Not only is training relative but apparently prediction evaluation
is quite relative as well. One thing I have been trying to do is to
acquaint the readership of this board with earthquake probabilities
and how to tell a useful from a meaningless prediction. I do not
intend to evaluate every prediction on the web, but in this case
we have a good example to deal with. I would like to compare the
prediction of "W" noted in 2-cents post above with the prediction
of Don for the Clear Lake area. I hope this comparison is somewhat
illuminating.

W's Prediction on another board:

The prediction from "W" was posted on October 24 at 15:44 (UT or PST? or
??) The prediction was for an earthquake of M 2.2-3.5 within 42 hours
(1.75 days) within 0.8 degrees (88 km) of 33.2N 117.9W.
First let us investigate the a-priori probability of "success" by
random chance given the current activity in the area at the time
of the prediction. From August 23 through October 24 there were 14
1.75 day windows which contained earthquakes which have matched the above
parameters of 34 total 1.75 day windows. The random chance of a
"success" would then be 14/31 = 0.45, about one chance in two of
a random success.
But the question is was this a successful forecast for the Compton
earthquake? The Compton earthquake occurred on October 28 at 08:27 PST
with Ml 3.8 at 33.922N 118.270W. This is 88.75 hours after the forecast
or 3.7 day later, about 2 days outside the forecast range; it is 0.8
degrees from the predicted epicenter and out of the expected magnitude
range by 0.3 units. This is a miss on at least two counts and very
close to a miss on the distance. In order to have had a "success"
we would have to extend the magnitude range to 3.8 and the time window
to 3.7 days. What is the random probability of an event in the "success"
range now?
There are still 14 windows (now of 3.7) days length in which a
successful quake occurs, but now there are only 17 possible windows
of 3.7 days length. The probability of randomly choosing a window
in which a successful quake occurs is then 14/17 = 0.82 or a success
in about 8 of 10 attempts. Pretty good odds if your the betting type.
This is one of those predictions which is somewhat correct and
looks good if something happens, as it nearly always will, but is
pretty useless otherwise.

Don's Clear Lake California Prediction on this board:

I would like you to compare this with the odds on the prediction
Don made for Clear Lake California. The prediction (Don prefers
the word forecast) was made Oct. 22 for a region within 40 km of
38.9N 122.7W for Ml 2.8-4.0, specified to occur on Oct 29 but
possible from Oct 27 - 31.
In this case the following earthquake is listed for this afternoon:

ML 3.1 2001/10/30 11:38:34 38.771N 122.732W 2.5 ESE of The Geysers, CA

In the same 61 day period (August 22-Oct 22) there were 3 five-day
periods which contained a "successful" earthquake which met all
criteria. Since there were 12 5-days periods in this time the probability
of a random success within 5 days is 3/12 = 0.25 or odds of about
1 in 4. On the other hand since the earthquake occurred within 1 day
of the stated date, the odds of a success within this time frame
are about 3/30 - about 1 in 10.
Was this a successful prediction, and was it of any use. The
earthquake lay within 14 km of the expected epicenter, within
the time window and within 1 day of the expected date of occurrence
and within the stated magnitude range and the prediction beat the
random odds considerably. Such a prediction could be
useful if a history of repeated successes of this type were available.