Testing methods for alarmbased earthquake prediction strategies - PowerPoint PPT Presentation

Title: Testing methods for alarmbased earthquake prediction strategies


1
Testing methods for alarm-based earthquake
prediction strategies

• Jeremy Zechar
• University of Southern California

2
Probabilisticpredictions
Alarm-basedpredictions

• Forecast rate of target events or probability of
  occurrence
• Standard evaluation and comparison via likelihood
  methods

• Forecast occurrence or non-occurrence of target
  events
• Several evaluation approaches
• Few attempts at establishing statistical
  significance
• Can be derived from probabilistic predictions

3
Deriving alarms from probabilisticpredictions
Low threshold ????
Medium threshold ????
High threshold ????
(after Holliday et al 2005)
4
Assumptions (or at least what I have in mind)

• The hypotheses were interested in testing lead
  to forecasts of this nature we expect one or
  more target earthquakes (earthquakes with
  magnitude gt target magnitude) in a given space.
• If were deriving alarms from a probabilistic
  forecast, were interested in evaluating
  individual alarm sets and a range of derived
  alarm sets.
• For the examples in this presentation, alarms are
  fixed at the beginning of an experiment and do
  not expire until the end of the experiment.

5
Binary prediction, binary outcome
Space
Time
6
In other words

• Having given the number of instances
  respectively in which things are both thus and
  so, in which they are thus but not so, in which
  they are so but not thus, and in which they are
  neither thus nor so, it is required to determine
  the special quantitative relativity subsisting
  between the thusness and the soness of the
  things.
• M.H. Doolittle, 1888

7
Contingency table (Finley 1884)
8
Scalar performance measures

• Some measures permit an optimal score to be
  obtained by a simple strategy.
• e.g., if one never declares an alarm, false alarm
  rate will be optimized.
• Some measures rely on explicit specification of
  negative-alarms.
• Otherwise, tester must infer negative alarms in
  order to count correct negatives.
• Introduces subjectivity
• Some measures consider a Type I error to have
  equal importance of Type II error.
• e.g., Critical Success Index combines false
  alarms and misses.
• Is this desirable?
• Considered alone, none of these measures seem
  ideal.

Joliffe and Stephenson 2003
9
Receiver Operating Characteristic (ROC) a
splitters definition

• False alarm rate b/(bd)
• Hit rate a/(ac)
• One set of alarms corresponds to a single point
  on ROC.
• Area under curve (AUC) is a common performance
  measure.
• All hits are considered equally good, all false
  alarms are considered equally bad.

10
What is the prior probability of target
earthquake within a given alarm space-time region?

• Standard estimate is given by Poisson
• Here, r is average target earthquake occurrence
  rate within alarm region based on catalog of
  maximal length and t is duration of the alarm
  (Jackson 1996).

11
Estimation by simulation (Jackson 1996)

• Create random catalogs and count successful
  alarms (hits and correct negatives)
• Construct frequency distribution of successes and
  sum those with as many or more successes.
• Each success counts equally, regardless of prior
  probability.
• What abt misses?
• How do we determine the prior probability for a
  missed event?

Repeat many times
12
Construct a simple model

• If a simple strategy can produce an alarm set
  that obtains as many hits as a more complex
  strategy alarm set, the effectiveness of the
  complex model is questionable.
• E.g., VAN analyses (in particular, Kagan 1996)

13
Molchan diagram

• Miss rate c/(ac)
• Fraction of alarm space-time is the space-time
  occupied by alarms divided by the total amount of
  space-time during experiment.
• Measure of space used to compute t is important.
  Options
• Map area
• Seismic intensity- weighted area

1
1
1
tmap1/4
tint1/2
(after Molchan and Kagan 1992)
14
t increases, n decreases
15

• Alarm sets that are significantly better than
  random will yield points below the confidence
  steps (outside the shaded regions). These
  intervals are independent of the measure of space.

16
Retrospective testing CA portion of national
seismic hazard map

• Define target eqks (Mgt5 in ANSS catalog)
• Define study region (lat 32 to 38.3, lon -123 to
  -115) and grid spacing (0.1x 0.1)
• Define experiment duration (2000-2009 inclusive)
• Details
• 17 target events occurred in the study region
  between 2000 and 15 May 2006.
• Earthquake rate predictions per each box obtained
  from Frankel 2002 codes, converted to
  probabilities using Poisson assumption.

17
USING MAP AREA FOR t No surprise NSHMP is
significantly better at predicting earthquakes
than someone throwing darts at a map of
California.
18
USING INTENSITY-WEIGHED AREA FOR t NSHMP does not
regularly yield points outside the confidence
regions. What can we say?
19
Introducing another level of abstraction

• To determine skill of crossing paths, we can
  compute the area under the Molchan trajectory.
• This describes predictive significance of all
  alarms produced up to the point of interest,
  rather than a single set of alarms.

20
Properties of Molchan trajectory area measure

• Perfect prediction yields 0 area, anti-perfect
  prediction yields unit area.
• Expected value of area for unskilled predictions
  is ½.
• Preliminary simulations indicate that area
  distribution for unskilled predictions quickly
  approaches a normal distribution as N, the number
  of target earthquakes, increases.
• Preliminary results suggest an exact analytic
  form for variance of area of unskilled
  predictions as a function of N.
• Together, these properties allow intuitive
  hypothesis testing using my ASS
• Area Skill Score (ASS) 1 Molchan trajectory
  area measure.

21
USING MAP AREA FOR t NSHMP does extremely well,
as we would expect.
22
USING INTENSITY-WEIGHED AREA FOR t We cannot
reject the null hypothesis that the NSHMP
trajectory was obtained by using a prediction
method with no skill.
23
Conclusions

• Deriving target earthquake alarms might provide a
  common ground for comparing forecast/prediction
  algorithms.
• There are a number of tools and performance
  measures available for testing alarm-based
  predictions.
• The questions CSEP wants to answer should drive
  selection development of testing procedures
  performance measures rather than the other way
  around.
• We may need to tailor existing performance
  measures and tools.