Can Earthquakes Be Predicted Algorithmically?

An anonymous reader with this story about a practical application of big data analysis as applied to the trove of sensor readings taken by satellites and by ground-based senosrs. A company called Terra Seismic says that earthquakes can be predicted 20-30 days before they occur, by sifting data for thermal, ionic, and other abnormalities in areas where quakes are considered likely. Says the linked article: "The company claims to have successfully predicted a number of earthquakes. For example, on 5th of April 2013, the firm issued a forecast for Japan. On 12th April 2013, an earthquake hit the identified area and 33 people were injured. On 4th June 2013, the firm again made a prediction for an earthquake in North Italy. On 21st June, an earthquake hit the identified area. On 3rd March 2013, the firm issued a forecast for an earthquake in Iran. Again, after 35 days, an earthquake hit the identified area."

Comments

Lists

[Anonymous Coward]

how about a list of times they predicted an earthquake and nothing happened.

without context its meaningless

Re:Lists

[BarbaraHudson]

Too much work. Just predict an earthquake in California within a week, every week. It'll be like predicting the sun rising tomorrow.

Re:Lists

[mspohr]

You win!
  Here's a map of California earthquakes for the last week:
http://scedc.caltech.edu/recen...

Re:Lists

[stephanruby]

how about a list of times they predicted an earthquake and nothing happened.

without context its meaningless

That's incorrect. They do give you all the context you need.

They say they can predict earthquakes between 20 days and 30 days before they occur. For the Japanese earthquake, they failed. For the North Italy earthquake, they failed. And for the Iran earthquake they failed. Three out of three failures, even taking into account their standard error rate of plus or minus 5 days, I'd say those failure rates are pretty significant.

Those guys should go into the earthquake insurance and earthquake protection business. It's as if someone was paying actively attention to their predictions, and was actively making sure that no earthquake would ever occur during any of the 10-days windows they predicted. Please don't tell me that's just a coincidence. Nobody can be that unlucky. The entire stock market should just follow these guys, and bet against them consistently.

Re:Lists

[Livius]

earthquake protection business

"Nice business you got here. It'd be a shame if an earthquake happened to it."

Re:Lists

[BarbaraHudson]

Other people have claimed the stats provided are useless if we don't have their failures as well. Missing the Nepal earthquake is one of those failures, since Nepal is also covered by the same satellites that they get their data from.

Plates can move because they are lubricated by superheated steam (which is why all the plates on Venus are locked in place). This is why tracking ground water levels is a good predictor. It's also why fracking increases earthquakes in the surrounding area. One way to monitor this from space is to look for small changes in local gravity caused by the "steam bubble" that slightly moves the ground up or down prior to the big energy release. They missed it. Totally. How many others of lesser magnitude did they miss? That's the test.

So, my (maybe too subtle) point was that without more complete data it's not all that great - like predicting an earthquake in California.

It's also pretty useless in most instances, since nobody's going to evacuate an area for a month or more based on a warning - again, about as useful as predicting an earthquake every week in California. Predict within a couple of days and the utility goes up. But predicting with an uncertainty of a month or more is about as useful to most people as a weather forecast that says it's going to rain some time in the next couple of months. Whether you're harvesting crops or deciding to bring along an umbrella, predictions with an accuracy of weeks or months are useless.

Re:Lists

[RockDoctor]

I don't know if they stated those predictions in public.

If it's not stated in public, then it is not worth the 30-m high letters in which it wasn't painted on the side of a building in down town Geneva.

Actually, there is a fair point there : there is no recognised forum for posting such predictions. And there are a lot of internet kooks out there who all think that they've got the perfect solution, but not one of them is willing to stand by a "prediction".

Some ground rules have been proposed about what constitutes a prediction. And then the kooks get involved and turn it into a kitten-in-a-washing-machine-with-a-broken-bottle experience for anyone who is in the least bit serious. Little things like : the prediction should be specific with respect to period to which it applies, magnitude of the earthquake predicted, and region that the prediction covers ; secondly, the prediction should be sufficiently precise that the pre-existing records for the area concerned would not predict that event just on statistics. "A magnitude 4.0 in southern California in the next 6 months" isn't a prediction, it's a racing certainty. "A magnitude 5.0 + on the eastern English Channel within the next 6 weeks" is a prediction (there was such an even in the late 17th century, IIRC, and hasn't been one since. So the occurrence of a predicted earthquake there would be pretty remarkable.) And finally, ALL your predictions need to be made public, and your method will be judged on the results of ALL of your predictions. (Some of the kooks use the "predict everything, everywhere, all the time" approach, and think that is effective.Your failed predictions will be counted along with your successes.)

Even getting agreement on these basic points - it's the kitten and the bottle into the washing machine again.

While all geological services are interested in such questions (including the BGS, in whose balliwick the Channel quake mentioned above falls), none of them see any reason for any general system to work. Why? Because they're geologists. As am I. So I can explain why they don't expect a generic system to work:

An earthquake occurs in a natural material which is inhomogeneous - in fact it has a structure that varies on scales ranging from the sub-millimetre to the multiple kilometre (I work at the sub-millimetre to sub-centimetre scale - people pay me to describe that inhomogeneity). The strength of such materials can be predicted in compression reasonably well - to within 20 to 50% ; but not so well in tension ; shear, combining tension, compression and structural homogeneity (absent - see above) is rather more difficult still. Earthquakes can occur because of either tension, compression or shear ; most often shear since it combines the others. Moving from the materials in which the failure occurs, consider the forces involved. They are, oddly enough, variable, because the distribution of forces depends on the action of large scale forces (weight, plate movements, tides, weather (including the last few centuries of rainfall and the last few minutes of barometric pressure)) which are delivered to the rock units that fail by a cascade of intermediate units, each one of which varies in stiffness (Young's modulus, for starters), in it's time variance of behaviour (some rocks "creep", others don't ; look up pictures of "chocolate boudinage", if you want to get a handle on how much rocks can vary) .. oh, and did I mention that the properties vary on scales form sub-millimetre to multiple kilometre?

So, how are we going to attack the problem. Clearly we need to map the rocks and the forces. But there is a problem. You see, rocks are generally opaque. Opaque to visible light ; opaque to anything with a shorter wavelength (and therefore able to measure the small scale variations), unless you can get the rock into a synchrotron beam or industrial X-ray machine. And them do the same for the next couple of millimetres, and then the next

And what of false positives?

[wbr1]

How often do their algorithms generate a false positive? If that rate is sufficiently high, the system is useless.

Even if it were used, people would likely begin to ignore it.

Re:And what of false positives?

[wbr1]

To add to my earlier comment (since I just read the crap article), not having an exact time limits utility as well, since people are not going to stay in shelters for 20-30 days.

I am half drunk and tired, can anyone find a better source than one that masturbates with buzz words?

Re:And what of false positives?

[NotInHere]

Both false positives and true negatives come with a cost. Calculate the probability with which a system is right, and you only have to do basic math to find out whether the prediction system gives you an economic or humanitarian advantage. As the humanitarian cost for false positives is very low compared to the economic one, it is very possible that there will be an unbalance between "most (economically or humanitarian) profitable strategies". Deciding between those can be I guess cause for some political debate.

Re:And what of false positives?

[Iamthecheese]

It limits utility but does not eliminate it. If the false positives are low enough people can make basic preparations. More importantly accuracy can be improved, possibly resulting in hours-long windows that will be extremely useful.

Re:And what of false positives?

[jklovanc]

If the false positives are low enough people can make basic preparations.

Anyone in an earthquake zone should already have made basic preparations.

Re:And what of false positives?

[im_thatoneguy]

It's not just economics it also strikes at the heart of whether or not the system actually *works*. If I create an algorithm that predicts an earthquake every 30 days in every major earthquake prone region I'll be right every single month because there is a major earthquake in one of the regions every month. It's like putting a bet on every horse in a race. You're guaranteed to win!

Sensitivity and Specificity

[ceoyoyo]

Sensitivity and specificity or it didn't happen.

Re:Sensitivity and Specificity

[oojah]

Baffled that this is modded as "Funny".

No.

[dskoll]

Betteridge's law.

Re:No.

[AK Marc]

We have an instinctive ability to do that math, an intuitive knack for maths and physics which is quite powerful. If we could apply it to school maths hardly anybody would ever fail.

Nope. It's not math. It's about feedback. You "guess" where it's going to land based on previous experience, and adjust as you see it coming. Try playing volleyball with a strobe light. You can see it, watch the path, but will have great trouble hitting it. Or take an adult with a math degree who hasn't played baseball. Hit a long fly ball towards them. They'll run in, as they'll feel like it'll fall short, but it will fly well over their heads.

Math is unrelated to catching. Catching is a feedback loop, combined with practice.

With what we know about how human brains intuitively work with basic physics

Nope. People don't have any intuitive knowledge of physics.

Re:Of Course

[dskoll]

Heisenberg says that you may never be able to get enough of the right type of information.

Quantum Mechanics says some things cannot be predicted.

Predictions

[CanadianMacFan]

Sure, I can say that I made predictions about earthquakes two years ago. I'd be impressed if they said today that in 10 days that there would be a magnitude 7.5 in a specific area and be right.

Re:Of Course

[TapeCutter]

Look up "n-body" problem for gravity, earthquakes are unpredictable for similar reasons.

Prediction vs forecast - the article gets it wrong

[wvmarle]

FTA:

With the recent Nepal earthquake claiming more than 6,000 lives, many of us have often wondered why earthquakes cannot be predicted the same way as Tsunamis or cyclones are predicted?

This already tells a lot on how much the authors of the article know about forecasting vs predicting - this opening line is wrong in so many ways. Tropical cyclones (which grow into typhoons aka hurricanes), tsunamis, tornadoes and other such natural events can not be predicted any more accurate than earthquakes.

Tropical cyclones can be predicted with a similar accuracy as earthquakes: this are the key areas, and they happen with that frequency. That's how much you can predict a cyclone to happen: Hong Kong is affected by about eight tropical cyclones per year, and about two a year will give rise to a T8 or higher signal. That's predicting: we've had years with five such typhoons hitting, and years without any hitting the city. When a cyclone forms (which is never predicted, only observed as it happens - like an earthquake is observed as it happens), meteorologists indeed are able to forecast with reasonable accuracy where it will head, and what strength it takes. This usually leaves a few days for people to react.

Tsunamis can be predicted with even less accuracy: when an earthquake or similar event has happened the presence of a tsunami can be measured, and a quick forecast can be made of when and where it will hit shorelines, and an alert may be issued. This leaves usually a few hours to half a day for people to react.

Tornadoes form without much warning, leaving often mere minutes for people to get out of the way and into shelters - if the alarms sound at all. They, too can not be predicted.

Earthquakes happen so fast, and end so fast, that there is nothing to forecast, no alarm to sound when it happens. By the time an alert is out, the quake is pretty much over.

And there we have the difference between prediction and forecasting. Forecasting is a lot more accurate by nature, as it is reacting to what is already happening, and works quite well for following slow processes such as the formation of a tropical cyclone. I'm used to know about an incoming typhoon a few days ahead, so plenty of time to prepare. Forecasting earthquakes, well, that doesn't work like that.

Re:Yes

[jbengt]

To be more accurate (or pedantic, depending on your point of view), only 20% of the deep well-injected waste water in Oklahoma is coming from fracking. The rest is coming from regular old traditional oil fields.