Earthquake Forecasting: Why Neural Point Processes Aren't There Yet

predicting earthquakes, you'd think the latest tech would have the edge. But it seems the old guard is holding strong. For decades, the epidemic-type aftershock sequence (ETAS) model has been a go-to for forecasting when and where earthquakes might hit. Enter Neural Point Processes (NPPs) with promises of greater flexibility and accuracy. Yet, reality checks out.

The Benchmark Bottleneck

Here's the kicker: the benchmarks used to measure NPPs are outdated. They suffer from data leakage and, astonishingly, skip the largest earthquake sequence in the region. It's like ranking sprinters without including Usain Bolt. No wonder NPPs aren't shining.

To shake things up, EarthquakeNPP steps in. It's a new benchmarking platform that pulls together global earthquake catalogs, the ETAS model, and evaluation methods straight from the seismology scene. Covering regions in California from 1971 to 2021, this platform aims to set a new standard.

Testing the Tech

Benchmarking trials using EarthquakeNPP show something wild: out of five NPPs tested, none beat ETAS. Let that sink in. The shiny new AI couldn't top a model that's been around for decades. So, what's the deal? Are NPPs just not ready for prime time, or is ETAS just that good?

This isn't just an academic spat. Real people in earthquake-prone zones could be relying on models that aren't up to snuff. If you're in California, you want your forecasting to be rock solid.

Looking Forward

Despite the current shortcomings, EarthquakeNPP isn't just a critique. It's a bridge for collaboration between seismologists and machine learning experts. That's something the field sorely needs. But let's be clear: the NPPs have yet to prove themselves. The labs are scrambling to refine these models, but can they catch up?

JUST IN: The leaderboard stays as is, for now. But EarthquakeNPP might just be the seismic shift the community needs to push innovation forward.