Location of Annie Millwood's attack

[JeffHamm]

Sigh. This is why one should take a pause when programming and developing routines before posting all excited like!. Detected a few errors in the calculations, but I think I've got them sorted now. There's still a few weights that need testing in a slightly more complicated set of routines, but I've gone with the simplified one here (they often produce very similar outputs, but under certain circumstances, will differ quite a bit. For example, in the C5 new version, it ends up with zone 1 in the old versions lower of the two eastern zones. This is a pattern that does occur sometimes with spatial layouts like the C5 (where they are fairly equally distributed around a circle). I'm not sure I've got the values quite right, though, as this is also a less common pattern, and it may be that the weightings are too high (this is part of what needs calibrating, and it's quite likely I've got the sensitivity too high). The less complicated version isn't as sensitive to this, and it's even possible is the more reliable routine of the two in general). Anyway, I'll stop making excuses and trying to deflect responsibility.

So, while this is still a work in progress, here's the updated and properly working versions of the new algorithm. It's now focusing in more on one region, and the basic pattern isn't changing as dramatically between the choice to include or exclude Millwood and Tabram (zone 1 shifts a bit north when they are included). Sagar's suspect remains a high priority suspect, and Levy is also in Zone 1 for the C5, and Zone 2 for the C5+2, so he would be considered high priority for further investigation. Barnett, for those who favour him, has moved into Zones 2 and 3, again, making him a top priority (which, of course, he was at the time and the police were satisfied of his innocence, but I know there are those today who disagree with that). The Graffiti is in a high priority search zone (2 and 1), and Hutchinson moves from 19 to 9 for the C5, and from 8 to 3 for the C5+2 (again, once he introduced himself into the investigation, his location, being so close to Zone 1, would boost his priority. The police spent a fair bit of time with him of course, but I'll leave it for you to decide if that means he was cleared or overlooked. The rest, while they jiggle around a bit, don't make any big changes.

Anyway, sorry for the mistake above. I've got another bit to build in at some point, which may or may not improve the performance of the predictions. I've still got to work out the proper weightings for some of the bits I've added, and will need to do an overall accuracy test soon (which takes a long time to run, so I've been putting that off while I add new routines)



And the updated results:

Geographical profile outputs:
Suspect ..............Zone (C5 / C5 + Tabram & Millwood / new C5/ new with 7)
Sagar’s Suspect .........( 1 / 4 / 1 / 4 ) - And the police were actually watching someone here!
Levy ........................ ( 4 / 2 / 1 / 2 ) - I know nothing of this suspect other than I spotted a post on them
Donston/Hospital ...... ( 6 / 15 / 11 / 12 ) - This also fits the "mad doctor/medical student" ideas
Druitt / Ludwig ......... ( 7 / 15 / 12 / 19 ) Druitt is suggested to have had access to Dr. Thyne's surgery, and given his cricket schedule, seems almost ruled out; Ludwig was ruled out as he was in custody on the double event
Bachert ................... ( 7 / 12 / 9 / 13 ) - Another suspect I know nothing about, other than he was on the vigilant committees and was a bit of a nuisance to the police at the time.
Gouston Str. Graffiti ..( 8 / 1 / 2 / 1 ) - this isn't a suspect, but we know JtR was here at some point after Eddowes - was he near home?
Barnet .................... ( 9 / 5 / 2 / 3 ) - Using Kelly's location
Tumblety ................ (15 / 31 / 15 / 23 ) There is some question as to whether he was here in 1888 or not
David Cohen ........... (17 / 17 / 17 / 17 )
Peabody House ....... (19/ 8 / 15 / 10) (there's no suspect here; I mislocated Hutchinson here before so I leave it in)
Hutchinson ............. (19 / 8 / 9 / 3)

These POIs fall in zones that would be considered pretty much "excluded" except * - but the suspect wasn't living there at the time:
Pizer ....................... (32 / 38 / 23 / 26) - Pizer was cleared (although officially identified as Leather Apron)
Klowoski ................. (34 / 21 / 23 / 14*) - Chapman, however apparently this wasn't his address in the Autumn of 1888 and he was on Cable Street, just south of the map (guessing in a 40+ zone)
Kosminski ............... (35 / 44 / 26 / 31)
Kaminsky ................ (40 / 53 / 28 / 37)

[Sam Flynn]

40+ zone for Klosowski sounds about right to me

[JeffHamm]

And finally, to update the version of the C4 (C5 without Stride). This shifts the focus from Sagar's suspect (now Zone 12, as is Levy) to Barnett (he's just in Zone 2, but there are slight penalties for being too close to a crime scene, which you can see as there's a bit of a dent in towards him; remember, you don't need a spatial profile to locate suspects or POIs who live with the victim, this is a tool that is used after excluding the usual suspects, like family members, etc; otherwise, the probabilities would just be spikes right on each crime location, which is sad, but true). The hospital drops to zone 16.

Anyway, I won't list all the suspects as many end up in areas which would be low priority search areas.

- Jeff

[JeffHamm]

40+ zone for Klosowski sounds about right to me

Well, remember, about 20% of the time the killer is a commuter. Zone 40 would be Zone 1 for commuters But, to be fair, the routines are not calibrated for actual commuters, and they have far fewer known predictors to use, and the relationships between how they spread their crimes out once they get to their crime zone becomes much less indicative of where the started their journey from. If JtR was a commuter, and traveled to Whitechapel from a distance, none of these are going to be informative of much other than, perhaps, indicating places to look for him when he's around, but not necessarily where he resides or works (the anchor point could be his work, his local pub, etc; quite often, though, it is the residence).

- Jeff

[JeffHamm]

So much for finally. I've toned back some of the weights in the more complicated routines and have redone the C5. This is looking closer to what I sort of expected it to do, which is, start to split the area of high interest between close to the bulk of the crimes, but there is another pattern slightly less common, which is for the offender to be closer to the less dense end of the crime zone (Dennis Rader, for example; the bulk of his crimes were quite far south, with two in Park City, where he lived.) The JtR pattern is a bit similar in some aspects to how Rader's crimes were distributed, just more east-west rather than north south. So, seeing a zone to the east start to pop up makes sense as the analysis should draw attention to both possible patterns. Previously, it was zooming in on that area at the expense of the more common "dense end" zone. I was thinking it would be the one to the north, closer to Nichols, but that's dropped down to a zone 3 peak.

It's late here now, so I won't detail the suspects in the format above, and just list them here: Sagar's suspect and Levy are in Zone 1, the Grafitti is in zone 2, the Hospital tucks into zone 4, Barnett is in zone 5, Bachert zone 9, Hutchinson and Tumblety zone 11, Druitt zone 14, David Cohen zone 17, the Peabody house zone 20, Pizer 22, Klowoski 23, Kosminski 25, and Kaminsky 28.



- Jeff

[Sam Flynn]

...split the area of high interest between close to the bulk of the crimes, but there is another pattern slightly less common, which is for the offender to be closer to the less dense end of the crime zone (Dennis Rader, for example; the bulk of his crimes were quite far south, with two in Park City, where he lived.)

Rader used a motorised vehicle, though I note that you said in your previous post that ~20% of killers are commuters, but I wonder whether that would apply to the 19th century, when access to private transport was significantly less common than it became in the latter half of the 20th? It would be fascinating to know how the "catchment areas" of Victorian killers compared to those of their more mobile successors.

[JeffHamm]

Rader used a motorised vehicle, though I note that you said in your previous post that ~20% of killers are commuters, but I wonder whether that would apply to the 19th century, when access to private transport was significantly less common than it became in the latter half of the 20th? It would be fascinating to know how the "catchment areas" of Victorian killers compared to those of their more mobile successors.

Yes, Rader had access to a car, but that just means the scale changes, he travels in kilometers (or miles, since he's in the US) and JtR works in meters/yards. The routines I've been working on use the crime pattern itself to deal with patterns invariant of the scale of physical measurement, so it recognizes a pattern whether the offender is travelling by foot, bike, or car, etc. Of course, I'm working with modern crimes, so most have access to vehicles, so your concern is not without merit and is something I eventually want to ensure is a safe assumption to make.

And remember, commuters are just offenders who reside outside the crime zone, which is defined as the smallest circle that encloses all the crimes (although this is often then expanded by increasing the radius of that circle by 10-25% on the assumption that the estimated crime zone is a bit smaller than it really is). A quick and dirty, but quite accurate, way of estimating this is to just take the two offense locations that are most distant from each other, and use them to define a circle (which one could expand 10-25%). Basically, an offender that lives outside the circle is a commuter and those who live inside are called "marauders" - the majority of offenders are marauders. Again, if you're travelling by foot, the crime zone will be smaller, so commuters will be closer. I tend to expand by 25% because "near commuters", those juts outside the crime zone, show similar patterns to definite marauders. Things start to break down after that, at least with the data set I have to work with.

That said, there could very well be differences in the percentage of commuters / marauders between now and Victorian times. If I were to guess, I would think commuters are more common now, not because they have cars etc (that just let's them expand the scale of their crime zones) but because crime detection is no longer a new thing, and commuting is an attempt to hinder investigation.

- Jeff

[JeffHamm]

Hi,

Ok, I've been working on some new ideas with the geographical profiling routines that I'm developing. What I do is analyse a bunch of cases I have, which are about 30 or so cases of serial arson in New Zealand. To be honest, serial arson is a pretty hard type of crime to geographically profile because a big part of the spatial decision making is not up to the offender (vacant lots, abandoned buildings, schools, etc, are limited in choice from which to choose). Also, I never expected the equations that I've been developing based upon arson to translate to other types of serial offending (serial rape and/or serial murder). However, as maps of those crimes were relatively easy to find on the net, I've had a go at checking out one or two, expecting it to show that "crime specific values" are required (so while the approach might work for serial rape and/or serial murder, the specific equations would be slightly different, and one would need a set of those types of crimes from which to obtain the values; another reason I started building up a collection of them). Surprisingly, though, the same routines that I've been using the arson cases to develop the routines seem to work quite well. I've been able to obtain the information I need from 8 cases of serial rape and/or murder. These are not used to set the parameters, but to test the values I obtain from the arson cases (otherwise, it's not really a test of generalization). And, yes, if you're wondering if that's perhaps a bit small for a proper test sample size, you are absolutely correct. But, for now, it's what I've got. Eventually I hope to continue to add more cases to this, which will give a better idea of how well it's doing.

Anyway, now that I've programmed in the new bits, there were a number of options to explore in terms of which model produced the most reliable set of results, and I've finished that with my 8 test cases (the one listed as "Canter & Larken, 1993) is from a paper on spatial crime scene analysis, and is a pattern of offenses during a serial rape case in the US, but I know nothing more about it. The good news is that the additions have improved the results, with 2 cases not changing zones (both were in zone 1 before, and still are), 5 cases improving between 1 and 6 zones (average 3), and only one case dropping a zone (DeAngelo went from zone 2 down to zone 3). Here's the table of results, the %Searched is basically just how much of the "chance search space" would you have to search if you simply followed the geographical output exactly. The 50% and 75% are just indicating the point where 1/2 and 3/4 of my test cases fall (so, 3/4 of the test cases fall in or below zone 7). All of them were located better than a random search, so that was encouraging (and the Phoenix case I had the map for since 2016 or so, but they've only recently made an arrest and I found the address in the court documents online. It was good to see he falls in the red zone as this was a new addition to the test cases (he also has an unusual pattern, so I'm not actually disappointed that he's in a fairly high zone - it's still better than chance, and he is unusual. Also, DeSalvo has tended to be elusive to the routines up until now, generally falling in zone 20 in my previous best models (Lonnie Franklin, the Grim Sleeper, also tended to be lowish, at zone 12 in the previous models).

So, here's the performance of the current best model I've got on solved cases, all far more recent than the JtR series. After that, I'll put this new model's profile maps for the C5+2, and the C5, and the C5 - Stride. I'll list the POIs from above, and their zone numbers, in that order. I'm going to re-arrange them based upon their C5 zone rankings if things change, so they might be in a different order from before. I'm writing this as I'm doing these analyses for the first time with the improved models, so it may be that they look pretty much as before (in fact, I'm not expecting any dramatic changes, but there could be some shifting about):

			Zone	% Searched
Dennis Rader (BTK)			1	3.01
Ted Bundy (offenses )			1	4.21
Canter & Larkin, 1993 figure			1	4.5
Joseph James DeAngelo			3	13.92	50%
Bruce McArthur (Toronto)			6	27.73
Lonnie Franklin (Grim Sleeper)			7	33.93	75%
Albert Desalvo (Bos. Strangler)			14	69.22
Phoenix Serial Shooter			15	73.7
------------------------------------------------------------------------------
	mean		6	28.78
	sd		5.73	28.68
	Median		4.5	20.83
	worse than chance		0

New Geographical profile outputs:

Suspect ………..….Zone (C5+2 / C5 / C5-Stride)
Sagar’s Suspect …………( 4 / 1 / 14) - And the police were actually watching someone here!
Levy ........................ ( 2 / 2 / 14 ) - I know nothing of this suspect other than I spotted a post on them
Gouston Str. Graffiti …. ( 1 / 3 / 5 ) - this isn’t a suspect, but we know JtR was here at some point after Eddowes - was he near home?
Barnet ...................... ( 3 / 3 / 2 ) – Using Kelly’s location

{ This is the end of the 50% cut off POIs --------------------------------- }
Donston/Hospital ………. ( 18 / 7 / 19 ) - This also fits the “mad doctor/medical student” ideas

{ This is the end of the 75% cut off POIs --------------------------------- }
Druitt / Ludwig ........... ( 16 / 9 / 22) Druitt is suggested to have had access to Dr. Thyne's surgery, and given his cricket schedule, seems almost ruled out; Ludwig was ruled out as he was in custody on the double event
Bachert ..................... ( 12 / 9 / 50) - Another suspect I know nothing about, other than he was on the vigilant committees and was a bit of a nuisance to the police at the time.
Hutchinson ................ ( 3 / 13 / 10 )
Tumblety ................... ( 30 / 16 / 54) There is some question as to whether he was here in 1888 or not
Peabody House ……………( 9 / 17 / 8) (there’s no suspect here; I mislocated Hutchinson here before so I leave it in)
David Cohen .............. ( 18 / 19 / 50)
{ ----------------------------------------------------------------------
These POIs fall in zones that get considered pretty much “excluded” except *:
------------------------------------------------------------------------ }
Klowoski .................... ( 13* / 26 / 49) - Chapman, however apparently this wasn’t his address in the Autumn of 1888 and he was on Cable Street, just south of the map (guessing in a 40+ zone)
Pizer ......................... ( 29 / 27 / 70) - Pizer was cleared (although officially identified as Leather Apron)
Kosminski .................. ( 35 / 30 / 68 )
Kaminsky ................... ( 40 / 32 / 51)

And the maps (interesting, the C5 map still has an easterly zone, but it's now the one up closer to Nichols, not the southern one. I tended to think the one that now remains made more sense in some ways (close to first in the series, and so forth), so I'm a bit chuffed to see it's now back:



- Jeff

P.S. No doubt, things will change again at some point.

[JeffHamm]

Hi again,

I've programmed up routines to do geographical profiling based upon Rossmo's routines, which is used in actual police investigations and has contributed useful information to a number of cases, although at times, like any model of complex behaviour, doesn't get it right every time (while that would be nice, human's are a bit more complicated than that). Rossmo's approach, and he was the first to really bring the idea of using offense locations to locate offender's "anchor point" into a usable format, can be thought of as making predictions about the minimum distance an offender has traveled to commit an offense. He also uses "Manhattan Distances" (which is based upon a square rather than a circle, like Euclidean distances), which is why you can see square hotspots around some of the offenses. I don't have time to summarize the suspects just now, but will try and do that later. When I compared my routines with Rossmo's on the cases above and a couple others that I have (another zone 1, and a zone 22 case, for both approaches), there wasn't a dramatic difference in the solutions for most cases (Rossmo had "better zones" for 3, there was no difference in zones for 2, and I had better zones for 5, but in terms of absolute percentage of area searched, it was pretty much an even split as Rossmo's approach had the two tied zones slightly higher in the zone). Rossmo's approach was better with DeSalvo (Boston Strangler) but missed quite badly with the Phoenix Case, and that's a case where the offender's minimum distance is much larger than in the other cases. What I'm trying to do is base the predictions on more predictors, and not just rely on the accuracy of the minimum distance estimation. So, I'm actually pretty pleased with how this project is coming along as my limited performance tests are showing that these routines are at least as good as what's used, and at the moment, shows a marked improvement with some of the cases where current routines will miss things.

Anyway, below are the results for the C5+2 (Millwood and Tabram), C5, and the C5 without Stride. The only comment I'll make is that Sagar's suspect and Hutchinson both look to be in high priority zones; meaning, they tend to fall close to the estimated minimum Manhattan distance from an offense (or a local cluster of offenses).

- Jeff

[Ginger]

My apologies if this has been dealt with already, and I've overlooked it, but I'm curious - what happens if the Goulston Street Graffito is treated as a crime scene? Obviously it wasn't a murder, yet if Jack wrote it, he was standing out on a public street doing something that could easily have put his neck in a noose (as he tended to do). If it's his work, then I'd have to think that the same factors that operated in his choice of murder scenes would have operated in his choice of a place to write his message.

[JeffHamm]

My apologies if this has been dealt with already, and I've overlooked it, but I'm curious - what happens if the Goulston Street Graffito is treated as a crime scene? Obviously it wasn't a murder, yet if Jack wrote it, he was standing out on a public street doing something that could easily have put his neck in a noose (as he tended to do). If it's his work, then I'd have to think that the same factors that operated in his choice of murder scenes would have operated in his choice of a place to write his message.

I've not done that in any of these, so you've not overlooked it. There are things to consider when entering secondary locations of interest, such as, once we enter the Graffito location, we know it's tied to the Eddowes case, so we're in a way weighting that case more than the others. But, if we include Stride, we also know those two events are tied together (meaning, if we include Stride in the series, then the fact that JtR then heads west to end up at Eddowes' location is a decision that is in part based on being in Stride's location to start with. In some ways, that could complicate the inclusion of Eddowes as that location could be less representative of the offender's anchor point (meaning, his home usually). If we include the Graffito, we are sort of doing the same again, but potentially distorting things more because Eddowes' location is already a bit contaminated. That being said, I don't think the distortions are really all that large in a practical sense most of the times. I think there is potentially a lot of information from direction of travel (Stride - > Eddowes -> Ghoulston street, shows parts of the path the offender chose to travel, combining that knowledge with the profile map is part of interpreting the maps.) Anyway, I've done the C5 analysis with the Graffito included, which you can compare with the previous C5 map (not the Rossmo versions). I've listed the suspects below, the old C5 zones first, and the new ones 2nd. I've reordered the list to correspond to this map (so the suspects are listed in order of priority as per the map below:



Geographical profile outputs:
Suspect …………Zone (C5 / C5 + Graffiti)

Gouston Str. Graffiti ( 3 / 2*) – this isn’t a suspect, but we know JtR was here at some point – was he near home? – in the 2^nd analysis this is entered as a geographical data point

Levy ........................ ( 2 / 2) - I know nothing of this suspect other than I spotted a post on them
Barnet .................... ( 3 / 2)
PC Sagar’s Suspect ..( 1 / 3) - And the police were actually watching someone here!
Hutchinson ............. ( 13 / 6)
Peabody House ……. ( 17 / 9) (there’s no suspect here; I mislocated Hutchinson here before)

Bachert .................. ( 12 / 12) – Another suspect I know nothing about, other than he was on the
vigilant committees and was a bit of a nuisance to the police at the
time.
Druitt / Ludwig ....... ( 9 / 13) Druitt is suggested to have had access to Dr. Thyne's surgery, and
given his cricket schedule, seems almost ruled out; Ludwig
was ruled out as he was in custody on the double event
Donston/Hospital ... ( 7/ 17) – This also fits the “mad doctor/medical student” ideas
David Cohen .......... ( 19 / 18)
Klowoski ................. (26 / 18) – Chapman, but he was on Cabel Street, moved here after JtR series

Tumblety ................ (16 / 30) There is some question as to whether he was here in 1888 or not
Pizer ....................... (27 / 32) - Pizer was cleared (although officially identified as Leather Apron)
Kosminski ............... (30 / 35)
Kaminsky ................ (32 / 36)


- Jeff

[Sam Flynn]

what happens if the Goulston Street Graffito is treated as a crime scene? ... If it's his work, then I'd have to think that the same factors that operated in his choice of murder scenes would have operated in his choice of a place to write his message.

Happily, we can ignore the controversial graffito in this context, I think, because the apron piece alone indicates that the killer was there at some point.

[JeffHamm]

Happily, we can ignore the controversial graffito in this context, I think, because the apron piece alone indicates that the killer was there at some point.

Yes, the question is, when did the apron get there? Did JtR drop it as he fled, and it was just not spotted for an hour or so, or did he bolt home, and then drop it there later, after having made it home and was now out on the streets again?

- Jeff

[Sam Flynn]

Yes, the question is, when did the apron get there? Did JtR drop it as he fled, and it was just not spotted for an hour or so, or did he bolt home, and then drop it there later, after having made it home and was now out on the streets again?

Personally, I think he jettisoned it en route from Mitre Square to the safety of his "home". Either way, it's probable that Goulston Street was handily placed for him.

[JeffHamm]

Personally, I think he jettisoned it en route from Mitre Square to the safety of his "home". Either way, it's probable that Goulston Street was handily placed for him.

I tend to lean towards that as well. It does appear to be taken to wipe his hands and knife, etc, and I think he just tossed it when done. I'm not convinced he stopped to write on walls.

- Jeff