Statistical forecast
Earthquake forecast
Experimental researchProbability that at least one earthquake above a given magnitude occurs in a region within a time window — computed from observed earthquake clustering and aftershock statistics. This forecasts rates. It does not predict individual earthquakes, and nothing can.
This is a forecast, not a prediction. The difference matters.
A prediction would name a date, a place and a magnitude. No method can do that, and a platform that claimed otherwise would be lying to you.
A forecast states how often earthquakes should be expected in a region over a window, given what has just happened there. That is a real, testable, useful quantity — it is what aftershock advisories are built on — and it is what this page computes.
A 30% chance of M6+ this month does not tell you an earthquake is coming. It tells you that in 100 months that looked like this one, roughly 30 produced such an earthquake and 70 did not. Never evacuate, or decline to, on the basis of a number on this page. Follow your national authority.
Region
Convergence of the Pacific and Philippine Sea plates beneath the Eurasian and North American plates. Among the densest seismic networks on Earth, so the catalogue is unusually complete.
ETAS model forecast
Issued 2026-10-27 15:39:32 UTC (26 min ago) · ETAS etas-mle-v1
| Probability of at least one | next 1 day | next 7 days | next 30 days |
|---|---|---|---|
| M5.0+ | 24% 95% CI 24% – 25% background 8.5% 3.14× background expected 0.28 | 78% 95% CI 77% – 79% background 46% 2.43× background expected 1.51 | 99% 95% CI 99% – >99% background 93% 1.95× background expected 5.22 |
| M6.0+ | 1.5% 95% CI 1.4% – 1.7% background 0.49% 3.14× background expected 0.02 | 8.1% 95% CI 7.4% – 8.7% background 3.4% 2.43× background expected 0.08 | 25% 95% CI 23% – 27% background 14% 1.95× background expected 0.29 |
| M7.0+ | <0.1% 95% CI <0.1% – <0.1% background <0.1% 3.14× background expected 0.00 | 0.47% 95% CI 0.40% – 0.54% background 0.19% 2.43× background expected 0.00 | 1.6% 95% CI 1.4% – 1.8% background 0.82% 1.95× background expected 0.02 |
- μ background/day
- 0.378
- A productivity
- 0.0426
- α magnitude scaling
- 1.44
- c offset (days)
- 0.0249
- p Omori decay
- 1.069
- n̄ branching ratio
- 0.81
Source: USGS ComCat; EMSC/SeismicPortal; GFZ GEOFON · Observed: 2026-10-27 15:30:28 UTC · Retrieved: 2026-10-27 15:39:32 UTC · 135,494 events in the catalogue snapshot; this forecast is stored and will be scored once its window closes
What this means
Generated from the stored model output — every figure below traces back to it
Observed
Prakruti holds 10,864 earthquakes of magnitude 4.5 or above for Japan and the Ryukyu arc, spanning about 15.7 years of observation.
The largest earthquake recorded there in the past year was magnitude 7.6, about 323 days ago.
The magnitude-frequency relation fitted to this region has a b-value of 1.26, meaning each unit of magnitude is roughly 18.0 times less frequent than the one below it.
Model output
ETAS (etas-mle-v1) estimates a 78% chance of at least one earthquake of magnitude 5.0 or above in Japan and the Ryukyu arc during the next 7 days.
That corresponds to an expected 1.51 such earthquakes: 0.62 from the region's long-run background rate and 0.89 from aftershock sequences of earthquakes that have already happened.
Allowing for uncertainty in the fitted magnitude-frequency relation, the probability lies between 77% and 79%.
How to read it
The forecast sits above the region's background level of 46%, which is what an active aftershock sequence looks like. Sequences decay over days to months.
A probability of 78% is a statement about how often a situation like this one produces such an earthquake — not a statement that one is coming. It also says that in 22% of such situations, none does.
Limitations
Backtested over 246 forecasts between 2022-01-02 and 2026-09-21, this configuration performs better than a constant-rate baseline: Brier score 0.1903 against 0.2408 for the baseline.
The fitted branching ratio is 0.81 — the mean number of directly triggered earthquakes per earthquake. A value at or above one would mean the fit is not physically plausible.
The model forecasts rates over a region. It cannot indicate which fault, which day, or what size, and most of the probability is carried by earthquakes near the magnitude threshold rather than well above it.
This explanation is generated from the structured model result, not written by a language model. Every numeric token it contains is checked against that result before it is shown — 29 figures were verified. No language model is used anywhere in the numerical path. See how the model actually performs →
EXPERIMENTAL RESEARCH — NOT A VERIFIED EARTHQUAKE PREDICTION OR OFFICIAL WARNING.
ETAS forecasts earthquake rates. It does not predict individual earthquakes, it is not an official warning, and no agency has reviewed it. Its measured performance — including where it fails to beat a constant-rate baseline — is published on the model performance page.
Reference forecast — Gutenberg-Richter with generic aftershock parameters
Computed 2026-10-27 16:06:08 UTC from 10,893 events of M4.5+ held for this region
| Probability of at least one | next 1 day | next 7 days | next 30 days |
|---|---|---|---|
| M5.0+ | 50% 95% CI 44% – 57% background 36% 1.55× background rate | 99% 95% CI 97% – 99% background 96% 1.37× background rate | >99% 95% CI >99% – >99% background >99% 1.24× background rate |
| M6.0+ | 4.7% 95% CI 4.1% – 5.7% background 2.4% 1.97× background rate | 25% 95% CI 21% – 30% background 16% 1.64× background rate | 64% 95% CI 56% – 74% background 52% 1.39× background rate |
| M7.0+ | 0.17% 95% CI 0.12% – 0.24% background 0.14% 1.23× background rate | 1.2% 95% CI 0.86% – 1.6% background 0.96% 1.23× background rate | 4.9% 95% CI 3.6% – 6.8% background 4.0% 1.22× background rate |
Read each cell against the background figure beneath it. The background is the long-run rate for this region; the forecast is that rate plus the decaying contribution of earthquakes that have already happened. Right now this region is running at 1.37× its background rate for M5.0+ over seven days — mildly elevated by the tail of earlier sequences, and well within the region’s normal range.
What is driving this forecast
Earthquakes contributing most to the elevation
For the M5.0+ within 7 days forecast
- Bonin Islands, Japan regionM6.53 d ago+0.5652 expected
- 102 km ENE of Miyako, JapanM7.4190 d ago+0.0774 expected
- 2025 Aomori Prefecture, Japan EarthquakeM7.6323 d ago+0.0670 expected
- Bonin Islands, Japan regionM5.32 d ago+0.0571 expected
- The 2026 Kumamoto Region, Japan EarthquakeM6.891 d ago+0.0476 expected
Every proposed input, measured
All twenty-four variables, computed from real sources at forecast time, each with its actual role stated
Seismicity
- Earthquake clusteringDrives the forecast10,893 events in the catalogue window
Earthquakes cluster in space and time. This is the single most robust short-term signal there is, and it is what the forecast is built on.
Source: Prakruti reconciled catalogue
- Aftershock sequencesDrives the forecastlargest recent event M7.6, 323 days ago
Every recent event contributes a decaying aftershock rate through the Omori-Utsu law. This is the dominant term whenever a large earthquake has occurred recently.
Source: Prakruti reconciled catalogue
- Foreshock-like activityDrives the forecastcurrent rate is 0.55× the regional baseline
A sustained rate increase raises the modelled rate through the same clustering term. It is not a reliable warning: most rate increases are not followed by a larger earthquake, and most large earthquakes are not preceded by one.
Source: Prakruti reconciled catalogue
- Regional seismicity rateDrives the forecast1.885 events/day at or above the completeness magnitude
The long-run background rate. It sets the floor that the forecast decays back towards once a sequence dies away.
Source: Prakruti reconciled catalogue
- Gutenberg-Richter b-valueSets model parameters1.255
The slope of the magnitude-frequency relation, fitted for this region. It governs how the rate at one magnitude extrapolates to another, and its uncertainty is what the forecast interval is built from.
Source: Fitted from the regional catalogue (Aki maximum likelihood)
- Tectonic and geographic settingSets model parametersConvergence of the Pacific and Philippine Sea plates beneath the Eurasian and North American plates. Among the densest seismic networks on Earth, so the catalogue is unusually complete.
Setting is captured quantitatively through the region-specific b-value and background rate, both fitted only from events inside this region. The description is the qualitative half.
Source: Region definition
- Seismic gapShown for context, not used0.5 years since the last M7+ in this region
Shown because it is asked about, and excluded from the model on purpose. The seismic gap hypothesis — that a segment quiet for longer is more due — has repeatedly failed forecasting tests, and treating elapsed time as a countdown is one of the more dangerous intuitions in this subject.
Source: Prakruti reconciled catalogue
- Deepest recent hypocentreShown for context, not used522 km
Indicates whether the region is producing slab seismicity as well as shallow events. Not a forecast input.
Source: Prakruti reconciled catalogue
Sun, Moon and planets
- Earth tides (relative lunar tidal potential)Tested — no signal found1.1664
The solid Earth really is stressed by lunar and solar tides, by a few kilopascals — thousands of times less than an earthquake's stress drop. A weak tidal modulation has been reported in some specific settings, but none was detectable in this catalogue.
Source: Computed (VSOP87 / lunar theory)
- Lunar phaseTested — no signal found200.3° (0 = new, 180 = full)
Measured and displayed, but excluded from the forecast. Prakruti tested this against its own declustered catalogue and found no detectable association; see the research lab for the numbers.
Source: Computed
- Lunar distanceTested — no signal found365,176 km
Measured and displayed, but excluded from the forecast. Prakruti tested this against its own declustered catalogue and found no detectable association; see the research lab for the numbers.
Source: Computed
- Solar and lunar eclipse timingTested — no signal foundnext solar 2027-02-06, next lunar 2027-02-20
An eclipse is a particular Sun-Earth-Moon alignment, so it coincides with a syzygy tide that is already covered by the lunar phase test — which found nothing. An eclipse adds no force of its own.
Source: Computed
- Lunar declinationShown for context, not used23.68°
Sets which latitudes see the largest tidal range. Not tested separately here, and no established association with earthquake timing.
Source: Computed
- Lunar position (RA / declination)Shown for context, not usedRA 3.36 h, dec 23.68°
Geocentric apparent position. Shown for completeness.
Source: Computed
- Solar position (RA / declination)Shown for context, not usedRA 14.11 h, dec -12.81°
The solar tide is about 45% of the lunar one and moves with the seasons. Not a forecast input.
Source: Computed
- Planetary positionsShown for context, not usedMercury 230°, Venus 209°, Mars 136°, Jupiter 144°, Saturn 10°, Uranus 65°, Neptune 2°
Geocentric ecliptic longitudes. Excluded on physical grounds before any statistics: the tide-raising force a planet exerts on Earth is smaller than the Moon's by many orders of magnitude — Jupiter at its closest is weaker than the Moon by a factor of order ten million.
Source: Computed
- Closest planetary angular separationShown for context, not usedMars-Jupiter, 7.36°
An "alignment" changes nothing about the forces acting on Earth: the planets are still at the same distances, and those forces are negligible either way.
Source: Computed
Space weather
- Solar activity (sunspot number)Shown for context, not used60 (monthly mean, 2026-09)
No established physical pathway to tectonic earthquakes. Solar activity acts on the magnetosphere and upper atmosphere; earthquakes originate kilometres underground under stresses built up over years.
Source: NOAA SWPC
- Solar flaresShown for context, not usedcurrent X-ray class B3.0
Flare classes are logarithmic (A, B, C, M, X). Relevant to radio, satellites and power grids; not to faults.
Source: NOAA SWPC / GOES
- Geomagnetic index (planetary Kp)Shown for context, not used2.00
Claims of geomagnetic earthquake precursors recur regularly and have not survived testing. Included so the real number is visible rather than asserted.
Source: NOAA SWPC
Atmosphere
- Atmospheric pressureShown for context, not used874.8 hPa
Indicative reading at the region's representative point, not a regional average. Atmospheric pressure changes surface load by of order one kilopascal — far below the stresses that drive tectonic earthquakes, and excluded from the model.
Source: Open-Meteo
- Atmospheric temperatureShown for context, not used4.7 °C
Indicative reading at the region's representative point, not a regional average. Atmospheric pressure changes surface load by of order one kilopascal — far below the stresses that drive tectonic earthquakes, and excluded from the model.
Source: Open-Meteo
- PrecipitationShown for context, not used0.0 mm
Sustained hydrological loading has been linked to seismicity in a few specific settings — reservoir impoundment most clearly. A single current reading carries none of that information and is not a model input.
Source: Open-Meteo
Storms
- Cyclone and typhoon activityShown for context, not usedno active system within 2,000 km
A cyclone's low pressure does unload the surface, and a link to slow-slip events has been proposed in a few settings. Nothing established for ordinary tectonic earthquakes, and Prakruti only ingests NHC basins, so absence here does not mean absence globally.
Source: NOAA NHC via Prakruti
- Recent storm historyShown for context, not used0 active system(s) currently tracked near this region
Historical storm tracks (IBTrACS) are not yet ingested, so this reflects only currently active systems in the basins Prakruti covers.
Source: NOAA NHC via Prakruti
Earth rotation
- Length-of-day variationShown for context, not used1.5272 ms (2026-10-25)
Excess length of day. A decadal correlation with global M7+ rates has been proposed and is disputed; it concerns multi-year global rates, not where or when a particular earthquake will happen, so it cannot inform a regional short-term forecast.
Source: IERS finals2000A daily solution
- Polar motionShown for context, not usedx 0.1471″, y 0.3180″
Wobble of the rotation axis relative to the crust. No established association with earthquake occurrence.
Source: IERS finals2000A daily solution
Method, in full
1. Fit the magnitude-frequency relation. Gutenberg-Richter says the number of earthquakes falls off logarithmically with magnitude: roughly ten times fewer for each unit. The slope, b, is fitted for this region alone by Aki maximum likelihood with a correction for magnitude binning.
2. Take the background rate. The fitted relation is extrapolated to each threshold, giving a long-run Poisson rate.
3. Add the aftershock contribution. Every earthquake in this region over the past year contributes a decaying rate of its own, following the Omori-Utsu law in the Reasenberg-Jones formulation (p = 1.08, c = 0.05 days, a = -1.67, b = 0.91). Productivity scales steeply with mainshock magnitude and decays slightly faster than 1/t.
4. Convert to a probability. Rates add, so the total expected count is the background plus every sequence. The chance of at least one event is then 1 − e−N.
5. Compare against the baseline. Every cell shows the background-only probability beside the forecast. Without it a probability is unreadable.
Limitations of this forecast
It cannot tell you where, or when, or how big. A regional rate says nothing about which fault, which day, or what size. Most of the probability in any cell is carried by events near the threshold, not by the large ones people worry about.
The aftershock parameters are generic. Reasenberg-Jones values derived for one region are being applied everywhere. Region-specific parameters perform better; deriving them needs a catalogue of studied sequences that Prakruti does not yet hold.
The interval understates the uncertainty. It propagates the b-value error only. Uncertainty in the aftershock parameters, in the magnitude of completeness, and in the region boundary is not included, so the true interval is wider than the one shown.
Mixed magnitude scales bias b. The catalogue mixes mb, mww, ml and others. Body-wave magnitudes saturate at the top of this range, which compresses the magnitude distribution and tends to push the fitted b upward. Values above about 1.2 here should be read with that in mind.
Regions are rectangles. Real tectonic boundaries are not. A box mixes settings with genuinely different behaviour, which blurs the fitted parameters.
This model has not been prospectively tested here. The underlying method has a track record in the literature and its outputs here are consistent with observed regional rates, but Prakruti has not yet run a formal prospective evaluation of its own forecasts against what subsequently happened. Until it has, treat these numbers as a well-founded estimate rather than a validated one.
EXPERIMENTAL RESEARCH — NOT A VERIFIED EARTHQUAKE PREDICTION OR OFFICIAL WARNING.
This forecast is a statistical estimate of earthquake rates. It is not an official warning, no agency has reviewed it, and it must never be used as the basis for an emergency decision. Read the full scientific limitations →
Source: Prakruti reconciled catalogue (USGS, EMSC, GFZ GEOFON) · NOAA SWPC · Open-Meteo · IERS · Retrieved: 2026-10-27 16:06:08 UTC · model: Gutenberg-Richter + Reasenberg-Jones / Omori-Utsu, generic parameters