WEBVTT
Kind: captions
Language: en

00:00:00.380 --> 00:00:02.580
[Music]

00:00:03.360 --> 00:00:08.960
[Silence]

00:00:09.420 --> 00:00:12.300
Good morning, everyone.
So just a very quick announcement.

00:00:12.300 --> 00:00:16.120
So, this Wednesday, during the
usual 10:30 a.m. time slot, we’ll be

00:00:16.120 --> 00:00:20.420
hearing from our very own Ross Stein.
And, in typical Ross fashion, he’ll be

00:00:20.420 --> 00:00:23.640
doing something pretty creative.
He’ll have a mixture of kind of

00:00:23.640 --> 00:00:28.410
more typical slides in addition to
kind of practical demonstrations.

00:00:28.410 --> 00:00:32.661
And he’ll be trying to incorporate
those demonstrations in order to explore

00:00:32.661 --> 00:00:38.640
why earthquakes are sporadic, although
the stress states are more steady.

00:00:38.640 --> 00:00:40.800
So that’ll be
pretty interesting.

00:00:40.800 --> 00:00:44.220
So now I’ll hand it over to
Sara to introduce Matt.

00:00:45.860 --> 00:00:46.860
- Hi, everyone.

00:00:46.860 --> 00:00:49.390
My name is – for those people who
don’t know me, I’m Sara McBride.

00:00:49.390 --> 00:00:52.760
I’m a Mendenhall Fellow at
the Earthquake Science Center.

00:00:52.760 --> 00:00:54.690
And I just want to quickly
introduce Matt Gerstenberger.

00:00:54.690 --> 00:00:59.700
I have worked with Matt,
I think, since 2011, 2012.

00:00:59.700 --> 00:01:05.600
And so I can tell you about Matt’s –
a little bit about Matt’s background.

00:01:05.600 --> 00:01:10.780
Matt comes from the
highly seismic zone of Kansas,

00:01:10.780 --> 00:01:13.500
which then inspired his work.
I’m just kidding.

00:01:13.500 --> 00:01:18.250
Kansas now gets earthquakes,
but not when he was growing up.

00:01:18.250 --> 00:01:22.240
Just two quick stories about
Matt I would like to share.

00:01:22.240 --> 00:01:26.200
When there was the Cook Strait/Lake
Grassmere earthquake –

00:01:26.200 --> 00:01:30.380
earthquake sequence in 2013 –
it was a doublet in New Zealand –

00:01:30.380 --> 00:01:33.780
there was a little town called
Seddon that has experienced

00:01:33.780 --> 00:01:36.460
most of the earthquakes
and the damage there.

00:01:36.460 --> 00:01:39.900
And we sent Matt down to do
a community meeting to talk about

00:01:39.900 --> 00:01:42.470
the forecasts and what
he thought would happen

00:01:42.470 --> 00:01:45.110
and what kind of aftershocks
people could expect.

00:01:45.110 --> 00:01:51.470
And I think, at about minute 35
of Matt’s talk, a 4.8 magnitude

00:01:51.470 --> 00:01:56.159
earthquake struck just off the coast of
Seddon, and the entire building shook.

00:01:56.159 --> 00:02:00.420
And Matt, so calmly and coolly,
said, yes, that’s the kind of

00:02:00.420 --> 00:02:05.090
aftershocks I was talking about.
And it was a really good moment

00:02:05.090 --> 00:02:08.580
of Matt being very calm during
what was a very stressful event

00:02:08.580 --> 00:02:13.660
for the people of Seddon.
And it made everyone laugh and relax.

00:02:14.349 --> 00:02:20.779
So I think I won’t share the other story
maybe until [laughs] later, but just

00:02:20.779 --> 00:02:24.769
a quick word. I wouldn’t be here
with the USGS without Matt.

00:02:24.769 --> 00:02:28.959
Matt was a Mendenhall in 2004
and 2005, and he inspired me

00:02:28.960 --> 00:02:33.800
to apply for the Mendenhall, which
I felt completely unqualified for.

00:02:33.800 --> 00:02:36.260
And he said, no, no, no.
You might be unqualified,

00:02:36.260 --> 00:02:39.700
but give it a go anyway.
So thank you, Matt, for doing that.

00:02:39.700 --> 00:02:43.280
And without further ado,
Matt Gerstenberger.

00:02:45.400 --> 00:02:48.340
- Thank you very much, Sara.

00:02:48.340 --> 00:02:50.819
Good morning,
everyone.

00:02:50.819 --> 00:02:54.700
So I will – I was going to talk
about some of the work that

00:02:54.700 --> 00:02:58.719
we’ve been doing in the last few years
in terms of responding to earthquakes

00:02:58.719 --> 00:03:03.859
on the forecasting side.
I’ll be pretty operationally focused,

00:03:03.860 --> 00:03:07.960
but I can get into a bit more of
the details if anyone’s interested.

00:03:09.269 --> 00:03:11.599
So I'll just take you through kind of
what’s been going on for the last

00:03:11.599 --> 00:03:14.819
decade or so and then just pick out
a few examples from the bigger

00:03:14.819 --> 00:03:18.279
earthquakes that you’ll obviously
know about, and then just some of

00:03:18.279 --> 00:03:22.650
the different methods that we – that
we’ve put information out, how we’ve

00:03:22.650 --> 00:03:27.380
come up with that, and then how that’s
been used by the government and others.

00:03:27.380 --> 00:03:32.150
So, first, it’s been a fairly
busy last decade after the

00:03:32.150 --> 00:03:36.829
second half of last century being
very quiet with only really

00:03:36.829 --> 00:03:39.320
one significant earthquake
during that time.

00:03:39.320 --> 00:03:47.780
Starting in 2003, it’s become quite
active with a lot of work necessary to do.

00:03:47.780 --> 00:03:51.409
In the – I guess, prior to 2010,
in terms of forecasting, we were

00:03:51.409 --> 00:03:56.029
doing internal forecasting within GNS
and went out for some specific clients,

00:03:56.029 --> 00:04:00.329
but we weren’t really – hadn’t done
anything public-facing to that point.

00:04:00.329 --> 00:04:03.520
And that all changed
with the Darfield earthquake.

00:04:03.520 --> 00:04:07.440
And that’s when we really got into a lot
of the types of information that I’ll show.

00:04:07.459 --> 00:04:10.170
And we’ve had these – I don’t think
this is everything that we’ve

00:04:10.170 --> 00:04:13.700
had to respond to during that time,
but these are the earthquakes

00:04:13.700 --> 00:04:16.370
where we’ve kind of –
that we’ve used to refine

00:04:16.370 --> 00:04:20.680
all of our forecasting information
and methods that we’ve put out.

00:04:21.660 --> 00:04:23.660
Luckily, it’s been a bit
quiet the last two years.

00:04:23.660 --> 00:04:26.140
Hopefully,
it’ll stay that way.

00:04:27.300 --> 00:04:29.760
So, as I mentioned,
prior to 2010, we weren’t really –

00:04:29.760 --> 00:04:31.870
we weren’t doing this –
again, I’ve already said all this,

00:04:31.870 --> 00:04:37.550
so I won’t go into this slide in too details,
but I guess the key learning in all of this

00:04:37.550 --> 00:04:40.750
is the using of multiple products
and multiple types of information

00:04:40.750 --> 00:04:45.629
has really been demonstrated, I guess,
to us that that’s what’s necessary

00:04:45.629 --> 00:04:49.960
to do to get the information out
and to get people to understand it.

00:04:51.490 --> 00:04:56.620
This one is just a schematic of the –
kind of the overall types of

00:04:56.620 --> 00:04:59.930
information that we put out.
First is kind of the number 1 there,

00:04:59.930 --> 00:05:02.500
the hybrid combination of
statistical forecast models.

00:05:02.500 --> 00:05:05.300
That’s our core methods.
If you think of operational earthquake

00:05:05.300 --> 00:05:08.640
forecasting, I guess that would be –
that would be it there.

00:05:08.650 --> 00:05:13.009
It’s not automated at this point, but it’s
quantitative, statistically optimized.

00:05:13.009 --> 00:05:16.310
It’s quick. We don’t have to
interact with this – well, we don’t

00:05:16.310 --> 00:05:20.090
have to make any judgment calls.
Everything just happens.

00:05:20.090 --> 00:05:24.440
And then, all the other products
we do kind of stem out from that.

00:05:24.440 --> 00:05:27.850
I’ll show some examples of
kind of hazard- and risk-based

00:05:27.850 --> 00:05:30.410
information that we put out.
And the hazard- and risk-based

00:05:30.410 --> 00:05:34.639
information is really kind of the key
part of what we’re doing now, I think.

00:05:34.639 --> 00:05:39.759
And then, there’s scenarios.
It’s really a much more qualitative

00:05:39.759 --> 00:05:43.259
collection of information that requires
a team of people to put together

00:05:43.259 --> 00:05:48.370
and a lot of judgment, but that’s really
how we explain to both the public

00:05:48.370 --> 00:05:52.790
and government and others
what the particular OEF –

00:05:52.790 --> 00:05:54.389
the statistical
information might mean.

00:05:54.389 --> 00:05:58.340
What that – what that could –
how that could play out in the future.

00:06:00.540 --> 00:06:06.980
So the current situation with our
forecasting is it’s not – it’s not fully

00:06:06.980 --> 00:06:10.080
operational, meaning it’s not on its own.
We have to – we have to launch it.

00:06:10.080 --> 00:06:11.749
So I think we’re in
a similar method to things

00:06:11.749 --> 00:06:14.440
that have happened
in the past in the USGS.

00:06:14.440 --> 00:06:16.240
It’s not currently for
all of New Zealand.

00:06:16.240 --> 00:06:19.180
We have a system that can do this.
It’s just not online right now.

00:06:19.180 --> 00:06:22.900
And it requires somebody to
hit a button and tell it to go.

00:06:22.900 --> 00:06:26.270
The scenarios – the more
descriptive methods – they take

00:06:26.270 --> 00:06:29.900
a fair bit of interaction. Those we
don’t usually get out within a day.

00:06:29.900 --> 00:06:32.240
It takes a lot of time and
understanding what’s in the earthquake,

00:06:32.240 --> 00:06:35.240
understanding what analogs have
happened in the past in that region

00:06:35.249 --> 00:06:38.810
and understanding the tectonics
and geology of the area a bit better.

00:06:38.810 --> 00:06:42.009
We have pretty good
workflows in place for this,

00:06:42.009 --> 00:06:44.100
so we can do everything
very efficiently.

00:06:44.100 --> 00:06:47.620
But that’s taken a lot of time to
set up and understand, you know,

00:06:47.640 --> 00:06:51.000
as soon as you get people
changing, that all changes.

00:06:51.000 --> 00:06:54.400
And then we – new stories
go on for months and years.

00:06:54.400 --> 00:06:58.931
After each one of these,
it’s been a lot of time involved in that.

00:06:58.931 --> 00:07:03.419
And the final one there  is just,
all of the things above take time.

00:07:03.419 --> 00:07:06.780
But really, what’s been the
biggest surprise, I guess, for me,

00:07:06.780 --> 00:07:09.539
is just the considerable amount of
time with direct engagement with

00:07:09.539 --> 00:07:13.319
both government and the
engineering communities and others

00:07:13.319 --> 00:07:17.330
that goes on for a year or two after
these events, that – that’s really –

00:07:17.330 --> 00:07:23.600
that’s where most of the effort in this
kind of OEF space is going for us.

00:07:24.360 --> 00:07:28.400
So we’re currently trying to improve all
this and get a fully automated system.

00:07:28.400 --> 00:07:30.940
Hopefully, we’ll have that
up in the next few months.

00:07:30.940 --> 00:07:34.020
It presents IT challenges,
as people in this room know,

00:07:34.020 --> 00:07:36.800
and we’re trying to work through
some of those, of how we make sure

00:07:36.810 --> 00:07:39.770
that people are reliant on this,
and that we’re actually going to

00:07:39.770 --> 00:07:44.460
get it to them when they need it.
I won’t go into any details there.

00:07:45.440 --> 00:07:49.270
So just into the forecasting methods
themselves, I guess our overriding

00:07:49.270 --> 00:07:52.720
philosophy has always been,
and has become even more so,

00:07:52.720 --> 00:07:56.439
that no single model or data can
really represent everything that we know

00:07:56.439 --> 00:08:01.490
about future earthquake occurrence.
So we best – we get our best forecasts

00:08:01.490 --> 00:08:04.849
by using multiple models and really
trying to capture that uncertainty

00:08:04.849 --> 00:08:09.559
and be able to get that into the forecast
and also communicate that uncertainty.

00:08:09.560 --> 00:08:13.060
And using different
data sets to get into that.

00:08:13.060 --> 00:08:16.060
And then I’ll talk
a little bit more about that.

00:08:17.200 --> 00:08:21.009
So this is the hybrid statistical
model that’s evolved.

00:08:21.009 --> 00:08:23.069
It started out
very much the same.

00:08:23.069 --> 00:08:27.060
In Christchurch, there’s been
modifications made through it in time.

00:08:27.060 --> 00:08:30.169
But we’re essentially
doing the same core thing.

00:08:30.169 --> 00:08:32.670
And it’s a hybrid source model,
so it’s got three different components,

00:08:32.670 --> 00:08:37.360
where you’d think of the components in
terms of time, possibly with short-term,

00:08:37.360 --> 00:08:41.789
which is days to weeks to years.
Medium-term, which is years to decades.

00:08:41.789 --> 00:08:45.600
And then the long-term,
when you get out from there.

00:08:45.600 --> 00:08:48.019
I mean, ideally, this would be
a fully objective model based

00:08:48.019 --> 00:08:51.100
purely on statistical testing.
But, I guess, as people know, you can’t

00:08:51.100 --> 00:08:53.589
really do that no matter what we do.
Anywhere in this space, there’s

00:08:53.589 --> 00:08:58.920
a fair bit of subjectivity that comes into
that, and I’ll talk a little bit about that.

00:08:59.840 --> 00:09:03.699
On the – on the short-term side,
some of you are probably familiar

00:09:03.699 --> 00:09:06.240
with the step model.
That’s the same one that was

00:09:06.240 --> 00:09:08.540
running in California for a while.
So that’s – I would call that

00:09:08.540 --> 00:09:11.610
really an ETAS-class model.
It’s the same type.

00:09:11.610 --> 00:09:16.080
It’s just more of an algorithmic approach
to ETAS than a mathematical approach.

00:09:17.400 --> 00:09:19.399
We combine that, then,
with these EEPAS models,

00:09:19.399 --> 00:09:22.800
which I’ll talk a bit
about more in a minute.

00:09:22.800 --> 00:09:25.721
You can – you can think of EEPAS
as similar to ETAS, but it’s got

00:09:25.721 --> 00:09:29.230
completely different scaling laws in it.
And then we combine that with

00:09:29.230 --> 00:09:32.529
a long-term, so that’s where we
expect things to get to in anywhere

00:09:32.529 --> 00:09:36.850
from a decade to 25 years,
just depending on what’s going on.

00:09:36.850 --> 00:09:39.079
And you can see,
on this figure over here,

00:09:39.079 --> 00:09:42.410
this is looking at the short-
and the medium-term together.

00:09:42.410 --> 00:09:44.670
This is the – and this is
for a single location.

00:09:44.670 --> 00:09:46.510
Sorry, you probably
can’t read these numbers.

00:09:46.510 --> 00:09:48.779
But we’ve got
orders of magnitude difference

00:09:48.779 --> 00:09:51.570
in expected rates that
come out of the models.

00:09:51.570 --> 00:09:59.250
And those all get combined together at
each kind of space-time-magnitude step.

00:09:59.250 --> 00:10:01.540
And this is the range and the
uncertainty in the background model.

00:10:01.540 --> 00:10:07.080
So this long-term – we also have a
range of models that we – that we put in.

00:10:09.700 --> 00:10:14.000
The medium-term one is probably
something that we do a little bit

00:10:14.000 --> 00:10:17.009
differently than down here at the USGS
and also other places in the world,

00:10:17.009 --> 00:10:21.199
and that’s this EEPAS model.
So it’s based on work that

00:10:21.199 --> 00:10:24.590
David Rhoades and Frank Evison started
back in the ’70s with what they were

00:10:24.590 --> 00:10:27.990
calling the precursory scale increase.
So really looking at coming from

00:10:27.990 --> 00:10:32.720
swarm-type activity and what the
influence is that on the future.

00:10:32.720 --> 00:10:36.670
And they developed these scalings –
or they – I guess they produced

00:10:36.670 --> 00:10:40.740
these scaling laws from different
data sets around the world.

00:10:40.740 --> 00:10:43.300
This has got – sorry,
those are really hard to read,

00:10:43.300 --> 00:10:46.649
but it’s got from lots of different
tectonic regions around the country.

00:10:46.649 --> 00:10:50.500
And it really looks at – kind of you
can think of it as looking at the swarm.

00:10:50.500 --> 00:10:54.490
So it looks at kind of a cumulative
magnitude for a region for a given time,

00:10:54.490 --> 00:10:58.300
and then it looks at a space
and a time window around that.

00:10:58.300 --> 00:11:02.920
And, on the X axis on each one
of these is that kind of – I think it is –

00:11:02.920 --> 00:11:07.440
is that precursory magnitude.
So that average magnitude for

00:11:07.440 --> 00:11:10.459
a region is the easiest way to think of it.
As you can see, as you –

00:11:10.459 --> 00:11:13.779
as that magnitude goes up,
the magnitude of the

00:11:13.779 --> 00:11:18.600
kind of future earthquakes
go up with that very –

00:11:18.600 --> 00:11:20.820
I guess very well-constrained,
I would say.

00:11:20.820 --> 00:11:26.060
And in time too, so as the magnitudes
go up, the looking ahead time goes up.

00:11:26.060 --> 00:11:28.440
And it goes out to
kind of decades out here

00:11:28.440 --> 00:11:30.820
for when you’re looking
at magnitude 7s and 8s.

00:11:30.820 --> 00:11:34.980
And then the area scales
as you would expect as well.

00:11:34.980 --> 00:11:38.209
So the strongest information –
the forecast information and the

00:11:38.209 --> 00:11:43.380
highest probability gains from this really
come in kind of the 2- to 25-year range.

00:11:43.380 --> 00:11:47.639
It’s been tested globally in
CSEP testing centers around the world

00:11:47.639 --> 00:11:52.050
and lots of other different ways,
and it’s a very skillful model.

00:11:52.050 --> 00:11:59.149
It may seem very similar to
ETAS in that it’s – every earthquake

00:11:59.149 --> 00:12:02.810
is influencing the probability of
future earthquakes in the region.

00:12:02.810 --> 00:12:06.709
But you can’t reproduce
ETAS behavior – sorry,

00:12:06.709 --> 00:12:10.360
you can’t reproduce EEPAS
behavior within any test catalog.

00:12:10.360 --> 00:12:13.680
So it’s capturing – EEPAS is
capturing some process that’s

00:12:13.680 --> 00:12:18.720
not seen in the ETAS model that
goes out for longer time periods.

00:12:18.720 --> 00:12:22.750
And my personal opinion, I guess,
is this is a very important model, or

00:12:22.750 --> 00:12:26.500
these processes are important processes
that we’re not capturing in other ways.

00:12:26.500 --> 00:12:29.490
And they’re really at timeframes
that we’re finding that the end users

00:12:29.490 --> 00:12:32.839
are very much interested in.
This is the – this is the type of time scale

00:12:32.840 --> 00:12:37.959
that people are wanting from us, so this
has been a really useful model to us.

00:12:39.420 --> 00:12:42.640
Going on the long-term –
so we can go out 2 to 25 years, say,

00:12:42.649 --> 00:12:44.709
with the EEPAS model,
but we – people are wanting to know,

00:12:44.709 --> 00:12:48.540
when are we going to the –
to get to the background level?

00:12:48.540 --> 00:12:50.759
How do we merge this with the
National Seismic Hazard Model?

00:12:50.759 --> 00:12:53.870
Things like that.
So it ensures that the forecasts are

00:12:53.870 --> 00:12:59.600
spatially complete and continuous
with the – with the national model.

00:13:00.540 --> 00:13:05.000
So no matter what we – no matter what
we do, the time-independent models,

00:13:05.010 --> 00:13:07.089
they’re called – the long-term model –
they’re not really time-independent.

00:13:07.089 --> 00:13:09.290
They’re dependent
on the time length of the catalog

00:13:09.290 --> 00:13:11.410
and the data
that goes into them.

00:13:11.410 --> 00:13:14.170
Each decision you
make there in terms of

00:13:14.170 --> 00:13:17.420
what time period you use
affects what your forecast is.

00:13:18.020 --> 00:13:21.800
And so it’s the same for, say,
a 50-year forecast from a hazard model.

00:13:21.810 --> 00:13:25.540
So we include these variations
into the modeling that goes in.

00:13:25.540 --> 00:13:28.850
And we do that using these hybrid
combinations of multiple models.

00:13:28.850 --> 00:13:32.571
And those have been demonstrated to
consistently provide the best forecasts,

00:13:32.580 --> 00:13:36.340
and I’ll show a bit more
what I mean by that.

00:13:38.800 --> 00:13:41.730
So there’s different
ways you can do hybrids.

00:13:41.730 --> 00:13:45.690
You can think of both step
and ETAS as hybrid models.

00:13:45.690 --> 00:13:47.889
Those are additive hybrids, maybe,
where you’re adding in some

00:13:47.889 --> 00:13:53.110
background component with some
time-varying component on top of that.

00:13:53.110 --> 00:13:55.440
And we’ve done a lot in that space,
but most things that we’re using now

00:13:55.440 --> 00:13:59.790
we would call a multiplicative hybrid.
And the basic idea there is,

00:13:59.790 --> 00:14:03.470
you have some base model –
in this case, this is the

00:14:03.470 --> 00:14:05.380
Helmstetter smooth
seismicity model,

00:14:05.380 --> 00:14:08.769
and you have some other variant
that you – that you scale that with.

00:14:08.769 --> 00:14:12.480
So here we have the NeoKinema model.
And you’re essentially just multiplying

00:14:12.480 --> 00:14:17.430
this model by this model,
or however you want to do it.

00:14:17.430 --> 00:14:21.990
And that’s – your forecast is based on
that, and that scaling is done in some

00:14:21.990 --> 00:14:26.260
statistical optimization period, and then
you test that in some other period.

00:14:26.260 --> 00:14:31.300
And we’re seeing very large gains
in doing that, and there’s no restriction

00:14:31.310 --> 00:14:33.879
in the type of information
that you can include in this.

00:14:33.879 --> 00:14:37.200
And this is some of our earlier work
where we look into basically different

00:14:37.200 --> 00:14:41.959
versions of smooth seismicity –
some that included the fault locations,

00:14:41.959 --> 00:14:47.040
some that included slip rates,
and various other information

00:14:47.040 --> 00:14:51.070
that you can combine together.
And each – the idea is that each one

00:14:51.070 --> 00:14:55.199
of those bits of information have
some knowledge about what our

00:14:55.199 --> 00:14:58.510
earthquakes are going to be in the future.
And you get your best forecast

00:14:58.510 --> 00:15:01.660
by combining that all together
in some sort of optimized fashion.

00:15:01.660 --> 00:15:06.440
So you’re looking at as many different
ideas as you can to get in there.

00:15:06.440 --> 00:15:10.089
And then, some of the first hybrid
modeling work we did was based on

00:15:10.089 --> 00:15:13.240
this sets of models, some of –
as well as some others.

00:15:13.240 --> 00:15:18.300
And we were getting quite significant
gains in the forecasting skills of the –

00:15:18.300 --> 00:15:22.760
of the kind of the standard smooth
seismicity models by doing this.

00:15:23.620 --> 00:15:27.380
Then, more recently,
we started looking at the impact

00:15:27.389 --> 00:15:29.430
of including
strain rates into that.

00:15:29.430 --> 00:15:31.690
So this is purely strain rate
that we’re looking at here.

00:15:31.690 --> 00:15:35.000
This is shear strain, rotational strain,
and dilatational strain.

00:15:35.000 --> 00:15:39.260
There may be – you could argue
how these are calculated and so on,

00:15:39.260 --> 00:15:42.720
but the idea here is
this is just purely a strain rate.

00:15:42.720 --> 00:15:47.560
There’s no information about
earthquake rate in these maps.

00:15:47.560 --> 00:15:50.700
And you essentially just, then,
combine that strain rate together

00:15:50.709 --> 00:15:54.949
with whatever earthquake rate
information that you have so you get a –

00:15:54.949 --> 00:15:58.079
just a simple smooth seismicity model,
and you combine the two in some

00:15:58.079 --> 00:16:02.850
sort of statistically optimized fashion.
And this is probably the biggest

00:16:02.850 --> 00:16:06.410
change we’ve seen in
how good our forecasting models

00:16:06.410 --> 00:16:07.769
are out of anything
that we’ve done.

00:16:07.769 --> 00:16:15.170
We’re getting quite significant gains
in how well, spatially, earthquakes

00:16:15.170 --> 00:16:19.100
are forecast, when you start to include,
particularly, the shear strain.

00:16:19.100 --> 00:16:21.690
So, sorry, this is a –
quite a busy slide,

00:16:21.690 --> 00:16:28.160
but this is just to show, if we looked at
the work we did – so this, on the right,

00:16:28.160 --> 00:16:36.200
this is a comparison in – this is just in an
optimization period, I think, right here.

00:16:36.200 --> 00:16:40.360
And the way you read this is,
each one of these, the dot is kind of the –

00:16:40.360 --> 00:16:43.720
a test score with the
uncertainty around that,

00:16:43.730 --> 00:16:48.149
all relative to the –
to some sort of zero line here.

00:16:48.149 --> 00:16:55.879
And basically, the more to the right
that result is, the better the forecast is.

00:16:55.879 --> 00:16:58.380
And all the work that we were
doing prior to including strain rate,

00:16:58.380 --> 00:17:02.829
the best model was down here.
And as soon as we started adding in

00:17:02.829 --> 00:17:07.290
shear strain rate and looking at how we
could optimize these things, all of the

00:17:07.290 --> 00:17:11.960
models started performing significantly
better and more significantly better.

00:17:12.990 --> 00:17:16.760
We stopped the optimization in 2012,
so we don’t have a particularly

00:17:16.760 --> 00:17:20.430
long testing period yet.
This particular result was just

00:17:20.430 --> 00:17:25.530
looking at 2012 to 2015, so we had
50 earthquakes of 5 or greater in that.

00:17:25.530 --> 00:17:30.250
And, again, if we get – in that testing
period, we get very large significant

00:17:30.250 --> 00:17:34.870
gains by using the strain
rate information into the models.

00:17:34.870 --> 00:17:36.810
So we’re putting all of
this kind of hybrid model

00:17:36.810 --> 00:17:39.140
information into the
long-term forecasting.

00:17:39.140 --> 00:17:43.031
So sometimes that comes in
in 10 years. Sometimes it starts to

00:17:43.031 --> 00:17:45.370
come in in 20 years or later.
But we’re seeing really

00:17:45.370 --> 00:17:50.140
big improvements in how good
the modeling is based on that.

00:17:52.380 --> 00:17:55.260
So I don’t know if most people are
familiar with the Collaboratory for

00:17:55.260 --> 00:17:59.580
the Study of Earthquake Predictability,
or CSEP, out of SCEC.

00:17:59.580 --> 00:18:02.460
Everything we’re doing,
we’re kind of centered on CSEP.

00:18:02.470 --> 00:18:04.620
We have our own tests that we
do outside of that, but we’re

00:18:04.620 --> 00:18:10.730
heavily involved with SCEC.
And that testing has been

00:18:10.730 --> 00:18:15.000
very valuable to us in all the forecasting
information that we’ve put out.

00:18:15.840 --> 00:18:19.100
I mean, it’s – in the Jordan et al. report,
it was highlighted as something

00:18:19.100 --> 00:18:21.910
that’s necessary to do, and
I guess I would agree with that.

00:18:21.910 --> 00:18:24.240
I mean, there’s weaknesses in tests
that we don’t want to overstate the

00:18:24.240 --> 00:18:27.630
value of the – coming from the test,
but there’s a lot we can learn from them.

00:18:27.630 --> 00:18:30.010
And it’s been critical to
the development of our models.

00:18:30.010 --> 00:18:32.700
So understanding, say,
how we do these optimization

00:18:32.700 --> 00:18:35.820
and hybrid combinations,
just as an example, confidence.

00:18:35.820 --> 00:18:39.530
So we’re putting forecasts
out that are having pretty big impacts,

00:18:39.530 --> 00:18:42.570
as you’ll see later on.
So giving us confidence that we’re

00:18:42.570 --> 00:18:46.710
doing the right thing here, that there’s
skill in the models that we’re putting out.

00:18:46.710 --> 00:18:50.770
And credibility.
This one’s maybe not as good

00:18:50.770 --> 00:18:57.350
as it should be, coming from the test,
and I think that’s one where we have

00:18:57.350 --> 00:19:01.050
a lot of work to do, and that’s
these test results that you can put out,

00:19:01.050 --> 00:19:02.970
just like you can imagine
with that last slide I showed you.

00:19:02.970 --> 00:19:04.580
Those are pretty hard for
people to get their heads around to –

00:19:04.580 --> 00:19:07.410
people who are technical.

00:19:07.410 --> 00:19:11.220
So trying to communicate
these test results is challenging.

00:19:11.220 --> 00:19:12.760
And then even when people
then understand them,

00:19:12.760 --> 00:19:16.450
they just maybe flat-out don’t
want to – want to believe them.

00:19:16.450 --> 00:19:19.990
So we really need to be working on
how we test these models that are

00:19:19.990 --> 00:19:24.530
in more relevant space to the end users
of the forecasts so we can give them

00:19:24.530 --> 00:19:27.101
information and say, this is what
you’re doing with the forecast

00:19:27.101 --> 00:19:30.070
information that we’re putting out.
And this is how we can show you that

00:19:30.070 --> 00:19:35.060
it’s actually going to improve how you
make those decisions based on that.

00:19:35.060 --> 00:19:36.890
And then we just need to
be able to figure out how to

00:19:36.890 --> 00:19:40.900
communicate these better.
That’s always a big challenge.

00:19:42.780 --> 00:19:45.760
So shift gears just a little bit.
So that was kind of the more

00:19:45.760 --> 00:19:51.470
standard kind of OEF-type
forecasts that we’re putting out.

00:19:51.470 --> 00:19:56.000
And then, following Kaikoura,
we had to do something a little bit

00:19:56.000 --> 00:19:59.800
different that I’ll just talk about.
I don’t know if you just remember

00:19:59.800 --> 00:20:03.170
your New Zealand tectonics,
or if you looked at it before.

00:20:03.170 --> 00:20:05.880
But if we’re looking from –
New Zealand from north down to south,

00:20:05.880 --> 00:20:09.640
here’s the Hikurangi
Margin right along here.

00:20:09.640 --> 00:20:12.650
Capital city of Wellington is right here.
So the Hikurangi is about

00:20:12.650 --> 00:20:16.000
20, 25 kilometers beneath
Wellington right there.

00:20:16.000 --> 00:20:18.760
Then it transfers down into the
Alpine Fault and goes offshore

00:20:18.760 --> 00:20:24.440
down there again. But Kaikoura
all happened right around in here.

00:20:25.380 --> 00:20:28.940
And when that happened,
it had quite a big hit on –

00:20:28.940 --> 00:20:31.690
when Kaikoura happened,
it had quite a big hit in terms of

00:20:31.690 --> 00:20:35.420
both static and dynamic
stress changes on the Hikurangi.

00:20:35.420 --> 00:20:42.390
And pretty much immediately, big
chunks of the Hikurangi started to slip.

00:20:42.390 --> 00:20:46.820
So slow-slip patches
off the East Cape along here.

00:20:46.820 --> 00:20:50.480
A little bit down-dip
in the Manawatu in here.

00:20:50.480 --> 00:20:54.880
And then off of Kapiti
down-dip from Wellington.

00:20:54.880 --> 00:20:58.760
So Wellington sits right here, and then
down-dip from Wellington out here.

00:20:58.760 --> 00:21:01.840
And then, this – all the –
all the afterslip – whatever you

00:21:01.840 --> 00:21:04.140
want to call it – afterslip, slow slip,
also happened at this point.

00:21:04.140 --> 00:21:05.520
We didn’t have
good instrumentation there,

00:21:05.520 --> 00:21:08.040
so we didn’t know
this was going on.

00:21:08.800 --> 00:21:11.460
So a couple of interesting things here.
One, we had never –

00:21:11.460 --> 00:21:13.810
we don’t have – obviously
don’t have that many years

00:21:13.810 --> 00:21:18.920
of observations of slow slip in
New Zealand – only 20, 25 years.

00:21:18.920 --> 00:21:22.320
But during that time, we had never
seen all those patches go together.

00:21:22.320 --> 00:21:25.930
We had seen them go relatively
regularly from every six months

00:21:25.930 --> 00:21:29.720
to every two years kind of thing,
but never all at the same time.

00:21:29.720 --> 00:21:30.890
Maybe it’s not a
surprise they did,

00:21:30.890 --> 00:21:33.760
since they did take a bit hit,
but we hadn’t seen it.

00:21:33.760 --> 00:21:38.610
And also, in the time of observations
that we have, this patch right here

00:21:38.610 --> 00:21:42.180
has been completely locked.
There has been no movement on that.

00:21:42.180 --> 00:21:45.100
So there were concerns raised about,
what does it mean for this

00:21:45.100 --> 00:21:47.510
locked patch, which is
essentially underneath the city

00:21:47.510 --> 00:21:52.820
of Wellington, if we’ve got all this
movement going on around it?

00:21:53.990 --> 00:21:57.300
We told the government and the public
and so on that this was going on, but we

00:21:57.310 --> 00:22:01.480
didn’t really know what to – what else
we could say beyond that at that point.

00:22:01.480 --> 00:22:04.330
And the government was
not impressed, necessarily,

00:22:04.330 --> 00:22:06.740
that we didn’t –
couldn’t give them probabilities.

00:22:06.740 --> 00:22:09.200
We had done such a good job
over the last decade of putting

00:22:09.200 --> 00:22:12.630
probabilistic information out that
they wanted us to be able to say

00:22:12.630 --> 00:22:16.140
something quantitative about that.
And that was for reasons about

00:22:16.140 --> 00:22:19.080
fast-tracking big decisions
that they already had going on.

00:22:19.080 --> 00:22:25.370
So all thinking about – I guess
resilience of Wellington city

00:22:25.370 --> 00:22:27.810
and how they deal with major
roading infrastructure and so on.

00:22:27.810 --> 00:22:31.820
They had big decisions to make there
that they wanted some guidance on.

00:22:34.220 --> 00:22:38.320
So we had – initially, we had
about a two-week project that’s –

00:22:38.320 --> 00:22:40.410
they were obviously
on a very tight timeframe.

00:22:40.410 --> 00:22:43.900
So we had to make some information
quickly, so we did that first in

00:22:43.900 --> 00:22:47.620
December of 2016, where we
basically got a group of people

00:22:47.620 --> 00:22:50.160
together and evaluated
all the evidence that we could find,

00:22:50.160 --> 00:22:52.280
that I’ll go through
in a little bit here.

00:22:52.280 --> 00:22:56.370
And then, in – last year,
we went through in a bit more detail

00:22:56.370 --> 00:23:00.010
and did some new model development
with the larger panel of kind of

00:23:00.010 --> 00:23:04.240
international people to try to get
our heads around this a bit better.

00:23:04.690 --> 00:23:09.600
So we essentially evaluated this range
of models and evidence to come up with

00:23:09.600 --> 00:23:14.390
what – some sort of quantitative numbers
of what we thought might happen.

00:23:14.390 --> 00:23:18.990
And it was done in this series of
two workshops – one at SCEC last year

00:23:18.990 --> 00:23:21.420
and then a final one in November
in Wellington last year.

00:23:21.420 --> 00:23:23.300
And this was really –
the result of this was we had to

00:23:23.300 --> 00:23:25.380
give some numbers
to the government.

00:23:25.380 --> 00:23:28.670
And we used a structured
expert elicitation procedure

00:23:28.670 --> 00:23:32.920
with expert calibration,
which I’ll talk about in a minute.

00:23:34.580 --> 00:23:38.320
So we had a lot of different models
that we went through and a lot of

00:23:38.320 --> 00:23:41.800
different observations, and I won’t
go into the details on all of this,

00:23:41.800 --> 00:23:47.400
but we have a large catalog now of
behavior of New Zealand earthquakes

00:23:47.400 --> 00:23:49.970
in the region of slow-slip events
when slow slip is occurring

00:23:49.970 --> 00:23:53.970
and in times after that.
So just – I mean, a simple observation is,

00:23:53.970 --> 00:23:58.200
on average, we had a doubling of rate –
doubling of numbers of events during

00:23:58.200 --> 00:24:02.700
times of slow-slip earthquakes
that slowly tails off after that.

00:24:02.700 --> 00:24:04.980
We have our statistical
forecasting models, so the models

00:24:04.980 --> 00:24:08.470
that I was showing you early on.
We have about 10-plus models

00:24:08.470 --> 00:24:11.930
based on different ideas.
So we had all that as evidence.

00:24:11.930 --> 00:24:15.570
We did various things
related to equivalent magnitudes

00:24:15.570 --> 00:24:18.420
of the slow-slip
earthquake magnitude.

00:24:18.420 --> 00:24:22.400
So scaling of, say, Omori rates
or something else based on that.

00:24:22.400 --> 00:24:25.910
So we did that based on Omori
or based on rate-and-state.

00:24:25.910 --> 00:24:29.020
So it’s really just scaling down
their productivity based on

00:24:29.020 --> 00:24:33.940
what you see on the strain rates
or on the aftershock rates.

00:24:34.740 --> 00:24:37.140
We have Russell Robinson’s
ARTs simulator, which some of you

00:24:37.140 --> 00:24:41.340
may be familiar with.
So he’s got a – he uses static

00:24:41.340 --> 00:24:44.410
and dynamic stress changes.
He used a fault model of central

00:24:44.410 --> 00:24:47.010
New Zealand, and he runs that
for millions of years so we could

00:24:47.010 --> 00:24:50.860
look into that and see what types of
interactions he was seeing in that.

00:24:50.860 --> 00:24:56.130
We have a simple physical simulator,
which I’ll talk about next.

00:24:56.130 --> 00:24:59.720
We have paleoseismic data.
We have very limited paleoseismic

00:24:59.720 --> 00:25:04.140
data of megathrust – so the main
goal here was, I guess I should say,

00:25:04.140 --> 00:25:07.240
was earthquakes in central New Zealand,
but obviously there was a focus

00:25:07.240 --> 00:25:09.400
on the megathrust and
what could happen there.

00:25:09.410 --> 00:25:12.960
And we have very little data on
what has happened in the past.

00:25:12.960 --> 00:25:14.440
There’s some new
data coming out,

00:25:14.440 --> 00:25:17.600
but it’s on very, very limited
numbers of data points.

00:25:17.600 --> 00:25:22.120
And then, as a baseline for all of this,
was the National Seismic Hazard Model.

00:25:22.120 --> 00:25:25.890
So just more examples of that I don’t –
I don’t need to go into.

00:25:25.890 --> 00:25:30.480
They’re more, I guess, just describing
what I talked about in the last slide.

00:25:31.400 --> 00:25:36.960
So this simple physical model, this was
largely work done by Yoshi Kaneko.

00:25:36.960 --> 00:25:42.590
And it’s really just trying to get at how –
if we make some basic assumptions,

00:25:42.590 --> 00:25:47.130
how can we get at the impact of these
slow-slip earthquakes on the megathrust

00:25:47.130 --> 00:25:51.060
in terms of stress changes and what that
might mean for a future earthquake.

00:25:51.060 --> 00:25:54.510
So really, the first step was then
to estimate the stress changes

00:25:54.510 --> 00:25:58.950
on that locked patch, which is
obviously spatially variable.

00:25:58.950 --> 00:26:01.750
And then develop a synthetic
earthquake catalog over a million years.

00:26:01.750 --> 00:26:07.290
So in that, you’re sampling from a range
of possible stress drops on megathrust

00:26:07.290 --> 00:26:12.150
earthquakes and a range of recurrence
intervals where the recurrence intervals

00:26:12.150 --> 00:26:15.470
were largely constrained by
the limited paleoseismic data we had,

00:26:15.470 --> 00:26:19.550
but with very large
uncertainties on that distribution.

00:26:19.550 --> 00:26:22.970
So you create a million-year catalog –
something such as you see here,

00:26:22.970 --> 00:26:28.560
where you’re then putting in all the
stressing rate from the region into that.

00:26:30.540 --> 00:26:35.640
And then you apply –
randomly apply stress perturbations

00:26:35.640 --> 00:26:39.120
from a Kaikoura-like event.
And the idea is that there is some

00:26:39.120 --> 00:26:42.330
pre-determined threshold in
each one of these cycles that –

00:26:42.330 --> 00:26:46.650
randomly pre-determined thresholds,
and if that stress perturbation is then

00:26:46.650 --> 00:26:50.410
pushing it over, that’s kind of
some very baseline – pushing it over

00:26:50.410 --> 00:26:53.900
that threshold – that’s a very baseline
estimate of what your probability

00:26:53.900 --> 00:26:57.250
could be given absence
of any other information,

00:26:57.250 --> 00:27:00.780
which is the situation
that we were in.

00:27:01.950 --> 00:27:06.480
So this is just showing the results of that.
So here are the stress changes

00:27:06.490 --> 00:27:10.140
on the locked patch.
It’s obviously variable.

00:27:10.140 --> 00:27:12.730
So each dot here
corresponds to a mean

00:27:12.730 --> 00:27:17.850
of 50 runs of
10 million-year simulations on that.

00:27:17.850 --> 00:27:21.470
And the vertical corresponds to the
plus-or-minus 1 standard deviation error

00:27:21.470 --> 00:27:25.420
on the annual probability of
a large earthquake occurring

00:27:25.420 --> 00:27:30.160
on the megathrust following
one of these stress perturbations.

00:27:30.160 --> 00:27:31.950
And you can see –
so this is looking at years

00:27:31.950 --> 00:27:36.100
from the occurrence
of the Kaikoura-like event.

00:27:36.100 --> 00:27:40.900
And you can see very quickly,
those probabilities are dropping off.

00:27:40.900 --> 00:27:46.980
So, within – I mean, the stress
changes from the slow-slip events,

00:27:46.980 --> 00:27:50.120
related to everything that else is
going on is just – is relatively small.

00:27:50.120 --> 00:27:54.490
So it’s not having a big impact
in this particular model.

00:27:54.490 --> 00:27:56.940
And within – probably
within two years, it’s almost

00:27:56.940 --> 00:28:03.400
back down to the background level.
So any stress there is is getting absorbed.

00:28:03.400 --> 00:28:06.230
But interestingly, or maybe
not so interestingly, but really,

00:28:06.230 --> 00:28:10.200
what it seems to be coming down to is
that the ratio of the – it’s all controlled

00:28:10.200 --> 00:28:14.450
by the ratio of the total stressing rate
over the stress drop of the earthquakes.

00:28:14.450 --> 00:28:17.540
And that’s the simple
back-of-the-envelope calculation

00:28:17.540 --> 00:28:20.800
that fell out of this is that’s how
you can get at these probabilities.

00:28:20.800 --> 00:28:24.630
So we’re looking at just
probabilities of – within the first year

00:28:24.630 --> 00:28:28.620
or two of between 2 and 4% per year,
which are elevated over what

00:28:28.620 --> 00:28:31.900
we would normally expect
given the limited data we have,

00:28:31.900 --> 00:28:37.660
but, depending on your perspective,
not high or not low.

00:28:39.420 --> 00:28:43.100
Okay, so now the problem is,
we have – I’ll take a bit of a segue here.

00:28:43.100 --> 00:28:46.340
We’ve got all these different models
and different bits of evidence.

00:28:46.340 --> 00:28:47.920
None of them do
we fully believe in.

00:28:47.920 --> 00:28:50.450
I mean, a very simple
model that we’re looking at.

00:28:50.450 --> 00:28:52.530
Maybe the best-constrained
are the statistical models.

00:28:52.530 --> 00:28:55.720
None of those explicitly include
slow-slip earthquakes, although they

00:28:55.720 --> 00:28:59.130
would be there in some of the data sets.
But we have to come up with an answer

00:28:59.130 --> 00:29:01.340
from that and do it in
a fairly tight timeframe.

00:29:01.340 --> 00:29:05.100
So I think this is something that we
do a lot in our sciences, obviously.

00:29:05.100 --> 00:29:09.520
And the answer for
that is expert judgment.

00:29:09.520 --> 00:29:14.860
How you do the expert judgment itself is
a significant contributor to uncertainty.

00:29:14.860 --> 00:29:19.060
It’s often used in applied science,
but I think we could do a lot better.

00:29:19.060 --> 00:29:22.060
And the way it’s often done,
it’s very ad hoc.

00:29:22.060 --> 00:29:24.540
There’s a lot – there’s entire
bodies of research looking into,

00:29:24.540 --> 00:29:27.880
how do – how to do
expert elicitation.

00:29:27.880 --> 00:29:30.480
I think that we have
a lot to learn from that.

00:29:30.480 --> 00:29:32.480
Uncertainty is something
we want to capture in this.

00:29:32.490 --> 00:29:34.990
It’s not something we want to remove.
Quite often, when we do expert

00:29:34.990 --> 00:29:38.240
elicitation, we’re trying to come to
some sort of consensus, which means

00:29:38.240 --> 00:29:43.370
we’re necessarily ignoring uncertainty.
As everyone knows, there’s big

00:29:43.370 --> 00:29:46.620
personality interactions and social
interactions that affect what happens

00:29:46.620 --> 00:29:52.230
within an elicitation, and quite often,
they’re not very transparent.

00:29:52.230 --> 00:29:54.810
So this is something we’ve
had a lot of experience in in this area

00:29:54.810 --> 00:29:57.280
and other areas in
the last decade or more.

00:29:57.280 --> 00:30:01.360
So I’ll just give a quick example
of the types of things that we do.

00:30:02.120 --> 00:30:07.020
But I guess what – as I was just saying,
what’s behind this is that it’s –

00:30:07.020 --> 00:30:10.520
in a lot of the research, and in
our experience, it’s not really

00:30:10.520 --> 00:30:14.320
reasonable to expect consensus
when you have something like this.

00:30:14.320 --> 00:30:17.520
And the uncertainty
should be modeled, not excluded.

00:30:17.520 --> 00:30:19.760
So it’s all based on what’s
called a rational consensus,

00:30:19.770 --> 00:30:22.560
and that’s – you get the experts together,
and maybe they’re not going to agree

00:30:22.560 --> 00:30:25.480
with the result in the end. Somebody
may very much not agree with it.

00:30:25.480 --> 00:30:28.940
But everybody’s agreed to the process
of how you’re getting to that result.

00:30:28.940 --> 00:30:31.640
And that’s
where you start.

00:30:31.640 --> 00:30:35.090
So we do something based on
the Cooke method, or the

00:30:35.090 --> 00:30:38.220
classical model that’s been
around for a little while now,

00:30:38.220 --> 00:30:41.070
and that’s based –
called calibrating of the experts.

00:30:41.070 --> 00:30:43.750
Really, what that means is,
we’re testing the experts.

00:30:43.750 --> 00:30:51.540
And the tests are not to get at how
well they know the answers or how well

00:30:51.540 --> 00:30:54.410
they know the information, necessarily.
It’s really trying to get at

00:30:54.410 --> 00:30:57.790
the confidence of the experts.
So how good are the experts

00:30:57.790 --> 00:31:02.310
at estimating the uncertainties
in their own knowledge?

00:31:02.310 --> 00:31:04.260
So are they overconfident?
Are they not?

00:31:04.260 --> 00:31:05.990
And how broad are their
uncertainties, and so on?

00:31:05.990 --> 00:31:10.380
So really, you ask them a set of
subject-relevant questions,

00:31:10.380 --> 00:31:13.500
where subject-relevant is quite broad.
I won’t go into the details there.

00:31:13.500 --> 00:31:15.960
And then you ask for
the 80% confidence bounds.

00:31:15.960 --> 00:31:20.810
And you say, if you do that over
20 questions, you would expect

00:31:20.810 --> 00:31:25.080
a perfectly well-calibrated expert –
if you give 20 questions, if they’re 80%

00:31:25.080 --> 00:31:29.310
confidence bounds, they’d look like this.
They’d have 10%, so two answers,

00:31:29.310 --> 00:31:31.790
kind of on the low side outside
of their confidence bounds.

00:31:31.790 --> 00:31:35.900
You’d have 10% above.
And then pretty evenly distributed

00:31:35.900 --> 00:31:39.210
in the middle with whatever that is –
16 of the answers falling in there.

00:31:39.210 --> 00:31:44.250
Now, this is a perfectly well-calibrated
expert, and this isn’t what you often see.

00:31:44.250 --> 00:31:48.370
Most often, you see most of the answers
falling out here, most of the answers

00:31:48.370 --> 00:31:52.040
falling out here, with – very common
for experts to be overconfident.

00:31:52.040 --> 00:31:55.860
So it’s really that understanding
that you’re trying to get at.

00:31:55.860 --> 00:31:57.760
And then there’s a score
developed based on this

00:31:57.760 --> 00:32:01.020
and based on the width of
their uncertainty and so on.

00:32:01.020 --> 00:32:03.240
And the questions
are quite complicated.

00:32:03.240 --> 00:32:05.070
They’re not something you
can just go calculate.

00:32:05.070 --> 00:32:10.900
And so if we were doing one in hazard,
for example, we might say – if we were

00:32:10.900 --> 00:32:16.630
doing one here, what is the difference
in the 10% in 50-year PGA expected

00:32:16.630 --> 00:32:21.260
for this site when you use
a declustered catalog

00:32:21.260 --> 00:32:24.410
for the background model versus
a non-declustered catalog?

00:32:24.410 --> 00:32:26.440
So quite complicated.
Lots of steps they have to go through.

00:32:26.440 --> 00:32:28.540
So something very similar to
what we’re actually doing

00:32:28.540 --> 00:32:31.960
in the elicitation, where there are
answers that we just don’t know.

00:32:31.960 --> 00:32:35.160
And here’s one example from this
particular elicitation that we did.

00:32:35.160 --> 00:32:38.790
This is a very good question.
We had a lot more detail in it than this.

00:32:38.790 --> 00:32:42.650
But – so we had slow-slip
experts from around the world in this,

00:32:42.650 --> 00:32:45.210
from lots of
different countries.

00:32:45.210 --> 00:32:47.690
And the question related –
was related to, what was the

00:32:47.690 --> 00:32:52.590
slow-slip rate in millimeters per hour
prior to the Tohoku-Oki?

00:32:52.590 --> 00:32:56.770
And sorry, these are terrible graphs,
but each one of these is an expert.

00:32:56.770 --> 00:32:59.250
Here is their best guess and
their uncertainty bounds.

00:32:59.250 --> 00:33:01.500
You can see everybody’s –
we’re on a log scale here,

00:33:01.500 --> 00:33:05.000
so everybody’s scattered
all over the place.

00:33:05.000 --> 00:33:10.980
And the true answer was 6, right here.
So you can see about half of the experts

00:33:10.980 --> 00:33:15.540
got it within their confidence bounds.
Half did not.

00:33:15.540 --> 00:33:18.460
Uncertainties are
very large on this.

00:33:18.460 --> 00:33:23.520
So you’re essentially optimizing
how they – how they answer across

00:33:23.520 --> 00:33:28.360
all the questions and getting the
scores to give that best result.

00:33:28.360 --> 00:33:31.970
And there’s lots of evidence
that this does quite well.

00:33:31.970 --> 00:33:36.300
And it’s not only the process of –
it’s not just the weighting

00:33:36.300 --> 00:33:38.840
that makes the difference.
Actually, a big part of what makes

00:33:38.840 --> 00:33:40.800
a difference is actually putting
the experts through this process

00:33:40.800 --> 00:33:44.170
and really forcing people to think about,
how well do I actually know this?

00:33:44.170 --> 00:33:45.430
What are the questions
that I can ask,

00:33:45.430 --> 00:33:49.100
and how do I get at
solving this problem?

00:33:49.100 --> 00:33:52.670
So it’s been a really useful tool
in trying to make decisions quickly

00:33:52.670 --> 00:33:56.700
and getting the best we can
out of a group of people.

00:33:57.610 --> 00:33:59.280
So just some of
the key information there.

00:33:59.290 --> 00:34:01.450
Reducing key is a bias.
There’s so many different biases

00:34:01.450 --> 00:34:05.830
that go on when you’re in these expert
elicitations, and being aware of those

00:34:05.830 --> 00:34:08.369
and making the experts aware of
those and how you handle that.

00:34:08.369 --> 00:34:13.310
So experts from across a range of
experience and background are critically

00:34:13.310 --> 00:34:16.869
critical to exploring the possibilities.
So this gets into some kind of

00:34:16.869 --> 00:34:20.639
non-standard ways of doing
things in our science community.

00:34:20.639 --> 00:34:22.419
People don’t always like it,
and one example is,

00:34:22.419 --> 00:34:25.700
the students have been shown to be –
students in early – very early

00:34:25.700 --> 00:34:28.839
career researchers have been shown to
be very good experts in these situations.

00:34:28.839 --> 00:34:31.450
They don’t – a lot of us come in
knowing essentially what the answer is,

00:34:31.450 --> 00:34:33.419
and we have to be
pushed away from that answer.

00:34:33.420 --> 00:34:37.560
Whereas, the earlier career
researchers quite often come in

00:34:37.560 --> 00:34:41.609
with a bit more of an open mind
and ask very good questions.

00:34:41.609 --> 00:34:44.129
Anonymity of the response is crucial.
As you can imagine,

00:34:44.129 --> 00:34:47.990
there’s all kinds of career dynamics
and so on that can go on in these groups.

00:34:47.990 --> 00:34:50.720
Open discussion is key.

00:34:51.780 --> 00:34:55.720
Okay. That was enough on that.
But here is the result, then.

00:34:55.730 --> 00:34:58.069
In this case, we actually
were very well-constrained.

00:34:58.069 --> 00:35:01.150
So this is the probabilities of –
across experts of a

00:35:01.150 --> 00:35:05.170
magnitude 7.8-plus in one
year in central New Zealand.

00:35:05.170 --> 00:35:06.980
And then this is decade
for the same thing.

00:35:06.980 --> 00:35:10.890
You can see very low probabilities.
And everyone, in this case, was very

00:35:10.890 --> 00:35:14.950
much anchored on the statistical models.
People found it pretty hard to move away

00:35:14.950 --> 00:35:20.210
from the statistical models, which were
really the best-constrained in this case.

00:35:20.210 --> 00:35:23.740
So I’ll talk a bit more about how we
put that information out in a second.

00:35:25.960 --> 00:35:27.660
I don’t have that many slides,
by the way.

00:35:27.660 --> 00:35:31.420
I’ve got a lot of empty slides at the end.
Check the time.

00:35:31.420 --> 00:35:35.670
So I just – a little bit about the
information that we’ve put out.

00:35:35.670 --> 00:35:38.380
As I mentioned, we have these different
timeframes in the way we think about

00:35:38.380 --> 00:35:41.760
the model, but that’s actually how they –
how they get used as well.

00:35:41.760 --> 00:35:45.210
So, in the very short term, that hasn’t
been necessarily how they’ve been –

00:35:45.210 --> 00:35:48.390
the models have been used as much.
But in the – kind of the days-to-weeks

00:35:48.390 --> 00:35:52.910
scale, that’s where you might think
about it in search-and-rescue and so on.

00:35:52.910 --> 00:35:56.670
Buildings inspections within –
it says days/weeks here,

00:35:56.670 --> 00:36:00.050
but that’s really gone into
months in some cases.

00:36:00.050 --> 00:36:03.200
So they’ve been fairly heavy users of
some of the information that we put out.

00:36:03.200 --> 00:36:06.349
And then getting through different kind
of large governmental-scale decisions

00:36:06.349 --> 00:36:11.130
in terms of occupying regions or
getting in and out of particular regions.

00:36:11.130 --> 00:36:13.720
And then you – as you get into the
months and years where financial

00:36:13.720 --> 00:36:18.099
decisions and recovery planning
decisions start to come in.

00:36:18.099 --> 00:36:22.930
And then your long-term decisions
related to land zoning and so on.

00:36:22.930 --> 00:36:26.900
And the public information is – I guess
is really just across all aspects of this.

00:36:26.900 --> 00:36:30.920
So the forecasts have really been
used in lots of different ways.

00:36:30.920 --> 00:36:33.369
And I’ll just give a few examples
of the information that we put out.

00:36:33.369 --> 00:36:35.920
Oh, and I just had
a couple probabilities.

00:36:36.840 --> 00:36:40.380
Initially, there was huge
resistance to probabilities.

00:36:40.380 --> 00:36:43.890
Had some very upset ministers
back in 2010 who did not like

00:36:43.890 --> 00:36:46.400
the probabilistic
information that we put out.

00:36:46.400 --> 00:36:49.690
But that’s all changed
a lot in the last decade.

00:36:49.690 --> 00:36:52.750
We’ve been very insistent
about keeping – finding new ways

00:36:52.750 --> 00:36:55.190
to communicate it and how we can
better improve the information

00:36:55.190 --> 00:36:58.619
that we put out but still insisting on
including probabilistic information.

00:36:58.619 --> 00:37:02.710
As I was mentioning, that’s
something that’s requested of us now,

00:37:02.710 --> 00:37:05.839
is they’ve gotten used to
being able to make decisions

00:37:05.839 --> 00:37:08.880
based on information
that we provide them.

00:37:08.880 --> 00:37:10.480
So this is
just an example.

00:37:10.480 --> 00:37:14.430
Very similar to things that
you guys have done in the past.

00:37:14.430 --> 00:37:18.260
Just a table that is
broken down by magnitudes.

00:37:18.260 --> 00:37:21.540
Numbers of events that are expected.
Kind of the mean range.

00:37:21.549 --> 00:37:25.589
The uncertainties on that.
And then the probability.

00:37:25.589 --> 00:37:28.490
And for – there’s always a context
that we’re providing – I don’t think

00:37:28.490 --> 00:37:32.700
we did initially do this, but based on
work from Anne and Sara and others,

00:37:32.700 --> 00:37:35.559
there’s obviously something that
was missing was this context.

00:37:35.559 --> 00:37:39.289
And what’s been working really well
in the last five or so years is context

00:37:39.289 --> 00:37:41.460
of the building design standard.
Nobody really knows what the

00:37:41.460 --> 00:37:44.680
building design standard is, but they
understand that’s kind of what’s been

00:37:44.680 --> 00:37:49.980
accepted is what the region has been –
the buildings have been designed for.

00:37:49.980 --> 00:37:53.079
So they can understand what that means,
so we give essentially a scaling

00:37:53.079 --> 00:37:55.670
related to the building design
standard for the region.

00:37:55.670 --> 00:37:58.660
And one of the interesting –
I think the precision that we use

00:37:58.660 --> 00:38:01.829
on the probabilities is challenging.
We caught a lot of – a lot of flak

00:38:01.829 --> 00:38:04.820
from the engineering community
by having precision like this,

00:38:04.820 --> 00:38:08.060
which I can understand.
But it just shows the balance

00:38:08.060 --> 00:38:10.980
that you have to have because we’re
presenting to multiple audiences.

00:38:10.980 --> 00:38:14.980
And as soon as we lose that precision,
then we lose the ability to communicate

00:38:14.990 --> 00:38:18.160
the decay in the aftershock rates
and the decay in probabilities.

00:38:18.160 --> 00:38:21.230
And that’s been a key thing that’s
been hard to communicate.

00:38:21.230 --> 00:38:24.900
It seems intuitive to most of us, but
it’s a very – it’s a pretty tricky thing

00:38:24.900 --> 00:38:29.020
for people to get their heads around.
So that’s – we still have some thinking

00:38:29.020 --> 00:38:33.700
to do there, but that’s one reason
and one reason it gets difficult.

00:38:34.500 --> 00:38:37.640
Sorry, this one’s hard to read, but this is
just essentially showing the same thing.

00:38:37.640 --> 00:38:40.020
But on here, this is
from Christchurch days.

00:38:40.030 --> 00:38:43.299
This is just keeping – so this was
our forecast, numbers of events,

00:38:43.299 --> 00:38:46.960
and this is just saying, during
that time, here’s what happened.

00:38:46.960 --> 00:38:50.730
So we say we forecast this range of events
to occur, and this is how many occurred.

00:38:50.730 --> 00:38:54.840
So just putting out there, trying to be as
open and transparent about how well

00:38:54.840 --> 00:39:00.400
the model is actually doing. And this is
one that seemed to get a lot of interest.

00:39:01.400 --> 00:39:04.320
Just another
communication.

00:39:04.320 --> 00:39:07.839
Similar sorts of MMI maps that
I think that are talked about here.

00:39:07.839 --> 00:39:12.390
So this is the probability
of MMI VII within 30 days.

00:39:12.390 --> 00:39:17.359
This was a Kaikoura map from
fairly soon after the event.

00:39:17.359 --> 00:39:19.769
And then having to – when you
get something as large as Kaikoura

00:39:19.769 --> 00:39:22.480
covering a very large region,
you have to give multiple regions

00:39:22.480 --> 00:39:24.319
for that so people
can understand.

00:39:24.320 --> 00:39:28.740
One problem we really had in Kaikoura
was we put out numbers of events,

00:39:28.740 --> 00:39:31.180
and this was over hundreds
of kilometers, and people

00:39:31.180 --> 00:39:36.579
really struggled just to distinguish
where they were and not assuming

00:39:36.579 --> 00:39:40.089
that that probability was the same
everywhere within the region.

00:39:40.089 --> 00:39:43.359
So trying to get information like these
maps out to people was important.

00:39:43.359 --> 00:39:47.391
It was also difficult in very busy times.
People don’t necessarily have the time

00:39:47.391 --> 00:39:50.190
to go to the web page, so this is where
the direct engagement with the

00:39:50.190 --> 00:39:55.240
stakeholders became very important to
be able to communicate such things.

00:39:57.880 --> 00:40:01.780
Slow slop probabilities. Sorry.
[laughter]

00:40:02.340 --> 00:40:06.980
This was the table that we
put out last year based on this.

00:40:06.980 --> 00:40:11.200
So this – as I said, direct engagement
to the government, to the media,

00:40:11.200 --> 00:40:15.249
GeoNet web pages, and direct
engagement with insurance and so on.

00:40:15.249 --> 00:40:17.980
So lots of interaction over
this type of information.

00:40:17.980 --> 00:40:22.260
And you may notice that one thing
we did here for the – this was the

00:40:22.260 --> 00:40:26.029
first time we did this – we actually
had an uncertainty on our uncertainty,

00:40:26.029 --> 00:40:28.609
which is an uncomfortable
concept for many.

00:40:28.609 --> 00:40:34.630
But it’s something I think – we felt was
important, and it actually was something

00:40:34.630 --> 00:40:38.490
that had been requested of us quite a bit
over the previous year was people

00:40:38.490 --> 00:40:42.560
wanted to get a feeling for what
our confidence was in the

00:40:42.560 --> 00:40:46.880
probability that we were putting out.
And this was one way to show it.

00:40:46.880 --> 00:40:50.180
It seems a bit odd since the probability
should, in theory, be incorporating your

00:40:50.180 --> 00:40:55.140
uncertainty. But it gets into when we
start including subjectivity and so on.

00:40:55.150 --> 00:40:58.660
But that actually seemed to
have worked reasonably well

00:40:58.660 --> 00:41:01.720
to include that probability,
but I think it’s something that we –

00:41:01.720 --> 00:41:05.470
or that range, but it’s something
we need to understand more.

00:41:07.820 --> 00:41:10.000
Now, into some
slightly different products.

00:41:10.010 --> 00:41:15.859
So this is – after Kaikoura, there
were 10,000 or more landslides.

00:41:15.859 --> 00:41:18.250
And there’s not a lot of people
in that part of the country,

00:41:18.250 --> 00:41:21.270
but there’s really just one road that
goes through it, and it’s the

00:41:21.270 --> 00:41:24.349
main highway through the country,
and it was shut down continually

00:41:24.349 --> 00:41:27.720
and lots of repair guys out
working and trying to remove

00:41:27.720 --> 00:41:30.930
the landslides and so on.
And they requested from us to

00:41:30.930 --> 00:41:34.869
get fairly – almost daily updates,
I think, at one point, of probabilities

00:41:34.869 --> 00:41:37.790
of exceeding certain thresholds
of shaking that would then

00:41:37.790 --> 00:41:41.530
trigger landslides.
So it was coming from the geotechs.

00:41:41.530 --> 00:41:45.359
So we were updating these maps
that just show the probabilities

00:41:45.359 --> 00:41:47.560
of exceeding those thresholds,
and that’s something they were using

00:41:47.560 --> 00:41:53.700
in their – in their daily planning of
their crews that were out on the roads.

00:41:54.900 --> 00:42:01.099
This is a few here related to
the buildings inspections going on

00:42:01.099 --> 00:42:04.789
in the Wellington region.
Again, this is post-Kaikoura.

00:42:04.789 --> 00:42:07.550
So the context here is
building design standards.

00:42:07.550 --> 00:42:13.740
So in New Zealand, if buildings are
less than – judged to be less than 33%

00:42:13.740 --> 00:42:16.940
of new building standard, they’re said
to be earthquake-prone, which means

00:42:16.940 --> 00:42:21.220
they need to be retrofit, and there’s
various regulations around that.

00:42:21.220 --> 00:42:24.450
But this is just showing –
again, focusing on the decay.

00:42:24.450 --> 00:42:28.869
This is the probability of exceeding
ground shaking equivalent to 33%

00:42:28.869 --> 00:42:33.599
new design level in a three-month
time period and showing that – how that

00:42:33.599 --> 00:42:39.099
decayed over the three-month period.
And this is actually showing the

00:42:39.099 --> 00:42:43.460
relative risk – so essentially your relative
exposure and how that decays over time.

00:42:43.460 --> 00:42:47.749
So the highest – at this point, it was
kind of 8 to 10 times exposure

00:42:47.749 --> 00:42:53.920
for the buildings in this region,
given those particular thresholds.

00:42:53.920 --> 00:42:59.460
And then, the next step on top of that
was then to look at that at different

00:42:59.460 --> 00:43:02.359
spectral periods, essentially so
people could understand, looking at

00:43:02.359 --> 00:43:04.800
different buildings, what it
might mean for a building.

00:43:04.800 --> 00:43:07.369
And I think you – just ignore –
the uncertainty is looking at different

00:43:07.369 --> 00:43:10.940
ways to get at the ground
shaking using different GMPEs.

00:43:10.940 --> 00:43:14.440
But you can – really, looking at this –
for this particular time period,

00:43:14.440 --> 00:43:20.559
they’re essentially about
10 times more exposed to

00:43:20.559 --> 00:43:24.010
exceeding that 33% standard
in the region.

00:43:24.010 --> 00:43:27.000
And there’s a lot of unreinforced
masonry and so on in the

00:43:27.000 --> 00:43:29.420
Wellington region that they
really were concerned about.

00:43:29.420 --> 00:43:32.859
So they were trying to get their heads
around what to do about that with the –

00:43:32.859 --> 00:43:37.109
the risk was obviously higher than
what was expected at the time

00:43:37.109 --> 00:43:40.490
of the building design standards
because of everything that was going on.

00:43:40.490 --> 00:43:43.660
And they were trying to make
decisions about what to do for that.

00:43:43.660 --> 00:43:48.230
In the end, we had a map like this,
which is essentially just showing the

00:43:48.230 --> 00:43:55.319
probability of exceeding that – it’s the
ratio again, but related to that 33% NBS,

00:43:55.319 --> 00:44:01.210
New Building Standard, and comparing
it to the pre-Kaikoura hazard.

00:44:01.210 --> 00:44:05.749
Used our statistical optimized model –
sorry, normal OEF model.

00:44:05.749 --> 00:44:09.119
And so, in short-term,
hazard increases of up to 10 times.

00:44:09.119 --> 00:44:12.589
And so they didn’t – they went –
New Zealand is a small place,

00:44:12.589 --> 00:44:13.990
and things can
happen pretty quickly.

00:44:13.990 --> 00:44:18.049
They can make decisions that
might not happen other places.

00:44:18.049 --> 00:44:21.170
But they had – right before this had all –
before Kaikoura happened,

00:44:21.170 --> 00:44:25.289
they had put – essentially for this
region in the map, they had put in –

00:44:25.289 --> 00:44:27.530
had a 10-year timeframe where
building owners had 10 years

00:44:27.530 --> 00:44:32.760
to retrofit buildings to get them above
kind of a 60% standard, I think it is.

00:44:32.760 --> 00:44:34.630
I might have that wrong.

00:44:34.630 --> 00:44:39.420
And then, based on this information,
a little over a year ago,

00:44:39.420 --> 00:44:44.950
they gave all the building owners
in this region one year to retrofit

00:44:44.950 --> 00:44:49.559
unreinforced masonry facades
and parapets. So a big change.

00:44:49.559 --> 00:44:51.630
But they also put some
government funding in to do that.

00:44:51.630 --> 00:44:56.190
And that year is – I think it’s just passed,
and last I heard was about six weeks

00:44:56.190 --> 00:44:59.040
ago, and they had actually very good
success with only a few buildings left.

00:44:59.040 --> 00:45:04.529
Where they had had problems in the
past getting this to happen, they actually

00:45:04.529 --> 00:45:07.470
got some of those changes made
fairly quickly in the last year.

00:45:07.470 --> 00:45:11.660
That was a fairly big
impact from the forecasts.

00:45:12.420 --> 00:45:15.650
The building design standard in
Christchurch at the time – well,

00:45:15.650 --> 00:45:20.260
it still is – had about a 35% increase
based on the time-dependent hazard.

00:45:20.260 --> 00:45:23.480
So using a longer-term model that
included the clustering information,

00:45:23.480 --> 00:45:27.100
they increased the building
design standard from – by 35%.

00:45:27.100 --> 00:45:30.200
And that –
the reason was fairly low –

00:45:30.200 --> 00:45:34.990
moderate to low design standard
in the country prior to that.

00:45:36.480 --> 00:45:40.100
And actually, the entire
Christchurch region moved to

00:45:40.109 --> 00:45:44.789
risk-based planning at that point.
So obviously lots of rebuilding and

00:45:44.789 --> 00:45:48.300
things that had to happen in
Christchurch and a new plan for that,

00:45:48.300 --> 00:45:52.269
and it was all risk-based where
kind of the most base layer of that –

00:45:52.269 --> 00:45:55.160
obviously lots of information
on top of that – was the

00:45:55.160 --> 00:45:57.339
time-dependent hazard model.
So they essentially took this

00:45:57.340 --> 00:46:01.900
forecasting information into
their long-term plan for the region.

00:46:03.180 --> 00:46:05.280
Last thing – just a
couple of slides left here.

00:46:05.280 --> 00:46:10.450
I talked about scenarios earlier on,
so that’s where we take the

00:46:10.450 --> 00:46:12.589
quantitative information that’s
coming from the forecast,

00:46:12.589 --> 00:46:17.019
and we try to put that into, I guess,
a bit more of a qualitative way

00:46:17.020 --> 00:46:20.260
for people to understand
what it might mean.

00:46:20.260 --> 00:46:23.510
So we started doing these in 2013,
and I think they’ve been –

00:46:23.510 --> 00:46:26.619
they’ve shown to communicate
very well. So they add this additional

00:46:26.620 --> 00:46:32.000
context – what the probabilities
mean in terms of the local tectonics.

00:46:32.760 --> 00:46:37.620
And they are directly tied
to the OEF probabilities,

00:46:37.630 --> 00:46:40.200
which can make things
challenging on how to do that.

00:46:40.200 --> 00:46:43.839
But we then provide related
New Zealand or global examples

00:46:43.839 --> 00:46:47.059
relevant to the current situation.
So it really requires having

00:46:47.059 --> 00:46:50.289
people around who know
what happened here 80 years ago.

00:46:50.289 --> 00:46:54.040
What sort of clustering did we have here,
and what happened from that?

00:46:54.040 --> 00:46:57.000
Requires a very interdisciplinary
team and quick to access,

00:46:57.009 --> 00:46:59.530
like I was saying,
to the historical knowledge.

00:46:59.530 --> 00:47:03.980
It obviously introduces
subjectivity to the forecasts.

00:47:03.980 --> 00:47:07.319
But it seems to be worth it
in terms of communication.

00:47:07.319 --> 00:47:10.770
They’re not easy.
They take a lot of time, as Sara will

00:47:10.770 --> 00:47:13.940
know, to get the wording right on these.
It can be quite challenging.

00:47:13.940 --> 00:47:17.860
And they can also be quite
confusing if care is not taken.

00:47:17.869 --> 00:47:19.849
But they’ve
been very helpful.

00:47:19.849 --> 00:47:22.259
And this is just a
very abridged version of one.

00:47:22.259 --> 00:47:25.180
They use a few paragraphs
on the web page.

00:47:25.180 --> 00:47:28.769
So we’ve always stuck with
three scenarios, where it doesn’t –

00:47:28.769 --> 00:47:32.059
it doesn’t work, depending on
what the probabilities are.

00:47:32.059 --> 00:47:37.789
But usually the first one is very likely.
So we have the descriptive words,

00:47:37.789 --> 00:47:42.559
which have been quite troublesome,
and there’s been a lot of debate.

00:47:42.559 --> 00:47:45.989
We have a large team of social
scientists and communicators at GNS,

00:47:45.989 --> 00:47:50.340
and it’s – a lot of debate
about the right way to handle that.

00:47:50.340 --> 00:47:54.220
But very likely – in this case, over 90%
within 30 days that you essentially

00:47:54.230 --> 00:47:57.999
just have the normal aftershock
sequence and everything goes away.

00:47:57.999 --> 00:48:01.400
And then very unlikely – so this
was near the megathrust here.

00:48:01.400 --> 00:48:07.009
So this is where you start to bring in
potential for a megathrust earthquake.

00:48:07.009 --> 00:48:10.940
So I think at this point, there was
very little awareness in broader

00:48:10.940 --> 00:48:16.349
New Zealand of what the Hikurangi
was and what that actually could do.

00:48:16.349 --> 00:48:18.239
So this is where we
started to bring that in.

00:48:18.240 --> 00:48:22.100
And then Scenario Three
was extremely unlikely,

00:48:22.100 --> 00:48:25.420
but that was of a large
event on the megathrust.

00:48:25.420 --> 00:48:27.240
So it really depends on
what earthquakes occurred

00:48:27.240 --> 00:48:30.680
and what historical
analogs we have and so on.

00:48:31.640 --> 00:48:34.120
So finally,
just conclusions.

00:48:34.120 --> 00:48:38.660
There’s been a lot of government
and society interest in what we’ve

00:48:38.660 --> 00:48:42.650
been doing, and I guess that’s really
forced us to beyond just the short-term

00:48:42.650 --> 00:48:44.650
weekly forecasting
into the longer term.

00:48:44.650 --> 00:48:47.040
Really, that’s where
the interest has been.

00:48:47.040 --> 00:48:53.200
Been a strong demand
for probability-based forecasts.

00:48:53.200 --> 00:48:55.660
And also for us to
include our best judgment.

00:48:55.660 --> 00:48:58.041
I guess, as people know here,
you get those questions all the time

00:48:58.041 --> 00:49:02.500
that goes beyond what we can
really objectively quantify.

00:49:03.100 --> 00:49:07.520
But we’re using tested models,
models that we understand, and

00:49:07.529 --> 00:49:11.660
including the uncertainty from that.
And the array of needs from

00:49:11.660 --> 00:49:15.499
end users has been massive
differences in what they want,

00:49:15.499 --> 00:49:19.970
how they’re able to understand it,
and their requests to us.

00:49:19.970 --> 00:49:22.339
We’re working more
with social scientists.

00:49:22.340 --> 00:49:25.980
Hopefully keep doing that. Better
communication – that’s the tricky part.

00:49:25.980 --> 00:49:28.760
And then working to this
nationwide OEF system.

00:49:28.760 --> 00:49:33.630
Hopefully be online and getting –
the risk is where – the hazard and risk,

00:49:33.630 --> 00:49:35.180
I guess, is where
a lot of the interest is.

00:49:35.180 --> 00:49:42.860
So getting a lot of those layers in
easy-to-understand and useful ways.

00:49:43.680 --> 00:49:46.100
So what we have
still is pretty simple.

00:49:46.100 --> 00:49:49.160
Our models may seem a bit complicated,
but the information in them

00:49:49.161 --> 00:49:52.631
isn’t all that complicated.
But it’s been very – proven to be

00:49:52.631 --> 00:49:59.499
very useful for lots of decision-making
in response mode and, I guess,

00:49:59.500 --> 00:50:03.900
maybe almost more importantly,
in post-event planning and so on.

00:50:04.600 --> 00:50:08.780
A key thing for this is
pre-planning for us was essential.

00:50:09.840 --> 00:50:11.960
All of this has taken a lot of time,
and I think we wouldn’t have

00:50:11.970 --> 00:50:14.789
been able to do a lot of this if we hadn’t
thought through a lot of these processes

00:50:14.789 --> 00:50:18.509
beforehand and made plans for how
we would be interacting with each other

00:50:18.509 --> 00:50:22.190
and within the – within GeoNet
and having all the relationships

00:50:22.190 --> 00:50:24.380
with the people around the country.
New Zealand is obviously a

00:50:24.380 --> 00:50:27.390
different place with not that
many people compared to here,

00:50:27.390 --> 00:50:29.630
but having those relationships
and trying to build them

00:50:29.630 --> 00:50:34.280
during the crisis makes
things a lot more challenging.

00:50:35.309 --> 00:50:38.700
And it consumes way more time than
I ever thought it was going to be,

00:50:38.700 --> 00:50:41.049
and that doesn’t just
end a week or two out.

00:50:41.049 --> 00:50:44.619
It goes on sometimes for – say, for
Christchurch, the biggest hit on the

00:50:44.619 --> 00:50:49.130
forecasting side was maybe a year out
when it was taking a lot of our time.

00:50:49.130 --> 00:50:52.489
Interesting situation with Christchurch,
but that was the situation.

00:50:52.489 --> 00:50:57.400
Data quality issues are a problem.
So the GeoNet catalog is great,

00:50:57.400 --> 00:51:00.470
but there’s lots of inconsistency,
and we don’t fully understand

00:51:00.470 --> 00:51:04.720
the impact of the data quality
on our real-time forecasts.

00:51:05.060 --> 00:51:09.020
Just finally, communicating
time-department probability

00:51:09.020 --> 00:51:12.300
information is tricky.
And that’s within the technical

00:51:12.300 --> 00:51:16.440
communities, so between ourselves,
between the risk scientists who are

00:51:16.440 --> 00:51:18.289
using our information, there is
often miscommunications,

00:51:18.289 --> 00:51:20.760
and things get used incorrectly
with engineers and with the

00:51:20.760 --> 00:51:24.640
government and the public.
So I think we need ongoing

00:51:24.640 --> 00:51:27.980
interaction between the
communities to understand

00:51:27.980 --> 00:51:32.520
impacts of decisions that get
made in the time-dependent space.

00:51:32.520 --> 00:51:37.180
Transparency of what we’re doing is
critical, and that’s a continual battle

00:51:37.180 --> 00:51:43.220
I think we’ll always be facing.
A quick response is always demanded.

00:51:43.220 --> 00:51:45.420
And that was it.
Thank you.

00:51:45.420 --> 00:51:52.280
[Applause]

00:51:53.000 --> 00:51:55.560
- Any questions for Matt?

00:51:59.840 --> 00:52:03.700
- Matt, yeah, one thing in passing
about the model never going

00:52:03.700 --> 00:52:07.789
below the background, that I think
I remember being discussed

00:52:07.789 --> 00:52:12.670
at some workshops a long time ago.
So that basically – maybe that’s valid

00:52:12.670 --> 00:52:17.040
because the background is declustered?
Is that true?

00:52:17.040 --> 00:52:21.480
Because otherwise, I mean, essentially,
if you have a background that truly

00:52:21.490 --> 00:52:25.710
represents the long-term hazard,
and then you increase it from time to

00:52:25.710 --> 00:52:30.240
time, you’re basically saying that your
long-term hazard was underestimated.

00:52:30.240 --> 00:52:33.890
I mean, there’s – if you increase – if you
increase – have increases in some times,

00:52:33.890 --> 00:52:38.759
then you have to have decreases other
times if the – if the long-term is right.

00:52:38.760 --> 00:52:41.060
But maybe it’s because it’s declustered.
Any comments?

00:52:41.060 --> 00:52:43.019
- Well, there’s a –
in the background, there’s a range

00:52:43.019 --> 00:52:46.799
of declustered and
non-declustered that’s in there.

00:52:46.799 --> 00:52:50.150
And I think you’d have to be careful
how long you took out – I mean,

00:52:50.150 --> 00:52:52.739
that’s in the context of the timeframes
that we’re looking at where it’s

00:52:52.739 --> 00:52:56.210
not allowed to drop below. But I think if
we started looking at longer timeframes,

00:52:56.210 --> 00:52:59.460
like – yeah, I think that’s not
something that we’ve solved yet.

00:52:59.460 --> 00:53:03.580
- I think if you look at, like,
the UCERF3 ETAS models,

00:53:03.589 --> 00:53:07.589
that has some plots where he has
the long-term average of the model,

00:53:07.589 --> 00:53:11.049
and you can sort of see
the swings on both sides of that.

00:53:11.049 --> 00:53:15.940
- Yeah. Yeah.
- But I think there’s also a societal –

00:53:15.940 --> 00:53:20.630
or, an applications push to not go below
the long-term background. [chuckles]

00:53:20.630 --> 00:53:23.440
To never tell people they’re
safer than they – than usual.

00:53:23.440 --> 00:53:26.480
- Yeah, yeah, yeah.
That would be hard to communicate,

00:53:26.480 --> 00:53:28.960
and I guess you couldn’t deal with it
in terms of standards and so on.

00:53:28.960 --> 00:53:32.800
But in terms of the
science side, yeah, I agree.

00:53:34.500 --> 00:53:36.360
- Any other questions?

00:53:37.720 --> 00:53:43.080
[Silence]

00:53:44.240 --> 00:53:46.500
- Nice, thanks.
Is this on?

00:53:48.120 --> 00:53:50.320
Okay. Oh, yeah.

00:53:50.320 --> 00:53:56.819
How has your process changed
since Christchurch to today in terms of

00:53:56.820 --> 00:54:01.080
how you’re putting – engaging
and organizing yourselves?

00:54:05.909 --> 00:54:08.660
- I think probably the biggest thing
is what I was mentioning at the end,

00:54:08.671 --> 00:54:13.999
and that’s in terms of relationships
of knowing who to talk to and the right

00:54:13.999 --> 00:54:18.150
channels to go through to get information
to people and knowing who the people

00:54:18.150 --> 00:54:21.460
are that are in
need of information.

00:54:21.460 --> 00:54:25.940
Because one of the – I think, initially,
we were largely just putting

00:54:25.940 --> 00:54:30.240
information onto GeoNet with
less of the direct engagement.

00:54:30.240 --> 00:54:34.480
And I think we learned, particularly
in the last five years, that, as I said,

00:54:34.480 --> 00:54:36.499
when things are busy,
people don’t have time.

00:54:36.499 --> 00:54:38.940
And people that need the information
aren’t going to the web page.

00:54:38.940 --> 00:54:42.200
So they’re hearing things secondhand.
So quite often, we were getting

00:54:42.200 --> 00:54:45.940
information back to us that was wrong.
You know, people had

00:54:45.940 --> 00:54:47.869
misinterpretation because they
were hearing it secondhand.

00:54:47.869 --> 00:54:50.329
So it’s – I guess that’s
been a large change is

00:54:50.329 --> 00:54:54.860
more focus on
the direct engagement.

00:54:55.820 --> 00:54:59.690
And I think, within – just within
our own – within GNS and GeoNet,

00:54:59.690 --> 00:55:02.980
knowing who the right people to talk to
are is really – we’ve become much more

00:55:02.980 --> 00:55:06.760
efficient about that and just knowing
what the decisions are that need to be –

00:55:06.760 --> 00:55:09.980
get made, what timeframes we need
to get things done in, and so on.

00:55:09.980 --> 00:55:12.380
Is that what
you’re asking?

00:55:15.460 --> 00:55:17.640
- Any other questions?

00:55:21.840 --> 00:55:24.240
- Thanks, Matt.
So I’m not very well-calibrated

00:55:24.240 --> 00:55:29.440
on EEPAS, but has the last two years
of relative quiet changed the

00:55:29.440 --> 00:55:31.619
sort of medium-term
forecast from EEPAS much?

00:55:31.620 --> 00:55:37.180
And what types of users are sort of tuned
into those changes and using them?

00:55:38.509 --> 00:55:46.720
- So it’s – do I have the – it reacts
on a longer time scale than that.

00:55:46.720 --> 00:55:51.390
So you won’t be – it won’t be seeing,
I guess, a change in – immediately

00:55:51.390 --> 00:55:53.359
from what’s – the quietness
in the last two years.

00:55:53.359 --> 00:55:58.320
So that will come out – I guess you’ll
see – as new forecasts start to come out,

00:55:58.320 --> 00:56:03.460
you’ll see decreases in, say, 20 years
from that for the larger events.

00:56:04.420 --> 00:56:08.000
In terms of users, it’s really the
regional planning, the long-term –

00:56:08.009 --> 00:56:10.859
the building codes,
for example.

00:56:10.859 --> 00:56:15.279
Some of the information based on the
earthquake probabilities and so on.

00:56:15.279 --> 00:56:18.210
Information is more on –
or, detailed more on that time scale.

00:56:18.210 --> 00:56:22.299
Not necessarily the financial decisions,
but the more kind of infrastructure-type

00:56:22.300 --> 00:56:26.099
decisions are where that’s
having the bigger impact.

00:56:26.500 --> 00:56:27.900
- Thanks.

00:56:29.080 --> 00:56:36.080
[Silence]

00:56:36.680 --> 00:56:39.080
- So I am following up
on Andy’s question.

00:56:39.860 --> 00:56:43.060
I have a hard time understanding
when we would say

00:56:43.069 --> 00:56:46.450
the hazard is below
the long-term average.

00:56:46.450 --> 00:56:52.300
Can you provide examples of when
you might think that was true?

00:56:53.980 --> 00:56:56.360
- I don’t – I don’t think I can.
I don’t think that’s something

00:56:56.369 --> 00:56:58.690
I feel like I understand.
It’s certainly not something we

00:56:58.690 --> 00:57:03.140
have good models of, I don’t think,
of when we can say that the

00:57:03.140 --> 00:57:05.479
rates have really decreased.
I mean, we could – if we talk about

00:57:05.479 --> 00:57:14.829
just the – related to Jeff’s question,
if we see – if we rely purely on EEPAS,

00:57:14.829 --> 00:57:17.749
there certainly would be long-term
fluctuations coming from that where

00:57:17.749 --> 00:57:23.340
we might see decreases that would – that
could potentially get below background.

00:57:24.300 --> 00:57:27.569
And so we’re talking about – in the
decision space, I would – I don’t know

00:57:27.569 --> 00:57:31.880
when I would be comfortable to say that
it actually has gone below background.

00:57:31.880 --> 00:57:34.670
I mean, maybe scientifically there might
be some evidence of that, but I think –

00:57:34.670 --> 00:57:38.380
I don’t feel like it’s something
we have that well-constrained.

00:57:39.880 --> 00:57:42.540
[Silence]

00:57:43.900 --> 00:57:46.280
- [inaudible voices]

00:57:47.100 --> 00:57:50.800
- Yeah. My question is about
short-term versus longer-term use.

00:57:50.809 --> 00:57:55.400
So I totally agree that’s what we saw in
Christchurch, but I also thought that part

00:57:55.400 --> 00:57:59.240
of the reason was that there was more
time to deliberate what was going on.

00:57:59.240 --> 00:58:01.300
Whereas, in the short-term,
people weren’t actually set up

00:58:01.300 --> 00:58:03.600
to use the information.
In fact, in Christchurch, they didn’t

00:58:03.609 --> 00:58:08.099
even get the information in time to use it.
And it sounds like, with Kaikoura now,

00:58:08.099 --> 00:58:11.200
you have got people who are primed
and ready to use the information

00:58:11.200 --> 00:58:13.150
on a daily basis
straight afterwards.

00:58:13.150 --> 00:58:18.920
And so do you think there’s still more
potential to use it in the short term?

00:58:18.920 --> 00:58:23.779
But it would take – what would
it take to get people ready to use it?

00:58:24.700 --> 00:58:27.220
- And this is just my personal
feeling here, but I guess when I –

00:58:27.230 --> 00:58:33.309
we’re talking about short terms
in terms of the length of the forecasts,

00:58:33.309 --> 00:58:36.359
so not in terms of related to the main –
so we can – people are using

00:58:36.359 --> 00:58:39.700
information pretty quickly
after the main shock had happened.

00:58:39.700 --> 00:58:42.739
But if we’re looking at 24-hour forecasts,
I think maybe even Andy had said

00:58:42.739 --> 00:58:45.829
something before, that people just
assume that tomorrow is going to be like

00:58:45.829 --> 00:58:49.279
today. And I think that’s probably the
level of information that they need.

00:58:49.279 --> 00:58:54.420
But when they’re starting to look a week
out and more out, that’s when they’re

00:58:54.420 --> 00:58:59.480
able to make – they didn’t have the same
feeling for what that would mean.

00:59:00.080 --> 00:59:05.640
So that’s where it started to become
more relevant, or where – I’m not sure

00:59:05.650 --> 00:59:09.129
I see it becoming shorter than that,
but maybe others disagree.

00:59:09.129 --> 00:59:14.320
- Well, it seemed like the Christchurch
engineers thought they could have

00:59:14.320 --> 00:59:20.200
used it when they were coming in
to do USAR and that kind of work.

00:59:22.040 --> 00:59:25.600
But they didn’t – they weren’t
getting it fast enough then.

00:59:25.600 --> 00:59:27.280
- Right.

00:59:27.280 --> 00:59:31.620
So, in Kaikoura, they –
yeah, it was – yeah, interesting.

00:59:31.620 --> 00:59:34.320
We didn’t really have requests
for that short-term information.

00:59:34.320 --> 00:59:37.170
We were putting it out,
but we weren’t getting any

00:59:37.170 --> 00:59:41.520
additional requests
on that in Kaikoura.

00:59:43.540 --> 00:59:49.320
[Silence]

00:59:50.580 --> 00:59:55.060
- General communications thing.
So I – I think I’ve made – I’ve never

00:59:55.069 --> 00:59:58.749
hidden my dislike for those terms,
“extremely unlikely,” “very unlikely.”

00:59:58.749 --> 01:00:00.380
[laughter]

01:00:01.140 --> 01:00:03.780
Is there a – I know you’re
sort of forced into them.

01:00:03.780 --> 01:00:07.520
Is there a chance of getting
out of that or killing them?

01:00:08.260 --> 01:00:10.980
- Yeah. That’s absolutely
not my area of expertise.

01:00:10.980 --> 01:00:14.750
Maybe Sara can comment on this,
but my feeling is there needs to be

01:00:14.750 --> 01:00:19.400
something there. Because so many
people just struggle with the numbers.

01:00:20.040 --> 01:00:22.920
- Yeah, it’s – like, I can think of
doing something relative.

01:00:22.920 --> 01:00:26.300
This is what I just thought of.
You could say “very likely,”

01:00:26.300 --> 01:00:28.780
“less likely,”
“even less likely.”

01:00:28.780 --> 01:00:31.220
You know, compare the
things to each other.

01:00:31.230 --> 01:00:33.499
The problem is, when you say something
is extremely unlikely – I suspect if we

01:00:33.499 --> 01:00:37.220
went into our long-term hazard models,
we’d be telling people, well, actually,

01:00:37.220 --> 01:00:41.750
earthquakes are extremely unlikely.
And I don’t think that’s, you know,

01:00:41.750 --> 01:00:47.640
a message – some of this is reaction to,
we didn’t use the words, but we did

01:00:47.640 --> 01:00:51.700
right before the 7.3 in Nepal,
say that the probability of

01:00:51.700 --> 01:00:55.140
a 7 or greater that
week was 1 in 200.

01:00:55.140 --> 01:01:00.940
And, you know, but we recognized,
because of the high hazard that involved,

01:01:00.940 --> 01:01:05.140
that we didn’t want to put
words on it that downplayed it.

01:01:05.140 --> 01:01:07.030
Anyway, I don’t know.
I just think we really need to

01:01:07.030 --> 01:01:08.560
work more on that because I …
- Yeah.

01:01:08.560 --> 01:01:12.390
- I just blanch every time I see those.
- So do I.

01:01:12.390 --> 01:01:15.510
The context for it is so important.
And everybody’s got different

01:01:15.510 --> 01:01:18.820
contexts for how they’re seeing
the numbers, so yeah.

01:01:18.820 --> 01:01:22.579
- So it’s like, another way to do it
is saying, extremely unlikely,

01:01:22.580 --> 01:01:25.340
but really important if it happens.
- Yeah. Yeah, yeah, yeah.

01:01:25.340 --> 01:01:27.140
- Balance that probability
against the risk.

01:01:27.140 --> 01:01:30.640
- Yeah. Actually,
I think one of the issues here

01:01:30.640 --> 01:01:33.890
is that social
scientists don’t all agree.

01:01:33.890 --> 01:01:39.460
And I actually don’t completely
buy into the way that the wording

01:01:39.460 --> 01:01:43.079
is framed because it
doesn’t provide context.

01:01:43.079 --> 01:01:47.039
And also, it doesn’t
provide perspective.

01:01:47.039 --> 01:01:50.560
The scientific perspective –
you, a seismologist,

01:01:50.560 --> 01:01:53.630
you look at these numbers every day.
So when you see a change

01:01:53.630 --> 01:01:57.630
by 2 or 3 percentage points of
something that looks quite dangerous,

01:01:57.630 --> 01:02:00.410
you’re going to say,
that’s a really big deal.

01:02:00.410 --> 01:02:01.869
And that might be
a different context

01:02:01.869 --> 01:02:04.809
for someone who doesn’t understand
that that’s a really big deal.

01:02:04.809 --> 01:02:10.900
So framing it for a one-size-fits-all for
all perspectives, I think, is maybe not

01:02:10.900 --> 01:02:15.720
the best approach. So I think that
there’s more to unpack there.

01:02:15.720 --> 01:02:17.700
Not to critique the social
scientists who did that work

01:02:17.700 --> 01:02:20.580
because I think they
did very robust work.

01:02:20.580 --> 01:02:23.390
But that’s one of the reasons
why I always like the scenarios better

01:02:23.390 --> 01:02:25.940
because you can compare
them to each other.

01:02:25.940 --> 01:02:27.940
You can compare the
three scenarios to each other

01:02:27.940 --> 01:02:34.039
in relativity if you want, rather than
just sticking to the numbers.

01:02:34.040 --> 01:02:38.420
So I think providing
context is really critical

01:02:38.420 --> 01:02:40.900
to the communication
of these forecasts.

01:02:43.980 --> 01:02:45.800
- Let’s cut off
the discussion there.

01:02:45.800 --> 01:02:48.040
But I believe Matt will be
around for today and tomorrow,

01:02:48.040 --> 01:02:51.780
and he may still have a few open slots.
So if you’d like to meet with Matt,

01:02:51.780 --> 01:02:53.860
approach either Sara
or Matt and organize that.

01:02:53.860 --> 01:02:55.760
Great. Thanks a lot.
See you Wednesday.

01:02:55.960 --> 01:03:00.520
[Applause]

01:03:04.100 --> 01:03:27.760
[Silence]