Operational Earthquake Forecasting: Proposed Guidelines for Implementation Thomas H. Jordan Director, Southern California S33D-01, AGU Meeting 14 December 2010
Operational Earthquake Forecasting Authoritative information about the time dependence of seismic hazards to help communities prepare for potentially destructive earthquakes. Seismic hazard changes with time Earthquakes release energy and suddenly alter the tectonic forces that will eventually cause future earthquakes Statistical models of earthquake interactions can capture many of the short-term temporal and spatial features of natural seismicity Excitation of aftershocks and other seismic sequences Such models can use regional seismicity to estimate short-term changes in the probabilities of future earthquakes
Operational Earthquake Forecasting Authoritative information about the time dependence of seismic hazards to help communities prepare for potentially destructive earthquakes. What are the performance characteristics of current short-term forecasting methodologies? Probability (information) gain problem How should forecasting methods be qualified for operational use? Validation problem How should short-term forecasts be integrated with long-term forecasts? Consistency problem How should low-probability, short-term forecasts be used in decision-making related to civil protection? Valuation problem
Supporting Documents Operational Earthquake Forecasting: State of Knowledge and Guidelines for Implementation Final Report of the International Commission on Earthquake Forecasting for Civil Protection (T. H. Jordan, chair), Dipartimento della Protezione Civile, Rome, Italy, 79 pp., December, 2010. Operational Earthquake Forecasting: Some Thoughts on Why and How T. H. Jordan & L. M. Jones (2010), Seismol. Res. Lett., 81, 571-574, 2010.
Prediction vs. Forecasting Southern California An earthquake forecast gives a probability that a target event will occur within a space-time domain An earthquake prediction is a deterministic statement that a target event will occur within a space-time domain Rupture Probability on San Andreas System (WGCEP, 2007) RTP Alarm for California M 6.4, 15 Nov 2004-14 Aug 2005 (Keilis-Borok et al., 2004)
Prediction vs. Forecasting 1.0 Shannon Entropy 0.5 P < 0.2 P > 0.8 0 0 0.5 1.0 Earthquake Probability For operational purposes, deterministic prediction is only useful in a high-probability environment probabilistic forecasting can be useful in a low-probability environment
Operational Forecasting in California Organizations USGS - National Earthquake Prediction Evaluation Council (NEPEC) CalEMA - California Earthquake Prediction Evaluation Council (CEPEC) Operational forecasting tools Long-term models (WGCEP models; e.g., UCERF2) Short-term models (Reasenberg-Jones, STEP, ETAS, Agnew-Jones) Notification protocols Southern San Andreas Working Group (1991) California Integrated Seismic Network notifications For M 5 events, probability of M 5 aftershocks and expected number of M 3 aftershocks
Long-Term Earthquake Probability Models Uniform California Earthquake Rupture Forecast (UCERF2) 1680 1857 1906 UCERF2 30-yr Gain 5-yr Gain 0 1 2 3 Probability Gain UCERF2 1-day probability for M > 7 Coachella rupture: P = 3 x 10-5 UCERF2 ratio of time-dependent to time-independent participation probabilities for M 6.7 (WGCEP, 2007)
Southern California Short-Term Earthquake Probability (STEP) Model -0 + 11 13 hour month hours 2hour months Gerstenberger et al. (2005) Probability of Exceeding MMI VI http://earthquake.usgs.gov/earthquakes/step/ 2004 Parkfield Earthquake
Short-Term Earthquake Probability (STEP) Model Southern California + 1 hour STEP 1-day probability gain for MMI > VI shaking: G > 100 Gerstenberger et al. (2005) Probability of Exceeding MMI VI http://earthquake.usgs.gov/earthquakes/step/ 2004 Parkfield Earthquake
Summary of Probability Gains Method Gain Factor P max (3 day) SAF-Coachella Prospectively validated? Long-term renewal 1-2 1 x 10 4 No Medium-term seismicity patterns 2-4 2 x 10 4 Yes Short-term STEP/ETAS 10-100 3 x 10 3 Yes Short-term empirical foreshock probability 100-1000 3 x 10 2 No The probability gains of short-term, seismicity-based forecasts can be high, but the absolute probabilities of large earthquakes typically remain low, even in seismically active areas, such as California.
Operational Forecasting in California W. H. Bakun et al., Parkfield, California, Earthquake Prediction Scenarios and Response Plans. USGS OFR 87-192, 1987. Southern San Andreas Working Group, Short-Term Earthquake Hazard Assessment for the San Andreas Fault in Southern California, USGS OFR 91-32, 1991. * # instances many ~10 2
Operational Forecasting in California Earthquake forecasting in a low-probability environment is already operational in California, and the dissemination of forecasting products is becoming more automated Level-A probability threshold of 25% has never been reached Level-B threshold of 5-25% has been exceeded only twice (Joshua Tree and Parkfield) However, procedures are deficient in several respects: CEPEC has generally relied on generic short-term earthquake probabilities or ad hoc estimates calculated informally, rather than probabilities based on operationally qualified, regularly updated seismicity forecasting systems Procedures are unwieldy, requiring the scheduling of meetings or telecons, which lead to delayed and inconsistent alert actions How the alerts are used is quite variable, depending on decisions at different levels of government and among the public
The Validation Problem To be fit for operational purposes, short-term forecasting methods should demonstrate reliability and skill with respect to long-term (e.g., time-independent) models Forecasting methods intended for operational use should be scientifically tested against the appropriate data for reliability and skill, both retrospectively and prospectively All operational models should be under continuous prospective testing competing time-dependent models Collaboratory for the Study of Earthquake Predictability (CSEP) provides the infrastructure for this testing
CSEP Testing Regions & Testing Centers Southern California Western Pacific SCEC Testing Center Global EU Testing Center Zurich China Testing Center ERI Testing Center Los Angeles California Italy North-South Seismic Belt Beijing Japan Tokyo GNS Science Testing Center Testing Center Upcoming Testing Region Upcoming Wellington New Zealand
CSEP Evaluation of Short-Term Models in the California Testing Region (Rhoades T-test, M 3) Southern California STEP Model Reference forecast IG = 2.6, PG = 13.5/eqk IG = 0.3, PG = 1.35/eqk Information gain per earthquake
The Consistency Problem Spatiotemporal consistency is an important issue for dynamic risk management, which often involves trade-offs among multiple targets and time frames In lieu of physics-based forecasting, consistency must be statistically enforced a challenge because: Long-term renewal models are less clustered than Poisson Short-term triggering models are more clustered than Poisson Consistency can be lacking if long-term forecasts specify background seismicity rates for the short-term models (e.g., STEP) Seismicity fluctuations introduced by earthquake triggering can occur on time scales comparable to the recurrence intervals of the largest events Model development needs to be integrated across all time scales of forecast applicability Approach adopted by WGCEP for development of UCERF3
The Valuation Problem Earthquake forecasts acquire value through their ability to influence decisions made by users seeking to mitigate seismic risk and improve community resilience to earthquake disasters Societal value of seismic safety measures based on long-term forecasts has been repeatedly demonstrated Potential value of protective actions that might be prompted by short-term forecasts is far less clear Benefits and costs of preparedness actions in high-gain, lowprobability situations have not been systematically investigated Previous work on the public utility of short-term forecasts has anticipated that they would deliver high probabilities of large earthquakes (deterministic prediction)
The Valuation Problem Economic valuation is one basis for prioritizing how to allocate the limited resources available for short-term preparedness In a low-probability environment, only low-cost actions are justified Cost-Benefit Analysis for Binary Decision-Making (e.g., van Stiphout et al., 2010) Suppose cost of protection against loss L is C < L. If the short-term earthquake probability is P, the policy that minimizes the expected expense E: - Protect if P > C/L - Do not protect if P < C/L Then, E = min {C, PL}. Many factors complicate this rational approach Monetary valuation of life, historical structures, etc. is difficult Valuation must account for information available in the absence of forecast Official actions can incur intangible costs (e.g., loss of credibility) and benefits (e.g., gains in psychological preparedness and resilience)
Recommendations Utilization of earthquake forecasts for risk mitigation and earthquake preparedness should comprise two basic components Scientific advisories expressed in terms of probabilities of threatening events Protocols that establish how probabilities can be translated into mitigation actions and preparedness Public sources of information on short-term probabilities should be authoritative, scientific, open, and timely Authoritative forecasts, even when the absolute probability is low, can provide a psychological benefit to the public by filling information vacuums that can lead to informal predictions and misinformation Should continuously inform the public about the seismic situation, in accordance with socialscience principles for effective public communication of warnings Need to convey the epistemic uncertainties in the operational forecasts Alert procedures should be standardized to facilitate decisions at different levels of government and among the public, based in part on objective analysis of costs and benefits Should also account for less tangible aspects of value-of-information, such as gains in psychological preparedness and resilience
Conclusions Southern California Current short-term forecasting methodologies can provide nominal (unvalidated) probability gains up to 100-1000 Issue: unification of methodologies across temporal and spatial scales Operational forecasting procedures should be qualified for usage according to three standards for operational fitness Quality: correspondence between the forecasts and actual earthquake behavior Consistency: compatibility of methods at different spatial or temporal scales Value: realizable benefits relative to costs incurred All operational forecasting models should be under continuous prospective testing Issue: evaluation of operational forecasts in terms of ground motions Governments should develop and maintain an open source of authoritative, scientific information about the short-term probabilities of future earthquakes keep the population aware of the current state of hazard decrease the impact of ungrounded information improve preparedness Issue: decision-making in a low-probability environment
How Should the Time-Dependent Forecasts be Communicated to Decision-Makers? Here here or here? Earthquake Rupture Forecast Attenuation Relationship Shaking Intensity Loss P(S n ) P(IM k S n ) P(IM k ) P(L k IM k ) Hazard Risk Probabilistic Seismic Hazard and Risk Analysis
Southern California STEP Map for 2004 Parkfield Earthquake + 1 hour Probability of Exceeding MMI VI http://pasadena.wr.usgs.gov/step
CyberShake 1.0 Hazard Model (225 sites in Los Angeles region, f < 0.5 Hz) Southern California Uses an extended earthquake rupture forecast Source area probabilities Hypocenter distributions (conditional) Slip variations (conditional) Uses reciprocity to calculate ground motions for ~440,000 events at each site Psuedo-dynamic fault rupture 3D anelastic model of wave propagation CyberShake seismogram LA region CyberShake hazard map PoE = 2% in 50 yrs Graves et al. (2010)
CyberShake as a Platform for Short-Term Earthquake Forecasting Southern California (see poster by Milner et al., S51A-1926) Parkfield (M6.0) Sept 28, 2004 Parkfield, 2004 Los Angeles Bombay Beach, 2009 Bombay Beach (M4.8) Mar 24, 2009 Compute probability gain from Agnew & Jones (1991) model. Example: G = 1000 for R 10 km Apply probability gain to CyberShake ruptures and recompute ground motion probabilities for short interval following events. Example: 1 day
End Southern California