes or massive solar flares—are outliers.

Predicting these "extreme events" is a binary classification problem: will the value exceed a very high threshold $y_0$ in the future? While machine learning has been thrown at this problem (especially for solar flares), we’ve lacked a clear understanding of the theoretical limits of such predictions. This paper fills that gap by defining what an "optimal" predictor looks like for heavy-tailed data and calculating the maximum possible precision we can expect.

Methods

The authors bridge the gap between classical statistics and extreme value theory (EVT) using several key techniques:

  • Neyman-Pearson Characterization: Borrowing from hypothesis testing, they prove that the optimal predictor for an extreme event is a ratio of two conditional densities. This allows them to derive closed-form "best-case" predictors for various models.
  • Calibration: They define a "calibrated" predictor as one that raises alarms at the same rate the events actually occur. This provides a fair baseline for comparing different forecasting methods.
  • Extremal Precision: To measure success, they use the "upper tail dependence coefficient." This metric quantifies the probability that an event actually happens given that the model predicted it, specifically as the threshold for "extreme" goes to infinity.
  • Plug-in Estimators: For practical use in Autoregressive (AR) models, they developed an algorithm that uses Least Absolute Deviation (LAD) or Ordinary Least Squares (OLS) to estimate model parameters from training data and then applies the optimal theoretical threshold.

Datasets

The primary application uses GOES (Geostationary Operational Environmental Satellites) soft X-ray flux measurements.

  • The data covers the peaks of Solar Cycles 23 and 24 (2000–2002 and 2011–2014).
  • The raw one-minute intervals were processed into a series of hourly maxima, resulting in ~50,400 observations.
  • The authors modeled this data using FARIMA (Fractional ARIMA) models, which are specifically designed to handle both "long memory" (long-range dependence) and heavy tails (infinite variance).

Key Insights and Ideas

  • Heavy Tails are a Blessing (for Prediction): One of the most surprising theoretical findings is that in a Gaussian (normal) world, the precision of predicting an extreme event actually drops to zero as the event becomes more extreme. In contrast, in a heavy-tailed world, the precision remains positive. Essentially, the very "weirdness" of heavy-tailed data makes its extremes more mathematically linked to its past.
  • The "Baseline" is Hard to Beat: For many heavy-tailed models, a "baseline" predictor—which simply assumes an extreme event is imminent if the current value is high—is nearly as good as complex models that look at the entire history.
  • Theoretical Ceilings: By using multivariate regular variation, the authors derived a mathematical "speed limit" for prediction. This tells researchers exactly how much precision is being lost due to the inherent randomness of the system versus the limitations of their specific model.

Key Results

The authors compared their optimal AR and FARIMA predictors against a simple baseline across different "lead times" ($h$) and "extremeness" levels ($p$):

  • Short-term ($h=1$ hour): The baseline was remarkably strong, achieving a precision of ~0.49 for the 90th percentile. This means nearly half of the "alarms" were followed by an actual event.
  • Longer-term ($h=6$ hours): Precision dropped significantly. For the most extreme events (99th percentile), the AR model actually outperformed the baseline, but the overall precision was low (~0.12).
  • FARIMA Performance: While FARIMA models are theoretically better at capturing long-term trends, they only slightly outperformed the baseline at the 6-hour mark, suggesting that the "memory" of solar flux is quite volatile.

Limitations and Future Work

  • Univariate Constraint: The study only looked at X-ray flux. The authors acknowledge that "operational" forecasting (real-world use) would require covariates, such as magnetograms or solar images, to break through the theoretical limits identified here.
  • Stationarity: The models assume the time series is stationary (its statistical properties don't change over time). However, solar activity is famously cyclical and non-stationary, which might degrade model performance over long periods.
  • The "Rare Event" Data Gap: Estimating parameters for the "tail" of a distribution is notoriously difficult because, by definition, there are very few data points in that region. Future work could explore more robust ways to estimate these density ratios with sparse data.

Summary updated

Abstract

The prediction of extreme events in time series is a fundamental problem arising in many financial, scientific, engineering, and other applications. We begin by establishing a general Neyman-Pearson-type characterization of optimal extreme event predictors in terms of density ratios. This yields new insights and several closed-form optimal extreme event predictors for additive models. These results naturally extend to time series, where we study optimal extreme event prediction for both light- and heavy-tailed autoregressive and moving average models. Using a uniform law of large numbers for ergodic time series, we establish the asymptotic optimality of an empirical version of the optimal predictor for autoregressive models. Using multivariate regular variation, we obtain an expression for the optimal extremal precision in heavy-tailed infinite moving averages, which provides theoretical bounds on the ability to predict extremes in this general class of models. We address the important problem of predicting solar flares by applying our theory and methodology to a state-of-the-art time series consisting of solar soft X-ray flux measurements. Our results demonstrate the success and limitations in solar flare forecasting of long-memory autoregressive models and long-range-dependent, heavy-tailed FARIMA models.
BibTeX
@misc{verma2024optimal,
  author = {V. K. Verma and Stilian Stoev and Yang Chen},
  title = {{On the optimal prediction of extreme events in heavy-tailed time series with applications to solar flare forecasting}},
  year = {2024},
  eprint = {2407.11887},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.11887},
}
Figure