XPost: alt.ufo.reports

EXECUTIVE SUMMARY:
- We build the best possible model (the AI s/w can find) that
predicts US power outages.
- The model explicitly includes adjusted data from the NUFORC
sightings database.
- Statistical tests show while the model predicts outages from mostly
climate and other satellite data within +-10% the UFO sightings data
has no significant effect on outages.
- The nominal effect of sightings, that in any case shows as
"protective" (more sightings relate to slightly fewer outages), is
orders of magnitude less than annual wear and tear on the network
that shows up as stat significant in the model.


In previous posts we've looked at patterns in available data that
suggest ufo's may be flying defensively.

With an up-coming report to Congress (which may or may not be made
public at some point) supposedly including a detailed assessment of intelligence regarding possible threats posted by UAP we can do our
own assessment, given a smattering of knowledge of data science and
probability theory.

This and some subsequent articles will look at what evidence there is
that ufo's might pose any kind of threat. The preliminary executive
summary I can report here is -- mostly no. But there is some evidence
some kinds of mass animals deaths may be linked with increased
sightings. Not exactly a national security threat and perhaps
indicating a certain carelessness or lack of efficiency most (human)
societies are well acquainted with.

But here we'll look at whether there is evidence ufos have an
observable effect on national power supplies.

It's well known that some reports say ufos have been spotted hovering
over or in the vicinity of power lines. Other reports claim the
presence of ufos can be associated with loss of power to vehicles and electronic/electrical devices.

Can we see this in the data?

Short answer -- no.

The technique I'll describe below is one attempt to deal with a
"fringe topic" in science using AI techniques. In the case the AI has
simple ability to spot patterns from relatively small number of
examples as well as follow causal chains/trees and perform some
limited qualitative reasoning ("if you twiddle X then later Y changes").

The main advantage of AI versus RI is bias is potentially removed.
The algorithm proceeds to analyze all the data it has (my s/w at
present has a limited ability to google up new databases or facts to
test theories or assumptions it makes on the fly) to produce a
validated conclusion.

If we're asking a question like "does data X indicate an effect on
data Y" the s/w will attempt to build the best model it can to predict
the behavior of Y, such model including X. It can then robustly test
whether such a model actually shows X is necessary in the model, or
whether it seems to be involved simply by chance alone.

The model building involves all the data the AI has access to. As a
working data scientist I've collected quite a few data series over the
years. Many of them are from satellite surveys of surface temperature,
mass concentrations, sea height, plant color, etc. Some are curated
monthly series of global averages of various climatic data such as
atmospheric gas concentrations, humidity at certain heights in the
atmosphere, precipitation, snowfall or temperature, etc.

In all there are around 4,000 of them at this point.

So the exercise in this instance will try to build the best model
possible according to some algorithm and decide whether that model is
better WITH data on UFO sightings or not. If not then we might
concluded UFO data has little to do with whatever the model is
predicting.

So the UFO sighting data I'll use is a somewhat twiddled version of
the NUFORC database 1900-2020. "Twiddled" means certain biases are
ironed out according to standard algorithms and time discontinuities
in the series are taken into account. E.g. around 2006 NUFORC began
using an online report form that greatly changed the character of
sightings reports. I've therefore run the adjusted data through a
final phase that tries to present all sighting data as if it was
reported through the web reporting form. A statistically acceptable
adjustment appears to be multiplying total monthly sightings prior to
2006 by a factor of around 10.

The data on power outages come from a US DOE series available via
google spreadsheets. That was last updated in my database in 2017.

After some minutes of combining the ~4000 basic climate and other
variables together in many different ways and assuring the results
were statistically robust according to 3 different algorithms the best
model including the twiddled UFO sighting data turned out to have an "explanation power" of 64.5%. IOW about 65% of the month-to-month
variations in number of power outages across the US were predicted by
the model.

The final output from the "most robust" algorithm after all the mixing
and matching of variables was as follows:

REWEIGHTED LEAST SQUARES BASED ON THE LMS
*****************************************


VARIABLE COEFFICIENT STAND. ERROR T - VALUE P - VALUE
----------------------------------------------------------------------
date 1.27044 0.15410 8.24440 0.00000
x1 -0.00141 0.00329 -0.42995 0.66789
x2 -4.07119 0.45657 -8.91692 0.00000
x3 -7.19983 1.32116 -5.44964 0.00000
x4 1.13729 0.38024 2.99098 0.00329
x5 0.79549 0.20846 3.81610 0.00020
x6 4.00810 3.29246 1.21736 0.22552
x7 -2.69242 1.02530 -2.62597 0.00960
x8 1.02031 0.50537 2.01894 0.04540
CONSTANT -2299.93115 306.73410 -7.49813 0.00000

WEIGHTED SUM OF SQUARES = 2744.07031
DEGREES OF FREEDOM = 140
SCALE ESTIMATE = 4.42725
COEFFICIENT OF DETERMINATION (R SQUARED) = 0.64508
THE F-VALUE = 28.272 (WITH 9 AND 140 DF) P - VALUE = 0.00000
THERE ARE 150 POINTS WITH NON-ZERO WEIGHT.
AVERAGE WEIGHT = 0.94937

The model supposedly predicts the number of monthly power outages in a
given month within +-4.4 events. In "heavy" months this represents an
error of around 10%. (According to the DOE dataset Jul 2011 saw 51 outages).

The statistical tests are quite quite sure the model is predicting
power outages and not just guessing. The P-VALUE related to the
F-VALUE shows there is almost no chance this is just luck.

The "AVERAGE WEIGHT" produced by this s/w also shows about 95% of data
points run close to a straight line.

And, finally, around 65% of the month-to-month variation in power
outage events is predicted by the model.

Most tellingly the line labelled "x1" corresponds with the UFO
sighting data. Over in the P-VALUE columns associated with x1 is shows
around 67%. I.e. the odds are good the measured correlation between
UFO sightings and power outages is just a chance occurrence. In any
case, the BETA is nominally -ve at -0.00141 . I.e. for every 1000
sightings in a given month there is expected to be around 14 *less*
power outages than otherwise. If anything, the presence of lots of
UFO's seems to have a noisy protective effect on the US grid.

More worryingly, the "date" variable (in years + fraction due to
month) is positive. E.g. for each additional year the grid is showing
a trend of increasing power outages to the tune of around 1.3 per month.

Again, this argues the presence of how ever many UFO's represented by
the sightings data is having a smaller effect than the annual wear
and tear + repairs of the network.

While there may be some incidents that suggest otherwise, the overall
data shows UFO's are not generally associated with degrading the
performance of the US grid.

--- SoupGate-Win32 v1.05
* Origin: www.darkrealms.ca (1:229/2)