Home

CAOS_SEISMIC

Conditional probabilistic seismic forecasting — forecasts, never predictions.

CAOS_SEISMIC is an independent, open, honest research and education tool that reads the recent state of global seismicity and publishes bounded, calibrated, conditional probabilities of earthquakes over short horizons (1 day / 2 days / 7 days), for a region and a magnitude band, always shown next to the long-term baseline and evaluated prospectively against reality with the field-standard CSEP framework. It produces one inference per day, renders a continuous probability field (never alarm dots), and never issues an alarm, a countdown, a binary call, or a "safe" state.

The creed (carries into the product verbatim)

Earthquakes cannot be predicted, but their probability can be forecast — reported honestly, with uncertainty, evaluated against reality, never as an alarm and never as a promise of safety.

• Live site: https://seismic.fasl-work.com
• Source repository: https://github.com/fsantibanezleal/CAOS_SEISMIC

---

What CAOS_SEISMIC is — and is not

It is	It is not
A forecaster: every published number is a probability strictly in $(0, 1)$, scoped to a region × magnitude band × horizon.	A predictor: it never makes a deterministic "an earthquake will / will not happen" statement.
An honest research/education tool that complements official Operational Earthquake Forecasting (USGS, INGV, CSN, GeoNet, JMA).	An authoritative civil-protection alarm, nor a competitor to official agencies.
Calibrated and prospectively scored — the live reliability diagram ("when we said 5 %, it happened ~5 % of the time") is its central credibility artifact.	An accuracy/AUC leaderboard. AUC is banned as a primary forecasting metric (it is calibration-blind).
A continuous probability field with explicit horizon, magnitude threshold, baseline ratio, and uncertainty bounds.	A map of red "alert" pins. Red appears only on the calibration badge (model quality), never on the forecast field.

Following the ICEF definition (Jordan et al., 2011), a prediction is a deterministic statement that an event will or will not occur; a forecast gives a probability strictly between 0 and 1. Short-term probabilities of a large event "may vary over orders of magnitude but typically remain low in an absolute sense (< 1 % per day)." A single outcome neither validates nor invalidates a probabilistic forecast — a 3 % forecast is not wrong when the 3 % outcome occurs.

---

The thesis in one paragraph

Global context conditions local short-term forecasts. Deterministic earthquake prediction is effectively impossible — whether a small rupture cascades into a great one depends on unmeasurably fine detail of the crust (Geller et al., 1997), with self-organized criticality (Bak & Tang, 1989) as the leading explanatory framework. What is achievable is Operational Earthquake Forecasting (OEF): conditional probabilities that stay low in absolute terms even when the relative gain during an active sequence is one-to-three orders of magnitude. CAOS_SEISMIC trains on global seismicity and complementary global covariates, fits a calibrated space–time ETAS reference (the de-facto operational baseline any model must beat), and runs inference across many countries — both high-seismicity (Chile, Japan, Indonesia, Mexico, Türkiye, California, New Zealand) and low-seismicity (United Kingdom, Germany, Australia, Brazil) — precisely so the forecast can be compared and audited for bias toward high-seismicity zones. Any stronger model (a context-conditioned neural temporal point process) reaches the public map only if it beats ETAS in our own prospective CSEP harness and is calibrated. The single public exceedance formula $P = 1 - e^{-N}$ never changes; only the quality of the conditional intensity $\lambda$ improves.

---

How CAOS_SEISMIC works (at a glance)

flowchart TD
    A["Public data feeds (read-only)<br/>USGS ComCat · ISC-GEM · GCMT · regional FDSN networks"] -->|daily pull| B
    B["Offline daily job (one inference per day)<br/>hygiene → fit/condition ETAS → simulate ensemble → calibrate"] -->|write ONE compact artifact| C
    C["Compact artifact (committed, gzipped, few hundred KB – few MB)<br/>per-cell rates · baseline · P10/median/P90 bounds · CSEP summary · coverage mask · provenance"] -->|served static| D
    D["Static web viewer (Vite + React + TS)<br/>world probability FIELD · per-country drill-down · no-map summary · reliability diagram"]

Loading

Heavy compute runs once per day, offline; the runtime is a pure static viewer with no processing backend. The mandatory data-hygiene order is load-bearing: time-varying completeness $M_c(x,y,t)$ → moment-magnitude homogenization → dual-catalog declustering. A strict forecast clock hands the model only the catalog slice $(-\infty, t)$ at each issue time, making temporal leakage structurally impossible. See Pipeline and Technical-Architecture for the full flow.

---

Table of contents

This wiki is structured into deep, single-topic sub-pages. Each model and method has its own page — intuition and history, governing equations (with derivation), assumptions, parameter estimation, strengths and limitations, its role in operational forecasting, a worked illustration, and a References section with DOIs. The two index pages link the sub-pages with one-line summaries.

Overview

• Home — this page: what CAOS_SEISMIC is and is not, the creed, the thesis.
• Honest-Limits — the epistemics: why prediction is impossible, the absolute-vs-relative scale of probabilities, the L'Aquila communication lesson, and the worked Ridgecrest example.

Methodology & History

• Methodology-History — the scientific methodology, the project's design decisions, and how the deep-research synthesis became the authoritative spec.

Classical models

• Models-Classical — index of the analytical, physics-informed and statistical baseline.
• Gutenberg-Richter-Law — the frequency–magnitude relation; the magnitude term of every forecast.
• Omori-Utsu-Law — the modified Omori aftershock-decay law and its productivity.
• ETAS-Model — the Epidemic-Type Aftershock Sequence; the de-facto operational baseline to beat.
• Reasenberg-Jones-Model — the original operational next-day aftershock-probability model.
• STEP-Model — Short-Term Earthquake Probability; the production daily-map reference shape.
• EEPAS-Model — Every Earthquake a Precursor According to Scale; the medium-term scaling model.
• Smoothed-Seismicity — the time-independent spatial background $\mu(x,y)$ and mandatory null.
• Brownian-Passage-Time — the BPT renewal model for long-term fault recurrence.
• Rate-and-State-and-Coulomb — the mechanistic stress-transfer layer (Dieterich, $\Delta$CFS).

ML & analytical methods

• Models-ML — index of the analytical / machine-learning model family and the honest verdict.
• Temporal-Point-Processes — the unifying conditional-intensity framework behind every model.
• RMTPP — Recurrent Marked Temporal Point Process; the first neural TPP.
• Neural-Hawkes-Process — the continuous-time LSTM self-modulating point process.
• Transformer-Hawkes-Process — attention-based TPPs (THP / self-attentive Hawkes).
• RECAST-and-FERN — the two neural TPPs built specifically for earthquakes.
• CNN-Spatial-Models — CNN spatial forecasting and the DeVries (2018) cautionary tale.
• Graph-and-Recurrent-Networks — GNN/RNN/LSTM approaches and where they fit.
• Detection-vs-Forecasting — the hard line: detection is not forecasting.

Models employed

• Models-Employed — the models CAOS_SEISMIC actually runs: the ETAS reference + smoothed Poisson null, the tiled global conditioning by tectonic regime, and the gated neural challenger.

Data

• Data-Sources — the catalog spine and regional networks, the long-term homogeneous anchor, mechanism and geophysical enrichers, the tidal covariate, licensing, and attribution.
• Data-Types-and-Features — the catalog schema, derived features, and the hygiene order.

Architecture

• Technical-Architecture — the deploy architecture: GPU-laptop compute, git-as-data publishing, the static viewer, the compact artifact contract, and operational QA gating.
• Pipeline — the versioned pipeline DAG stage by stage, catalog hygiene, the forecast clock, and reproducibility via input-state snapshots.

Evaluation

• Evaluation-and-Tests — the CSEP credibility backbone: the two forecast representations, consistency tests (N / M / S / L / CL), comparison tests (information gain, T/W), proper scoring rules, the Molchan/ROC view, calibration, and leakage discipline.

Progress

• Changelog-and-Progress — the project history, phases, releases, and current status.

Reference

• References — the consolidated, deduplicated bibliography (every citation, with DOIs).
• Glossary — precise, sourced definitions of every domain term used across the project.

---

Honest framing, throughout

Every number is a probability in $(0, 1)$, scoped to a region, a magnitude band, and a horizon, reported with a real (epistemic + aleatory) uncertainty band and shown alongside its long-term baseline. The product never renders an alarm, a countdown, a binary yes/no, or a "safe" state. Blank on the map never means "safe" — it means no forecast / out of coverage. The product positions strictly as an independent tool that complements, and never replaces, official agencies. See Honest-Limits for the full epistemic basis.

This wiki is the authoritative technical record. It cites only canonical, peer-reviewed literature (with DOIs) and official USGS / ISC / CSEP documentation.