Machine Learning for Extreme Event Data Scarcity

130 concepts118 papers3 proposals

Standard machine learning models are inherently "average-case" engines. By minimizing mean squared error or maximizing likelihood, they prioritize the bulk of a distribution where data is plentiful. However, in high-stakes domains like flood forecasting, financial risk management, and autonomous driving, the most critical events exist in the "long tail"—the rare, high-impact occurrences that are, by definition, scarce in historical records.

The central challenge of this research direction is the extrapolation gap: how to build models that remain reliable in regimes where they have seen little to no training data.

The Failure of Average-Case Modeling

When training data is concentrated around average performance levels, models often suffer from a "regression to the mean." As Michael Omori notes in Chess Rating Estimation from Moves and Clock Times Using a CNN-LSTM, this bias leads to significant inaccuracies when predicting extreme outliers, such as top-tier experts, because their play styles are underrepresented.

In environmental science, this manifests as "blurriness" in forecasts. Haoran Bi et al. explain that deep learning models often underestimate high-magnitude phenomena because they prioritize more frequent, lower-intensity observations. Without specialized intervention, a model trained on 100 years of weather data may still fail to predict a "1,000-year flood" simply because the underlying physics of such an event have not been sufficiently sampled.

Bridging the Gap with Extreme Value Theory

To overcome this scarcity, researchers are integrating Extreme Value Theory (EVT)—a branch of statistics dealing with the asymptotic behavior of tails—into deep learning architectures. EVT provides a mathematical justification for extrapolating beyond the range of observed data.

One approach involves modifying the objective function. Haoran Bi et al. and the authors of Nowcasting of Extreme Precipitation Using Deep Generative Models developed a specialized "Extreme Value Loss" that forces the model to learn the distribution of rare peaks rather than just the average rainfall. Similarly, Nicola Gnecco et al. propose Extremal Random Forests (ERF) in Extremal Random Forests, which combine the local weighting of random forests with EVT to predict "black swan" events in high-dimensional datasets where traditional regression fails.

Generative Augmentation of the Tail

If the historical record is too short to provide enough extreme samples, another strategy is to synthesize them. Standard Generative Adversarial Networks (GANs) typically ignore outliers to improve overall sample quality. To counter this, Siddharth Bhatia et al. introduced ExGAN in ExGAN: Adversarial Generation of Extreme Samples. By incorporating EVT into the GAN framework, the authors enable the generation of realistic, hypothetical extreme samples—like financial crashes or catastrophic storms—that follow the mathematical properties of a distribution's tail.

This synthetic approach is also seen in hurricane forecasting. The framework described in HurriCast: Synthetic Tropical Cyclone Track Generation for Hurricane Forecasting generates "plausible but hypothetical" storm scenarios to help insurers assess risks in regions where the 150-year historical record is insufficient to capture the full spectrum of potential disasters.

Robustness Through Constraints and Quantiles

When data is too sparse for even generative models, researchers turn to structural and geometric constraints. Henry Lam and Clementine Mottet argue in Tail Analysis Without Parametric Models: A Worst-Case Perspective that by assuming the tail density is convex, one can derive robust, statistically valid bounds for rare events even with extremely sparse data.

In time-series forecasting, the focus shifts from predicting a single value to estimating extreme quantiles (e.g., the 99.9th percentile). Danielle C. Maddix and Youngsuk Park propose Incremental Spline Quantile Functions in Learning Quantile Functions without Quantile Crossing for Distribution-free Time Series Forecasting to ensure that these tail estimates remain consistent and monotonic, preventing the "quantile crossing" problem that often plagues models in data-poor regions.

Causal Inference in the Extremes

Finally, understanding why an extreme event occurs is as important as predicting it. Jiyuan Tan et al. bridge the gap between causal inference and EVT in Estimation of Treatment Effects in Extreme and Unobserved Data. By developing the Normalized Extreme Treatment Effect (NETE), the authors provide a way to estimate how interventions (like a new seawall or a policy change) affect the probability of rare, high-impact outcomes that have not yet been recorded in the observed data.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_extreme_event_data_scarcity_3f5a351e,
  title = {extreme event data scarcity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extreme-event-data-scarcity/rdr_3f5a351e04e5413f9539fe83111c65d3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_extreme_event_data_scarcity_3f5a351e_v1,
  title = {extreme event data scarcity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extreme-event-data-scarcity/rdr_3f5a351e04e5413f9539fe83111c65d3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake