Honest Limits

Honest Limits

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.

This page is the epistemic backbone of CAOS_SEISMIC. It states plainly what this system cannot do, what it can do, and why the difference matters more than any technical detail. Honesty here is not a disclaimer bolted on at the end — it is a design constraint that shapes every number the product publishes and every pixel of how it is shown. The cautionary history of the field (above all, L'Aquila) shows that the failure which destroys public trust and exposes people to harm is the miscommunication of probability, not the absence of a crystal ball.

Every claim below is anchored to peer-reviewed literature with a DOI, and every equation is real.

---

Table of contents

1. Why deterministic prediction is impossible
2. Prediction vs. forecast — the definitional split
3. Absolute probabilities stay small
4. A single outcome neither validates nor invalidates a forecast (Ridgecrest 2019)
5. The L'Aquila communication lesson
6. Operational forecasting is real, and honest (USGS OAF, OEF-Italy)
7. What this product commits to
8. References

---

1. Why deterministic prediction is impossible

Deterministic prediction — naming the time, place, and magnitude of a specific future earthquake with enough precision and reliability to justify an alarm — has been pursued for over a century with no reproducible success, and there are strong theoretical reasons to believe it is effectively impossible with foreseeable observational means.

1.1 The canonical statement (Geller et al. 1997)

The most-cited rejection of deterministic prediction is the Science perspective "Earthquakes Cannot Be Predicted" (Geller, Jackson, Kagan & Mulargia 1997). Its core argument is physical: faulting is a strongly nonlinear process, and whether a small rupture cascades into a large event depends on unmeasurably fine details of conditions in the Earth's interior. Any small earthquake therefore has some chance of growing into a large one, and which one does is set by sub-observational details. Reliable, deterministic alarms of imminent large earthquakes are, on this reasoning, effectively impossible. A companion review (Geller 1997) develops the case that proposed precursors are not reliably diagnostic and that the crust sits in a near-critical state.

An important nuance, stated honestly. Geller's "cannot be predicted" is a strong claim about deterministic prediction with current or foreseeable means. It is not a claim that all forecasting is futile — weak, probabilistic skill demonstrably exists, and that is exactly what Operational Earthquake Forecasting formalizes. This product cites Geller as the reason it does not do deterministic prediction, never as a claim that probability forecasting is pointless. Conflating those two claims would be dishonest in the opposite direction.

1.2 The physical grounding: self-organized criticality

The mechanistic version of Geller's argument is self-organized criticality (SOC). A slowly driven, spatially extended system with many degrees of freedom — the loaded crust — evolves spontaneously toward a stationary critical state with no characteristic event size, in which fluctuations are scale-invariant and appear as power-law statistics (Bak, Tang & Wiesenfeld 1987; Bak & Tang 1989). The Gutenberg–Richter frequency–magnitude law,

$$\log_{10} N(\ge M) = a - b,M,$$

is read as the signature of this critical state. In a critical system the eventual size of a rupture is not determined by observable large-scale precursors; it is decided by microscopic details as the rupture propagates. The system sits perpetually "on the edge," so a small slip and a great earthquake begin indistinguishably.

SOC is the leading explanatory framework, not settled physics. Characteristic-earthquake and partial-predictability views coexist in the literature. This product presents SOC as the leading framework, not as proven ground truth.

1.3 The field experiment that taught the lesson: Parkfield

The Parkfield prediction is the empirical lesson. A physically motivated recurrence model forecast that an ~$M6$ would rupture the same San Andreas segment within a few years of the late 1980s, publicly framed as a high-probability call by the early 1990s. 1993 came and went with no earthquake. The event finally occurred on 28 September 2004 ($M_w$ 6.0) — right place, right size, roughly 11 years late, and crucially with no diagnostic short-term precursors despite the most heavily instrumented fault segment on Earth (Bakun et al. 2005). Even with a good model and dense instrumentation, the timing was not predictable and no precursor reliably announced the event.

---

2. Prediction vs. forecast — the definitional split

The whole field organizes around one distinction, with the cleanest definitions from the ICEF report (Jordan et al. 2011):

Prediction = "a deterministic statement that a future earthquake will or will not occur in a particular geographic region, time window, and magnitude range" (a yes/no).

Forecast = "a probability (greater than zero but less than one) that such an event will occur" in that region/window/magnitude range.

	Prediction	Forecast
Output	Binary: will / will not occur	Probability strictly in $(0, 1)$
Decision posture	Alarm / no-alarm	Risk-weighted, graded
Falsifiability of a single call	Yes — it happened or it didn't	Only as a set over many trials (calibration)
Status in seismology	Not achievable reliably	Achievable, with measurable skill
This product	Never	This, and only this

Design consequence. The product never renders a single binary "earthquake yes/no" state, an alarm light, a countdown, or a single-event call. Every output is a probability with an uncertainty band, scoped to a region, a magnitude band, and a horizon (1 day / 2 days / 7 days). A forecast is evaluated as a set over many days (calibration), not by whether one event happened.

---

3. Absolute probabilities stay small

The single most important honesty constraint comes verbatim from the ICEF report: short-term earthquake probabilities "may vary over orders of magnitude but typically remain low in an absolute sense (< 1 % per day). Translating such low-probability forecasts into effective decision-making is a difficult challenge."

Both halves of that sentence are true at once, and both must be visible:

• The relative gain is real and large. During an active sequence, short-term models can raise the conditional probability by one to three orders of magnitude above the long-term background (a foreshock/aftershock "probability gain").
• The absolute number stays small. Even when "elevated," the probability of a large event in the next day is usually well under a few percent.

A useful way to hold both truths together:

We can multiply your sense of when risk is temporarily higher by 10–1000×, but even 1000× a tiny number is usually still a small number. We tell you that number honestly, with its uncertainty — we never tell you an earthquake is coming.

Design consequence. Every published number is shown next to its long-term baseline — the absolute probability and the ratio to background — so that "higher than usual" and "still small" are visible simultaneously. An unanchored figure is never shown. This directly answers both the ICEF $&lt; 1%/\text{day}$ reality and the over-reassurance failure of §5.

---

4. A single outcome neither validates nor invalidates a forecast (Ridgecrest 2019)

A probabilistic forecast cannot be judged "right" or "wrong" by one outcome. A well-calibrated $3%$ forecast should be realized about $3%$ of the time across many such situations; the other $97%$ of the time, nothing of that size happens — and both are consistent with the forecast being correct.

The cleanest worked example is the 2019 Ridgecrest sequence. After the 4 July 2019 $M$ 6.4 Searles Valley event, a fault-based ETAS system (UCERF3-ETAS) ran within roughly half an hour and gave about a 3 % chance of a larger event within the first week. Roughly 34 hours later, the $M$ 7.1 Ridgecrest mainshock struck (Savran et al. 2020).

A 3 % forecast is not "wrong" when the 3 % outcome occurs. $3% \ne 0%$. Low-probability events happen; that is what "3 %" means. Evaluating a single outcome as proof that "the forecast failed" — or, symmetrically, as proof that the model "predicted" the earthquake — is exactly the error the field warns against. Skill is calibration over many trials, not a single hit or miss.

Design consequence. The product publishes a live reliability diagram ("when we said 5 %, it happened ~5 % of the time"), logs every forecast immutably at issue time, and scores prospectively. The Ridgecrest example is carried into product copy as the canonical illustration. Cherry-picking a single hit or miss as proof of skill or failure is treated as a category error, in copy and in design.

---

5. The L'Aquila communication lesson

The field's defining cautionary tale is not a failed prediction — it is a communication failure.

On 6 April 2009, an $M_w$ ~6.3 earthquake struck L'Aquila, Italy, killing more than 300 people. A months-long seismic swarm had preceded it. On 31 March 2009 a national risk commission met to assess the swarm; afterward, public officials reassured the population that a major shock was unlikely, with a widely reported framing implying the swarm was discharging energy and reducing risk — which is not how seismology works (swarms modestly elevate short-term probability, but the absolute probability stays low). Residents who would otherwise have slept outdoors are alleged to have stayed inside on the strength of that reassurance.

In 2011, seven people were charged with manslaughter; in 2012 all were convicted, provoking an international scientific outcry. On appeal in 2014, the six scientists were acquitted; only the official who had made the most categorical public reassurance remained convicted (sentence reduced). The case was ultimately settled by Italy's supreme court.

The verdict has been widely framed as "a judgment not against science, but against a failure of science communication." The harm flowed from a false reassurance — treating a low-but-elevated probability as "no risk" — not from any failure to predict.

The field's institutional response was not to stop forecasting. It was to create the ICEF and push toward standardized, authoritative, honestly communicated Operational Earthquake Forecasting (§6).

The two-sided rule this product encodes. There are two symmetric communication failures, and both are dishonest:

• Over-reassurance — "no risk" / "safe" / "the swarm released the energy." (The L'Aquila failure.)
• Over-alarm — implying imminent certainty, an alarm, a countdown.

The antidote, baked into the product, is to always show the number (the probability) with its baseline (the background rate) and its uncertainty, so that "elevated" and "still low" are visible together — never an alarm, and never a promise of safety.

---

6. Operational forecasting is real, and honest (USGS OAF, OEF-Italy)

None of the above means forecasting is futile. Operational Earthquake Forecasting (OEF) — the authoritative dissemination of time-dependent, regularly updated probabilistic forecasts — runs today as real, scheduled public services, and they are honest about their limits.

• USGS Operational Aftershock Forecasts (OAF). Methodology lineage Reasenberg & Jones (1989) → Page et al. (2016), which added global generic parameters by tectonic regime and modeled intersequence variability so that forecasts carry uncertainty bounds, not point estimates. The public output gives, over defined windows, expected numbers of felt aftershocks and probabilities of damaging ones — probabilistic, updated as data arrive, and explicit about what it does not do.

• OEF-Italy (INGV). The first system fully complying with ICEF requirements, it produces gridded forecasts over all of Italy updated every midnight and after each $M_L \ge 3.5$ event, combining an ensemble of three clustering models (two ETAS flavors plus a STEP model). Its 10-year prospective validation (Spassiani, Falcone, Murru & Marzocchi 2023) found it broadly reliable — and, honestly, documented a systematic underestimation during the 2016–2017 Central Italy sequence caused by post-mainshock catalog incompleteness. That incompleteness is a known limit: right after a large mainshock — exactly the highest-stakes window — small events are buried in the coda, the magnitude of completeness $M_c$ spikes for hours to days, and a naive model under-forecasts productivity precisely when a large aftershock is most likely.

The forecast a non-homogeneous Poisson model publishes is the familiar exceedance probability,

$$P(\ge 1 \text{ event} \ge M^_) = 1 - e^{-N_{\ge M^_}},$$

where $N_{\ge M^*}$ is the expected count above the target magnitude over the horizon. This product emits the same form, scored under the same community framework (CSEP / pyCSEP).

This product's relationship to official OEF. CAOS_SEISMIC is an independent research and education tool that complements official agencies (USGS, INGV, CSN, GeoNet, JMA) — it is not an authoritative civil-protection alarm, and users should defer to official agencies for action.

---

7. What this product commits to

The honest framing is implemented, not merely promised. Concretely:

The product does, and says:

1. It produces conditional probabilities, not predictions — a probability in $(0, 1)$ for an event in a magnitude band, in a region, over the next 1 / 2 / 7 days.
2. It shows the number next to its long-term baseline, so "higher than usual" and "still small" are visible at once.
3. It shows probabilities changing daily as new data arrive, with the trajectory visible.
4. It states real uncertainty — an optimistic (P10) / expected / pessimistic (P90) triad, over-dispersion-aware (wider than a naive Poisson interval), with a staleness banner and a coverage mask.
5. It is evaluated prospectively — a public reliability diagram, CSEP-style scoring, the Ridgecrest lesson explained.
6. It complements, never replaces, long-term preparedness, engineering, and official agencies.

The product never does, and never renders:

1. A deterministic call — "an earthquake will strike at X on day Y." (Geller; Parkfield.)
2. An alarm / red-light / countdown implying imminent certainty. (Red is reserved for model quality, never danger.)
3. A single-event yes/no. Forecasts are probabilities, evaluated as sets.
4. A false reassurance — "no risk," "safe," "the swarm released the energy." (The L'Aquila failure.)
5. A cherry-picked single hit or miss presented as proof of skill or failure.
6. An unanchored number, or an implication that the daily probability is high. It almost never is.

The creed, carried verbatim into the app: Earthquakes cannot be predicted, but their probability can be forecast — and we report that probability honestly, with its uncertainty, evaluated against reality, never as an alarm and never as a promise of safety.

---

8. References

Canonical, peer-reviewed sources. DOIs given where registered.

1. Bak, P., Tang, C. & Wiesenfeld, K. (1987). Self-organized criticality: An explanation of 1/f noise. Physical Review Letters 59(4), 381–384. doi:10.1103/PhysRevLett.59.381
2. Bak, P. & Tang, C. (1989). Earthquakes as a self-organized critical phenomenon. JGR 94(B11), 15635–15637. doi:10.1029/JB094iB11p15635
3. Bakun, W. H. et al. (2005). Implications for prediction and hazard assessment from the 2004 Parkfield earthquake. Nature 437, 969–974. doi:10.1038/nature04067
4. Geller, R. J. (1997). Earthquake prediction: a critical review. GJI 131(3), 425–450. doi:10.1111/j.1365-246X.1997.tb06588.x
5. Geller, R. J., Jackson, D. D., Kagan, Y. Y. & Mulargia, F. (1997). Earthquakes cannot be predicted. Science 275(5306), 1616–1617. doi:10.1126/science.275.5306.1616
6. Jordan, T. H. et al. (2011). Operational Earthquake Forecasting: State of Knowledge and Guidelines for Utilization (ICEF Report). Annals of Geophysics 54(4), 315–391. doi:10.4401/ag-5350
7. Page, M. T. et al. (2016). Three ingredients for improved global aftershock forecasts. BSSA 106(5), 2290–2301. doi:10.1785/0120160073
8. Reasenberg, P. A. & Jones, L. M. (1989). Earthquake hazard after a mainshock in California. Science 243(4895), 1173–1176. doi:10.1126/science.243.4895.1173
9. Savran, W. H. et al. (2020). Pseudoprospective evaluation of UCERF3-ETAS forecasts during the 2019 Ridgecrest sequence. BSSA 110(4), 1799–1817. doi:10.1785/0120200026
10. Spassiani, I., Falcone, G., Murru, M. & Marzocchi, W. (2023). Operational Earthquake Forecasting in Italy: validation after 10 yr of operativity. GJI 234(3), 2501–2518.

Collaboratory for the Study of Earthquake Predictability (CSEP): https://cseptesting.org. USGS Operational Aftershock Forecasting: https://earthquake.usgs.gov/data/oaf/. The 2009 L'Aquila earthquake and its trials are extensively documented in the contemporaneous scientific press (Nature Geoscience, Science/AAAS, Eos).

---

See also: Methodology History — how short-term forecasting methodology evolved and how this product's methodology was reasoned out, from Gutenberg–Richter through ETAS and CSEP to the gated, context-conditioned neural model.