[HunchLab] Predicting Crime in Miami

Tim Beelen tim@diffalt.com
Wed Jun 24 09:31:00 PDT 2015

This is how it's pitched to the community:
> http://www.miamiherald.com/news/local/community/miami-dade/article19256145.html

FTA: The Miami police currently is using COMPSTAT, which does not 
predict where crime will happen insofar it tells you where it has been 
taking place. In addition they will start using HunchLab at some point.

An associate professor from Florida International University, Rob T. 
Guerette is expected to become their local expert on this piece of 
software. The person who wrote the grant for it is Lt. Sean MacDonald.
> http://cj.fiu.edu/people/faculty/rob-guerette/

The article claims that similar software has "... helped prevent and 
stop property crimes, and is now being tested on gun crimes."

Which makes me curious as about the kind of heuristics that they are using.

HunchLab apparently produces maps showing small areas where specific 
crimes are likely to be committed.

This is not a new turn of events, as Miami-Dade’s robbery division uses 
IBM's Blue PALMS to solve cold cases. The software connects to a 
database of every crime ever documented by Miami-Dade police. Detectives 
enter the details of an unsolved crime and the program produces a list 
of 20 suspects.

Now the part of the Nuevo Herald's article that cough my attention is 
the moment that it starts speculating about it's effectiveness.
The last few sentences juxtapose the potential volatility of it's 
predictions with who is responsible for it's "effectiveness".

HunchLab uses a wider dataset than the rough equivalent PredPol. 
Annotated with the official final statement that the tool will only be 
"...as good as the officers using it."

Putting the burden of proof of it's effectiveness squarely on the 
shoulders of the officers forced to use it. Which is, in my opinion, 
ridiculous. Since the software is supposed to predict the crime and not 
the other way around.

I highly suspect that it grabs a bunch of data, normalizes it for the 
use with a map and starts looking for some correlation coefficient and 
then looks if it's statistically significant. I.e. it's a null 
hypothesis-- exactly what Larry Samuels, CEO of PredPol tries to warn 
them about expecting weird results. Since that is exactly what you get 
when you grow your dataset and statistically analyze the date using 
correlation coefficients + phi.

Anyway. Who build this software? And what does it do?

More information about the cypherpunks mailing list