Magnitude-Dependent Automatic Alarm Model

[j-zhang]

Are we optimizing on the magnitude-dependent automatic alarm model? Is this considered the simple model?

As a first step, should we be trying to reproduce Bradley Luen's dissertation analysis / magnitude-dependent automatic alarm model? Section 5.4: Automatic Alarms and ETAS Predictability

[j-zhang]

My new understanding is that there are various models we could be looking at, including ETAS, automatic alarm, magnitude-dependent automatic alarm model, poisson. The simple Stark model may be referring to the simple automatic alarm strategy listed in the table on last slide of Prediction section in Professor Stark's slides.

Also, based on the slides, it seems like some steps mentioned were:

• "Replicate Luen's dissertation results; try to replicate the plots using etas.f and retas.f as well."
• "Choose periods for training and for testing our enhanced MDA approach" 
• "Frame the tuning of MDA as an optimization problem"
• "Develop an algorithm to do the tuning"

So we would likely be looking at multiple models and different subgroups within analyzers could be looking into each one.

[aculich]

These are definitely the kinds of questions we should be asking at this stage. I am going to forward this question to Prof. Stark for some feedback and to get his thoughts & guidance as our work begins to take shape over the next few days.

[pbstark]

Yes, the project is to optimize the magnitude-dependent alarm strategy by choosing something other than a power law for window length as a function of magnitude.

I've thought about how to start exploring what to consider. I think a good first step is to bin the catalog into magnitude ranges of 0.5, then plot the empirical CDF of inter-arrival times in each range of magnitudes.

A similar plot for the ECDF to the second event after an event, the third event after an event, etc., will also be instructive, for reasons given below:

I'm not sure whether we want to base the ultimate approach on quantiles of that interarrival distribution or not. For instance, it might be much more efficient not to turn on any alarm at all unless there's a sufficiently large event--rather larger than the minimum magnitude we are trying to predict. If so, we might want the window for big events to be long enough to be expected to catch several aftershocks, not just the first aftershock, since the aftershocks will not themselves open "alarm" windows if they are smaller than the threshold we are using.

Does this make sense?

Philip