- What is forecasting (in contrast to prediction)?
- What is probabilistic forecasting?
- Are earthquake forecasts currently being made?
- What methods are currently used for earthquake forecasting?
- How are probability forecasts typically validated in scientific and financial communities?
- How does OH validate its forecasts?
- Why did OH choose this approach to validate its forecasts?
- What magnitude earthquakes can OH forecast? What is currently available?
While the dictionary defines forecasting to be a synonym for prediction, we consider forecasting to be a specification of the odds, or probability, of an earthquake occurring at a given location, during a given time window, within a given magnitude range. By contrast, we consider a prediction to be the specification that an earthquake either will, or will not, occur at a given location, during a given time window, within a given magnitude range. A forecast is therefore a statement of probability, whereas a prediction is a binary statement. An individual forecast can never be validated by a single observation, but a forecast method can be validated by many observations. By contrast, an individual prediction can be validated by a single observation.
Probabilistic forecasting involves the computation of a spatial probability density function together with a temporal probability, leading to a conditional probability. The latter is the probability that an earthquake of a specified magnitude will occur, conditioned on the observation that no earthquake has occurred in the recent past. An example of a forecast probability statement might be that "there is a 40% probability that an earthquake having a magnitude between 6.5 and 7.0 will occur within a 20 km radius around location X during the next 3 months."
Yes. The official forecast for the state of California is a collaboration between the US Geological Survey, the California Geological Survey, and a large group of scientists from universities and commercial companies. These forecasts are used to set earthquake insurance rates in California.
The United States Geological Survey, through its Working Groups on California Earthquake Probabilities, has been developing long term earthquake forecasts for regions in California since 1988. These forecasts are based on data describing historic averages of major earthquakes as well as paleoseismic geologic data, obtained from trenching studies on active fault traces, and instrumental data. The result of these studies are 30 year probabilities for major earthquakes, typically having magnitudes M > 6.7, on major earthquake faults in California. The process by which these probabilities are computed is an extensive consultation and collaboration among over 100 scientists lasting several years. Expert opinion plays a significant role as well. An analysis of these forecasts was recently published in the scientific journal Nature ("Shaking Up Earthquake Theory, Nature, volume 46, pp. 870-872, 15 October, 2009).
Other forecasts include the Accelerated Moment Release (AMR) method, pioneered by scientists in the United States and Europe; the Epidemic Type Aftershock Sequence (ETAS) method developed by scientists in Japan, the United States and Europe; M8 and the Reverse Tracing of Precursors (RTP) method developed by scientists in Russia; the Psi method developed by scientists in New Zealand; the USGS STEP method for computing 24 hour probabilities of ground shaking due to aftershocks of major earthquakes; and a variety of methods based on earth strain and ground deformation measures.
The OH forecasts use methods that are based on ideas that have been in the peer reviewed literature for the past decade or more. These methods are data driven, using the ANSS catalog of earthquakes from online sources, together with well known observational laws including the Gutenberg-Richter relation and the Omori-Utsu aftershock frequency law. When the parameters in these laws are fit to the past observations, future probabilities can be computed. An example of a method of this type is the USGS STEP method, which is used to compute 24 hour aftershock probabilities. The advantage of the OH method is that it can be applied in a uniform way world-wide, and does not require local knowledge of the geology. Methods similar to this have been repeatedly tested over the past years, and continue to be tested by a variety of statistical techniques. The resulting forecasts may differ from the official forecasts as developed by the US Geological Survey, which uses the method as described above.
OH forecasts are fully time- and space-dependent, so calculated earthquake risk will change both as a result of changing earthquake probabilities, as well as changing human exposure. Even without the occurrence of a large earthquake and its aftershocks, earthquake probabilities can increase or decrease by as much as several per cent per month in active areas.
Forecasts are validated by a process called backtesting as well as a process called monitoring. In backtesting, data from the past are divided into a training period (prior data) and a testing period (posterior data). Forecasts are made using prior data to forecast events that occur during the posterior interval. The accuracy of these forecasts are then scored by a variety of statistical tests. Forecasts that achieve a pre-determined level of accuracy are considered to be validated at the observed confidence level. In monitoring, actual real time forecasts are computed, then actual events are observed. The results are scored using the same types of statistical analysis. Many researchers consider monitoring to be a higher level of validation than backtesting, since the “answer” is not known in advance. However, monitoring can take many years to determine the accuracy of a forecast method, whereas backtesting typically leads to an answer within days, weeks, or a few months at most.
Open Hazards validates its forecasts using the same types of statistical testing that are used in the weather/climate/financial forecasting communities. These tests are used to determine resolution, the ability of a forecast to discriminate between alternative outcomes; reliability, whether the predicted frequency of events matches the observed frequency of events; and sharpness, whether events tend to occur at high forecast probabilities, and no events tend to occur at low forecast probabilities, in contrast to methods in which events tend to occur near average values of probability.
Open Hazards feels that it is best to use testing procedures that have become standardized by extensive use in other fields, rather than inventing new statistical tests whose uses and properties are not well understood.
The Open Hazards method generally computes probabilities for the occurrence of earthquakes in space and time, and for magnitudes typically larger than M > 5.0. In addition, the methods can be more specifically applied to compute probabilities for other magnitude ranges. For example Open Hazards methods have been applied to compute the probabilities for great earthquakes having magnitudes M > 8.0 in regions such as Indonesia where these occur.
