Talivio uses per-region band ML models trained via an expanding-window backtest (2000–present). For each seismic region and magnitude band, an algorithm competition (up to 5 candidates: LightGBM, RandomForest, ExtraTrees, GradientBoosting, CalibratedLR) selects the best single model using walk-forward cross-validation with hard same-region temporal negatives — preventing geographic classification leakage. Regional ROC-AUC values range from 0.62 to 0.99 across 0 global regions. CSEP-compatible forecast export for independent validation. Weekly automatic retraining.
The regional band model architecture trains a separate ML model for each (region, magnitude band) pair using an expanding-window backtest. Hard same-region negatives (temporal: same coordinates, different time; spatial: offset within the same seismic zone) are used to prevent geographic leakage. A per-band algorithm competition (up to 5 candidates — LightGBM, RandomForest, ExtraTrees, GradientBoosting, CalibratedLR; the candidate pool varies by band) selects a single winning algorithm via cross-validated ROC-AUC, then fits it with sigmoid (Platt) probability calibration. The global prediction endpoint retains the 102-feature ensemble (Coulomb stress, ETAS, GR b-value, multi-scale SRA, H3 spatial connectivity) for backwards-compatible API access.
Each band has a different feature dimensionality, data density, and dominant physical mechanism. In sparse bands (M6+) the seismic cycle and physical parameters take precedence, while in dense bands (M4–5) statistical precursors carry more weight.
Highest event rate; serves as a precursor indicator for larger events. Micro-cluster density and foreshock patterns are key signals.
Most frequent damaging class; high training data availability. Foreshock density and quiescence anomaly are the dominant signals in this band.
Magnitude range capable of causing damage. The 24-month linear regression of the b-value trend is the leading precursor signal in this band.
Potential for serious structural damage and casualties. Seismic cycle analysis with inter-event normalisation plays the decisive role.
Least training data; hardest prediction class. Physical parameters (slip rate, last rupture date, fault length) gain greater weight.
The ML feature vector contains 102 core features (94 temporal/mechanical + 8 spatial) grouped into the categories below; magnitude bands add band-specific extras on top (118–125 total per band). Coulomb stress transfer, the ETAS statistical model and multi-window SRA were added in v4.0. The v5.0 FFT spectral block was replaced in v5.1 by a Morlet-wavelet signal-analysis block, since seismicity time series are non-stationary (Torrence & Compo, 1998); IERS 2010 solid-earth-tide computations also entered in v5.0.
| Feature | Physical Meaning | Source |
|---|---|---|
| Fault distance | Distance to nearest active fault segment (km) | GEM Global Active Faults |
| Coupling ratio | Fault coupling ratio (0=creeping, 1=fully locked); elastic energy accumulation indicator | Fault DB + GNSS |
| Slip rate | Long-term slip rate of the fault (mm/yr); rate of energy accumulation | GEM / AFAD Fault DB |
| Cross-fault velocity | GNSS cross-fault differential velocity (mm/yr); plate motion difference across fault | GNSS microservice |
| Seismic gap | Time elapsed since last major earthquake (years); identifies mature segments | Historical catalogue |
| Last rupture | Time since last surface rupture; basis of seismic cycle calculation | Historical catalogue |
| Feature | Physical Meaning | Source |
|---|---|---|
| Cumulative ΔCFS | Cumulative Coulomb stress effect of historical M≥5.5 earthquakes (bar) | Computation + USGS/AFAD |
| Recent ΔCFS | Largest individual Coulomb stress change in the recent period (bar) — triggering potential | Computation |
| Contributing count | Number of historical sources contributing to the cumulative ΔCFS at the target | Computation |
| Feature | Physical Meaning | Source |
|---|---|---|
| ETAS density | ETAS model instantaneous seismicity rate (events/day/km²) | Computation (Ogata 1988 form) |
| Background rate | Background seismicity rate μ (excluding clustering); tectonic loading indicator | Computation |
| Triggering ratio | Triggered/total ratio; high value = active aftershock sequence | Computation |
| Contributing count | Number of prior events contributing to the triggered intensity | Computation |
Rates are computed over seven windows (3d / 7d / 14d / 30d / 90d / 180d / 1y) plus derived ratios, acceleration, quiescence/elevated flags and the Z-score below.
| Feature | Physical Meaning | Source |
|---|---|---|
| Short-term rate | Regional earthquake count over the short term; short-term activity | USGS + AFAD |
| Medium-term rate | Earthquake count over the medium term; medium-term activity | USGS + AFAD |
| Long-term rate | Earthquake count over the long term; long-term background | USGS + AFAD |
| Activity ratio | Short/long-term activity ratio (SRA). Detection of uplift or quiescence | Computation |
| Quiescence Z-score | Deviation of activity from historical mean. Anomaly detection (Habermann, 1988) | Computation |
| Feature | Physical Meaning | Source |
|---|---|---|
| b-value | Gutenberg-Richter b-value (Aki 1965 MLE); low b = high stress | Seismic catalogue |
| b-value trend | Time-series slope of b-value; decreasing trend signals stress increase | Computation |
| Precursor density | Density of small nearby earthquakes (foreshock density) | USGS + AFAD |
| Moment deficit | Accumulated/expected seismic moment ratio; ≥1 indicates energy surplus | Fault DB + Catalogue |
| Feature | Physical Meaning | Source |
|---|---|---|
| GNSS deformation | Crustal deformation rate computed via spatial interpolation from GPS stations (nstrain/yr) | NGL Nevada Geodetic Lab |
| Tidal index | Combined Moon+Sun tidal stress (normalised 0–1) | Ephemeris computation |
These IERS-2010 quantities are computed by the geophysics engine and feed the trigger component of the 4-component risk model and the geophysics API; the ML feature vector carries the consolidated tidal_stress_index (§3.6) rather than these raw quantities directly.
| Feature | Physical Meaning | Source |
|---|---|---|
| Volumetric strain | Volumetric strain computation based on IERS 2010 conventions (nanostrain). Typical amplitude: 50–120 ns | Computation (IERS 2010) |
| Tidal Coulomb | Coulomb projection of tidal stress onto the fault plane (kPa). Depends on fault geometry | Computation |
| Feature | Physical Meaning | Source |
|---|---|---|
| rolling_entropy_30d | Shannon entropy of the 30-day activity window (0–1); rising = unpredictable, potential precursor | Computation |
| rolling_variance_ratio | Short (30d) / long (180d) rolling-variance ratio; activity becoming erratic | Computation |
| wavelet_low_high_ratio | Morlet-CWT low- (scale>30d) / high-frequency (scale<14d) energy ratio; energy shifting low | Computation |
| sma_ratio | Short (7d) / long (90d) moving-average ratio of the count series | Computation |
| activity_trend | Linear-regression slope of the last 90 days; positive = acceleration | Computation |
| short_rolling_entropy | 14-day window entropy at tighter radius — foreshock signal | Computation |
| short_variance_change | Recent-14d vs prior-14d variance change | Computation |
Fault locking, slip deficit, and stress transfer from neighbouring earthquakes. Weighted average of three sub-components.
ETAS model seismicity density estimate and short/long-term activity ratio (SRA). Operational forecasting standard.
b-value decrease, seismic quiescence, and foreshock density. Weighted combination of four sub-signals.
Moon–Sun tidal stress; triggering potential when fault is near threshold. Scientifically small (~1–3%) but statistically significant.
Fault segments with ΔCFS > +0.1 bar (10 kPa) receive an automatic uplift in the mechanical component. King et al. (1994) triggering threshold.
If Mechanical ≥ 0.6, Statistical ≥ 0.5, Precursor ≥ 0.4, and Trigger ≥ 0.6 are simultaneously met, the system raises the Critical Window flag.
The earthquake catalogue is divided into spatial cells via a hierarchical hexagonal grid. Spatial features are computed using a multi-scale neighbourhood structure. Temporal leakage is prevented by leak-free binary search (only past events are used).
spatial_neighbor_rate_k1/k2: 90-day earthquake activity rate in near and extended neighbour cells.
spatial_max_mag_k1/k2: Maximum earthquake magnitude within 90 days in neighbouring cells.
spatial_activity_trend: Last 30-day vs. 90-day activity ratio — detection of acceleration or deceleration.
spatial_cluster_density: Ratio of active cells in the centre + K1 neighbours — measure of spatial clustering.
spatial_fault_smoothed_risk: Distance-weighted fault cell seismicity using exp(-dist/decay_km).
spatial_strain_gradient: Standard deviation of the activity difference between the centre cell and its neighbours.
Rather than a fixed ensemble, each (region, band) model is chosen by an algorithm competition: up to five candidates are cross-validated and the single highest-ROC-AUC algorithm wins. Training data is split chronologically (walk-forward): the competition and model fit use the earliest ~60% of samples, sigmoid calibration is fitted on the next ~20% (a held-out calibration slice), and all reported metrics come from the final ~20% — data strictly later in time than anything the model or its calibrator has seen. The candidate pool is band-dependent — e.g. LightGBM competes only in the denser M3-4/M4-5/M5-6 bands, GradientBoosting in the sparser M5-6/M6-7/M7+ bands.
n_estimators=300, learning_rate=0.06, num_leaves=31, max_depth=6, L1+L2 (reg_alpha=0.1, reg_lambda=1.0), is_unbalance. Competes in M3-4/M4-5/M5-6.
n_estimators=300, max_depth=8, min_samples_leaf=10, class_weight='balanced'. Reliable baseline in every band.
n_estimators=300, max_depth=8, max_features='sqrt'. Extra random splits add decorrelation; competes in every band.
n_estimators=200, learning_rate=0.05, max_depth=4, subsample=0.8. Sample-weighted; competes in M5-6/M6-7/M7+.
The winner is calibrated with sigmoid (Platt) scaling fitted on a held-out, temporally later calibration slice — never on the data used for algorithm selection. Sigmoid is monotonic (preserves ROC-AUC) and does not overfit on sparse bands the way isotonic can. The sparse M7+ path also uses sigmoid; when a band lacks enough positives for a reliable hold-out, the system falls back to cross-validated calibration and records this in the model metadata.
Daily CSEP-compatible grid forecasts (ASCII/CSV/XML) are generated for independent validation via cseptesting.org. GR-Poisson + ML risk overlay.
The multi-layer stacking ensemble used prior to v7.0. Retained here for reference.
| Algorithm | Version | Strengths | Band Advantage |
|---|---|---|---|
| XGB XGBoost | 2.x | Histogram-based splitting, L1/L2 regularisation, GPU support, missing value handling | M4–5, M5–6 (dense data) |
| LightGBM LightGBM | 4.x | DART regularisation, leaf-wise growth, scale_pos_weight, fastest training | M4–5 (speed + accuracy) |
| CB CatBoost | 1.x | Ordered boosting (target leakage prevention), native categorical features, low overfitting | M5–6, M6–7 |
| ET ExtraTrees | sklearn 1.x | Extra random splits; low variance, fast training | M6–7 (sparse data) |
| RF Random Forest | sklearn 1.x | Bagging, interpretability, Gini importance ranking, SHAP values | All bands (baseline) |
The platform draws on 8 different data sources. The historical catalogue is merged from multiple sources; after magnitude homogenisation (ML→Mw, Ulusay et al. 2004), aftershocks are separated using Gardner-Knopoff (1974) declustering. Record priority: AFAD > USGS > ISC > ISC-GEM > Ambraseys.
According to elastic rebound theory, a fault segment accumulates elastic energy proportional to the regional slip rate since the last major earthquake. McCann et al. (1979) classify mature and recently ruptured segments using the seismic gap criterion.
McCann et al. (1979): fault segments from which a long time has passed since the last major earthquake are flagged as "probable gaps". Quiet segment = accumulated stress.
Low b (b < 0.8) signals high crustal stress and an approaching large event (Wiemer & Wyss, 2002; Scholz, 1968). b = 1.0 is the long-term regional average.
Characteristic earthquake period ~250 years; westward-migrating rupture sequence (1939–1999). The Marmara segment has not had a major rupture for >300 years.
Mean recurrence interval ~300 years; the 2023 Kahramanmaraş earthquakes (Mw 7.8 + Mw 7.6) were simultaneous ruptures of multiple segments of this fault.
The PyEphem library uses VSOP87 analytical planetary theory to compute the positions of the Moon and Sun with nano-radian precision. Tidal stress is projected onto fault planes using Boussinesq load theory.
During New Moon and Full Moon periods (±5-day window) the Moon-Sun-Earth alignment maximises tidal stress. These periods are flagged as trigger pressure windows.
Near perigee the Moon's gravitational effect increases by 14% (1/r³ dependence). Perigee + Full Moon coincidence (Supermoon) produces the strongest tidal pressure.
The 18.6-year lunar orbital inclination oscillation (Saros cycle) creates a small (~1%) modulation in tidal stress; its effect is observed in long-term statistics.
Normalised: (F_Moon/r³_Moon + F_Sun/r³_Sun) × syzygy_factor. Syzygy factor: 1 + 0.2 × cos(2π × lunar_phase). Range: ~27 MPa (minimum) – ~65 MPa (maximum).
When a large earthquake occurs, stress redistributes in the surrounding crust. The Coulomb Failure Stress (CFS) change quantifies how close neighbouring faults are brought to failure. This mechanism has successfully explained many earthquake sequences including the 1994 Northridge and 1999 İzmit events.
The source double-couple's full static stress tensor is computed (Kelvin point-source solution; σ=λ·tr(ε)·I+2μ·ε) and the traction is resolved onto the receiver fault plane (shear + normal). Elastic parameters: μ=3.2×10¹⁰ Pa, ν=0.25, μ'=0.4. Homogeneous full-space; near-field saturated at the rupture half-length. (Half-space Okada-DC3D free-surface correction is a future refinement.)
Slip amount from magnitude: log₁₀(AD) = -4.80 + 0.69·M (metres). Rupture length: log₁₀(RLD) = -2.44 + 0.59·M (km). Rupture width: log₁₀(RW) = -1.01 + 0.32·M (km).
ΔCFS > +0.1 bar (10 kPa) triggering potential (Stein, 1999). Cumulative ΔCFS effect of all M≥5.5 earthquakes in the past 50 years is computed and projected onto target fault segments.
1999 İzmit (Mw 7.6) → Düzce segment ΔCFS: +3.2 bar → Düzce earthquake 87 days later. 1992 Landers → Big Bear triggering. 2023 Maraş multi-segment rupture.
Epidemic-Type Aftershock Sequence (ETAS) is a point process that models seismic activity as the sum of the background seismicity rate and the triggering potential of each event. Used by USGS, Italy INGV, and Japan JMA for operational earthquake forecasting.
The official U.S. Operational Aftershock Forecasting system is ETAS-based. Aftershock probability forecasts are published within 1 hour of each major event.
The 5 parameters {K, α, c, p, μ} are estimated by maximum-likelihood optimisation of the temporal ETAS log-likelihood (Ogata 1988) with L-BFGS-B, per region from past-only events (leak-safe). Falls back to a Reasenberg & Jones (1989) estimator when data are sparse or optimisation fails.
SRA = R_7day / (R_1yr/52). SRA>2: anomalous increase. SRA<0.3: seismic quiescence — Habermann (1988) potential precursor anomaly.
With M≥1.5 completeness, AFAD-driven ETAS parameterization uses 10× more events. Background rate estimation (μ) and b-value computation improve significantly.
Solid Earth tidal stress from the Moon and Sun is computed physically using Love numbers (h₂=0.6078, l₂=0.0847, k₂=0.2980) based on IERS 2010 conventions. Provides more accurate volumetric strain and Coulomb projection than the legacy 1/r³ approximation.
h₂=0.6078, l₂=0.0847, k₂=0.2980. Defines the elastic response of the crust to tidal forces. Degree-2 is the dominant component (Petit & Luzum, 2010).
Typical amplitude: 50–120 nanostrain. Maximum at New/Full Moon. Moon contribution 68%, Sun contribution 32%.
Tidal stress is projected onto Coulomb stress based on fault geometry (strike, dip). Standard crustal shear modulus used (Agnew, 2015).
Ide et al. (2016) Nature Geoscience: large earthquakes occur statistically more frequently during high tidal stress periods. Particularly on shallow thrust faults.
Extreme Value Theory (EVT)-motivated asymmetric loss functions heavily penalize missing
rare but catastrophic events like M7+ (False Negatives), as custom XGBoost/LightGBM objectives.
Status: these objectives are implemented and available in the codebase but are
not wired into the current production band competition — which instead handles class
imbalance via per-band negative-sampling ratios, class_weight='balanced' /
is_unbalance, and sigmoid calibration. The EVT/focal objectives remain an experimental track.
Baseline weight. Frequent earthquakes; false alarms and misses balanced.
Miss penalty increases. Earthquakes capable of causing damage.
FN far more costly; FP more acceptable. Serious damage potential.
False alarms are acceptable; missing an M7.5 is not (Coles, 2001).
The Molchan diagram is the internationally accepted method for scientific evaluation of earthquake forecast models. It measures model skill via the tradeoff between alarm rate and miss rate.
Excellent. Outstanding forecast skill. Far superior to random model performance.
Very Good. Statistically strong signal.
Good. Sufficient skill for operational use.
Each magnitude band evaluated with separate Molchan analysis. p < 0.05 means model is significantly different from random (Molchan, 1991).
Model performance is measured with a walk-forward expanding-window backtest from 1990 (catalogue
start) to present. The per-band backtester evaluates each candidate with a TimeSeriesSplit
whose fold count is data-driven (≈3–5, scaled to training-set size) rather than a fixed number.
Reported training metrics come from a chronological hold-out test slice (the most
recent ~20% of samples), and feature computation filters the catalogue to events strictly
before each sample's reference date (enforced in code and exercised by
test_leakage_audit.py). The system reports per-region, per-band metrics
(M3–4/M4–5/M5–6/M6–7/M7+).
Following an internal validation audit, the training and evaluation pipeline was revised. Several earlier published scores were optimistic due to methodological artefacts; the changes below make all future metrics strictly prospective-equivalent. Scores reported after this date are not directly comparable to earlier ones — expect them to be lower and more honest.
| Change | Before | After |
|---|---|---|
| Final train/test split | Random stratified shuffle — future events could inform training, inflating test AUC | Chronological walk-forward split: last ~20% of samples (in time) is the test set |
| Probability calibration | Sigmoid calibration fitted on the same data used for algorithm selection | Fitted on a separate, temporally later hold-out slice (~20% of the training window) |
| Decision threshold | Optimised on the test set (circular) | Optimised on the calibration hold-out; the test set is touched once, for reporting only |
| Base-rate correction | None — probabilities reflected the artificial 1:N sampling ratio | Prior-shift (odds-ratio) correction to the observed frequency of verified forecasts, applied once ≥30 verified outcomes exist per band |
| Backtest negatives | Synthetic feature vectors with hand-set "low-risk" ranges — label information leaked into features | Real catalogue locations through the production feature pipeline: quiet-period temporal negatives + sub-band background seismicity |
| Forecast verification | Outcome locked at first check; late catalogue entries could leave a true hit marked as a miss | 14-day grace window with miss → hit upgrade (never hit → miss); corrected labels are re-fed to online learning |
| Champion/challenger promotion | Auto-promotion on any AUC improvement (>1e-9) | Meaningful margins required (ΔAUC ≥ 0.01 temporal / ≥ 0.03 otherwise; ΔBrier ≥ 0.005); low-confidence models can never displace a champion |
| M7+ sparse band | In-sample metrics reported alongside other bands; isotonic calibration | Explicit insufficient_data / low_confidence flags (<20 positives), sigmoid calibration, excluded from automatic promotion |
A second revision focused on the prediction mechanism itself: feature parity between training and serving, an ETAS reference model as the skill baseline, grid-based forecast targets, and a physics-based renewal model for the data-starved M7+ band.
| Change | Before | After |
|---|---|---|
| Train/serve feature parity | Daily forecasts were generated without the seismicity, b-value, slip-deficit and seismic-cycle services and without the catalogue — ~40 catalogue-derived features silently collapsed to neutral defaults in production, so the model never saw the signals it was trained on | The full service stack and catalogue are wired into forecast generation; any service failure is logged as a critical monitoring event instead of degrading silently. Regional and on-demand trainers now use the same production feature pipeline as the global trainer |
| Skill baseline | Performance reported against a random/null model (AUC > 0.5) | A temporal ETAS model (Ogata 1988) is fitted to the catalogue by maximum likelihood (μ, K, c, p, α; multi-start Nelder-Mead). Backtests now report ETAS AUC and the ML model's per-event information gain over ETAS — the CSEP-community standard. A model that does not beat ETAS adds no information, whatever its standalone AUC |
| Forecast target | One point per region (the region centroid) | A 5×5 grid (0.75° spacing, ~330 km span) is scanned per region and band; the highest-risk cell becomes the forecast target, and the full risk grid is published for mapping |
| M7+ forecasting | ML classifier trained on 2–5 positive examples (statistically meaningless) | Physics-based time-dependent renewal model: Brownian Passage Time (Matthews et al. 2002; WGCEP practice) over per-fault recurrence intervals and elapsed time, with aperiodicity α=0.5. Used as the primary M7+ risk when the ML model is data-starved, blended (geometric mean) otherwise |
| Winner selection | Single algorithm-competition winner, even when the runner-up was statistically indistinguishable | If the top two candidates are within 0.01 CV-AUC, their calibrated probabilities are soft-blended — lower variance, same ranking quality |
| Uncertainty & diagnostics | Single point-estimate AUC; no feature-level diagnostics | Bootstrap 90% confidence interval on test AUC; automatic feature-importance audit (dominance > 0.5 flags possible leakage, zero-importance features are listed as pruning candidates) |
| Placeholder features | depth_km_norm and geodetic_locking_proxy were hard-coded
constants carrying zero information |
depth_km_norm from the real mean depth of regional seismicity;
geodetic_locking_proxy from fault slip rate × elapsed accumulation |
We compare two backtest methodologies: the Global Model (v1, trivial negatives from tectonically quiet zones) and the Regional Model (v2, hard negatives from the same seismic region). The regional approach forces the model to learn temporal prediction rather than geographic classification.
Expanding window 1996-2026. Negatives from stable zones (Central Asia, Sahara, etc.).
| Band | AUC | F1 | Assessment |
|---|---|---|---|
| M4-5 | 0.750 | 0.519 | Realistic |
| M5-6 | 0.948 | 0.716 | Inflated |
| M6-7 | 0.992 | 0.889 | Inflated |
| M7+ | 0.494 | 0.333 | Insufficient data |
0 regions, 100km radius, expanding window 2000–2026. Hard negatives from same seismic region. All active regions use per-region grid-search tuned params. Auto-updated from model registry.
| Region | M4-5 AUC | Hit% | M4-5 Yrs | M5-6 AUC | M5-6 Yrs | M6-7 |
|---|---|---|---|---|---|---|
| Loading... | ||||||
Loading regional summary...
The v1 model's near-perfect M5-6 and M6-7 scores are caused by trivial geographic separation between positives (active fault zones) and negatives (stable continental interiors). The model learns "is this location on a fault?" rather than "will this fault rupture soon?".
| Feature | Importance | Type |
|---|---|---|
| cross_fault_velocity_mm_yr | 0.186 | Geographic |
| strike_sin | 0.155 | Geographic |
| b_value | 0.106 | Seismological |
| strike_cos | 0.093 | Geographic |
| distance_to_fault_norm | 0.086 | Geographic |
4 of the top 5 features are geographic/fault geometry features — confirming the model classifies location, not timing. The v2 regional model eliminates this bias by using negatives from the same fault system, forcing temporal discrimination.
| Year | Train | Test (pos/neg) | AUC | Acc | Hit Rate |
|---|---|---|---|---|---|
| 2008 | <2008 | 2 / 8 | 0.188 | 50.0% | 0.0% |
| 2010 | <2010 | 3 / 12 | 0.417 | 66.7% | 0.0% |
| 2011 | <2011 | 2 / 8 | 0.563 | 80.0% | 0.0% |
| 2013 | <2013 | 3 / 12 | 0.750 | 73.3% | 0.0% |
| 2016 | <2016 | 3 / 12 | 0.639 | 80.0% | 66.7% |
| 2018 | <2018 | 2 / 6 | 0.417 | 75.0% | 50.0% |
| 2019 | <2019 | 5 / 20 | 0.930 | 88.0% | 80.0% |
| 2020 | <2020 | 3 / 12 | 0.361 | 66.7% | 33.3% |
| 2025 | <2025 | 8 / 32 | 0.820 | 77.5% | 75.0% |
Overall: AUC=0.619 (11 testable years, expanding-window, hard same-region negatives). 2019 shows AUC=0.930 (pre-Marmara seismicity buildup); 2025 AUC=0.820 (recent activity spike). Istanbul has sparse M4-5 events (median 3/yr), making per-year estimates high-variance.
| Year | Pos/Neg | AUC | Hit% | Year | Pos/Neg | AUC | Hit% |
|---|---|---|---|---|---|---|---|
| 2000 | 3/12 | 0.444 | 0% | 2015 | 3/12 | 0.667 | 33% |
| 2002 | 2/8 | 0.063 | 0% | 2017 | 3/12 | 0.417 | 0% |
| 2003 | 2/8 | 0.438 | 0% | 2018 | 4/16 | 0.531 | 50% |
| 2007 | 4/16 | 0.719 | 75% | 2019 | 2/8 | 0.375 | 0% |
| 2011 | 2/8 | 0.125 | 0% | 2020 | 2/8 | 0.188 | 0% |
| 2012 | 9/36 | 0.710 | 44% | 2021 | 2/8 | 0.625 | 50% |
| 2013 | 6/24 | 0.868 | 33% | 2022 | 5/20 | 0.520 | 20% |
| 2014 | 2/8 | 0.500 | 50% | 2023 | 60/50 | 0.716 | 78% |
| 2024 | 17/50 | 0.734 | 71% | ||||
Maraş 2023–2024 (post-Kahramanmaraş M7.8 sequence): AUC=0.716/0.734 — the model correctly identified elevated risk in the aftershock-rich period. 2013 shows AUC=0.868 (Doğanyol aftershock sequence). Strategy: neg_ratio=2, temporal_ratio=0.6 (grid-search tuned). Overall AUC=0.796 over 19 test years.
How does Talivio's regional model compare to published prospective earthquake forecast benchmarks tested through the Collaboratory for the Study of Earthquake Predictability (CSEP)? Below we compare against the best RELM (Regional Earthquake Likelihood Models) results from the California testing center and ETH Zurich's Switzerland testing center.
| Model / Study | Region | Best AUC | Area Skill Score | Negative Strategy | Validation |
|---|---|---|---|---|---|
| FCN Deep Learning (GJI 2024) | California | — | 0.882 | Same-region | Pseudo-prospective |
| ETAS Italy / OEF-Italy (Taroni 2023) | Italy | — | 0.70 | Same-region | Prospective 10yr |
| STEP (Gerstenberger et al. 2005) | California | ~0.65 | — | Random cells | Prospective (RELM) |
| ETAS (Helmstetter et al. 2007) | California | ~0.68 | — | Seismicity rate | Prospective (CSEP) |
| EEPAS (Rhoades & Evison 2004) | New Zealand | ~0.70 | — | Precursory swarms | Prospective (CSEP) |
| DeVries et al. 2018 (Google Brain) | Japan | 0.849 | — (inflated) | Geographically distinct | Retrospective |
| Talivio (this work) | 0 regions | — | — | Hard same-region temporal + spatial |
Expanding-window no future leakage |
ASS = Area Skill Score (Molchan diagram diagonal area). FCN 0.882 and ETAS Italy 0.70 from published papers. Talivio ASS computed live from prospective forecasts — grows as more forecasts are verified.