Super excited to see this post, specially as I've just spent the last week testing the library, and it's been awesome.
Also, I can attest that it can get quite computationally expensive, so the upcoming examples will be quite welcome.
Currently I've been running the test data (guilford-nc, downloades from Hugging Face) locally on my laptop. What are the hardware requirements (RAM, CPU) that would be recommended for running the test data?
We haven't really established that lately, but it's something we should nail down. One major thing I want to do is to figure out what enrichments are actually just kind of optional and and not all that important in most cases, so we can trim down the "default path" as much as possible.
Street frontage enrichments, for instance, are particularly expensive compared to the predictive power they yield. Also, we currently are generating horizontal cluster ids for three categories (general, land-only, improvement-only) and the second categories aren't currently being used for much, so we could do some trimming there too.
Also, when training your predictive models, I think there are some strong best practices--like you probably don't need to run full parameter search on all three gradient boosted trees (LightGBM, XGBoost, Catboost). Probably best to just start with just LightGBM and see how that performs, then spend time on your dataset and variable selection before worrying too much about trying to juice up the algorithmic side.
ALSO: We probably need to curate some *small* datasets, specifically because they will run much faster!
Ah, that makes sense, yes we should totally nail down which enrichments have the most predictive power.
I just joined the discord channel, thanks for the invite! I'm looking forward to continue my tests and hopefully finding other counties with good data to use.
I was discussing (debating) Georgism with a friend recently who kept advocating for investment opportunities in land along side things like company stock. My takeaway (which I’d love to see @Lars Doucet and @Greg Miller flesh out) is that having individuals invest in and benefit from increasing land values is bad for society whereas increasing stock values are good for society. We need a wholesale mind-shift to a world that advocates for cheaper costs to use land. The land can still be valuable but the cost to access the land needs to go down over time. In this world, homelessness decreases with time (more market-rate cheap housing available), poverty goes down (poor people spend less time traveling to jobs), etc.
Some have said that Blackrock is just Privatized Georgism: just buy their stock on the stock market! Okay, that’s a bit besides the point. In general, I think there’s a divide in conversations i have on land value/wealth/housing policy space. There are those who either explicitly or in their rhetoric want to create more people with “wealth” via land ownership and appreciating land value. I think it’s hard (impossible?) to get everyone into that bucket, let alone without still large inequality. Therefore, I prefer the second option: publicize the benefit of land and return land value to the community that makes it value. To your point, this drives down land values (while increasing holding costs), and creates the incentives we want.
100% To me the challenge is a failure of imagination, i.e, Kahneman's narrative fallacy (https://fs.blog/narrative-fallacy/). I'd love to see stories about what the world would look like if the cost of housing went down over time so that normal citizens could imagine the tremendous benefits and reasonable costs. If they bought in to the outcome (i.e., "why" we'd want cheaper housing costs) then they'd ask "How?". Then we've got them.
This is awesome -- I can't wait to read it. I'm actually trying to listen to your articles while I run but it seems that you've disabled(?) Substack's feature to allow readers to play text as audio. Could you pls enable it?
Wow!
Super excited to see this post, specially as I've just spent the last week testing the library, and it's been awesome.
Also, I can attest that it can get quite computationally expensive, so the upcoming examples will be quite welcome.
Currently I've been running the test data (guilford-nc, downloades from Hugging Face) locally on my laptop. What are the hardware requirements (RAM, CPU) that would be recommended for running the test data?
We haven't really established that lately, but it's something we should nail down. One major thing I want to do is to figure out what enrichments are actually just kind of optional and and not all that important in most cases, so we can trim down the "default path" as much as possible.
Street frontage enrichments, for instance, are particularly expensive compared to the predictive power they yield. Also, we currently are generating horizontal cluster ids for three categories (general, land-only, improvement-only) and the second categories aren't currently being used for much, so we could do some trimming there too.
Also, when training your predictive models, I think there are some strong best practices--like you probably don't need to run full parameter search on all three gradient boosted trees (LightGBM, XGBoost, Catboost). Probably best to just start with just LightGBM and see how that performs, then spend time on your dataset and variable selection before worrying too much about trying to juice up the algorithmic side.
ALSO: We probably need to curate some *small* datasets, specifically because they will run much faster!
Ah, that makes sense, yes we should totally nail down which enrichments have the most predictive power.
I just joined the discord channel, thanks for the invite! I'm looking forward to continue my tests and hopefully finding other counties with good data to use.
I was discussing (debating) Georgism with a friend recently who kept advocating for investment opportunities in land along side things like company stock. My takeaway (which I’d love to see @Lars Doucet and @Greg Miller flesh out) is that having individuals invest in and benefit from increasing land values is bad for society whereas increasing stock values are good for society. We need a wholesale mind-shift to a world that advocates for cheaper costs to use land. The land can still be valuable but the cost to access the land needs to go down over time. In this world, homelessness decreases with time (more market-rate cheap housing available), poverty goes down (poor people spend less time traveling to jobs), etc.
Some have said that Blackrock is just Privatized Georgism: just buy their stock on the stock market! Okay, that’s a bit besides the point. In general, I think there’s a divide in conversations i have on land value/wealth/housing policy space. There are those who either explicitly or in their rhetoric want to create more people with “wealth” via land ownership and appreciating land value. I think it’s hard (impossible?) to get everyone into that bucket, let alone without still large inequality. Therefore, I prefer the second option: publicize the benefit of land and return land value to the community that makes it value. To your point, this drives down land values (while increasing holding costs), and creates the incentives we want.
100% To me the challenge is a failure of imagination, i.e, Kahneman's narrative fallacy (https://fs.blog/narrative-fallacy/). I'd love to see stories about what the world would look like if the cost of housing went down over time so that normal citizens could imagine the tremendous benefits and reasonable costs. If they bought in to the outcome (i.e., "why" we'd want cheaper housing costs) then they'd ask "How?". Then we've got them.
This reminds me of an article I wrote about a year ago when I was blogging more personally... it's not the best written: https://peoplesland.substack.com/p/stories-from-the-city-that-shifted
This is awesome -- I can't wait to read it. I'm actually trying to listen to your articles while I run but it seems that you've disabled(?) Substack's feature to allow readers to play text as audio. Could you pls enable it?
https://support.substack.com/hc/en-us/articles/7265753724692-How-do-I-listen-to-a-Substack-post