Astral Codex Ten

IRS

22 min read

IRS

Article

IRS is a recurring organization in the Astral Codex Ten archive, appearing 12 times across 12 issues between March 18, 2021 and September 19, 2025. The archive places it in contexts such as “Strong protections separating investigation of tax fraud (IRS?)”; “the IRS has a staff size of 74,454”; “the cost of the IRS”. It most often appears alongside US, Scott, Bill Gates.

Metadata

  • Category: Organizations
  • Mention count: 12
  • Issue count: 12
  • First seen: March 18, 2021
  • Last seen: September 19, 2025

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Book Review: The New Sultan
March 18, 2021 · Original source
No direct inline source block was recovered for this mention.
Does Georgism Work, Part 3: Can Unimproved Land Value be Accurately Assessed Separately From Buildings?
December 11, 2021 · Original source
Gwartney says that when he was the assessment commissioner and chief executive officer in British Columbia, he had a staff of 690, and that this number has not changed significantly since then. British Columbia has a population of about 5 million, so that's 1 assessment officer for every 7,250 British Columbians. For context, the IRS has a staff size of 74,454, or about one IRS agent for every 4,425 Americans. I don't have data on how many property tax assessors the USA has in total, but the above slide suggests British Columbia's figure is on the high end. As for how you actually do assessments, sure, you can send out an army of assessors to value each and every property in your jurisdiction by hand. However, not only is that labor-intensive, it's also a recipe for inconsistency. Whatever method you're using to value properties needs to be consistent and standardized across all properties, so you don't have sharp discontinuities on the assessment map that are due solely to differences between Assessor Fred and Assessor Sally's personal methodologies. Thankfully, we're living in the modern age, and we have some fancy new tools at our disposal. 4. Modern Technology Georgists were doing split-rate assessments to allegedly good success long before the rise of the computer, such as J. J. Pastoriza's effort in setting up a Georgist tax regime in Houston, Texas in 1911. Today, we have spreadsheets, property value databases, GIS mapping visualizations, regression analysis, machine learning...the works. According to Gwartney, the Canadian province of British Columbia has revalued all its land and all its property on an annual basis simply by using computers and market analysis, ever since he first helped them set up their system back in 1975. Not every jurisdiction revalues their land this thoroughly and this often, but Gwartney says there is no significant technical or staffing barrier standing in the way. Gwartney has been retired for some time, so his seminar didn't cover all the latest cutting-edge techniques that have come out in the last few years. Let's look at some recent papers and see what new tools assessors have to play with. The first on my list is Land Value Appraisal Using Statistical Methods by Kolbe, Schulz, Wersing, and Werwatz (2019). This is a study on mass appraisal techniques using real estate transaction data from Berlin, Germany. It claims that not only are the results cheaper and faster to generate than those done by conventional property assessment methods, but they are also no less accurate than those done "by hand" by experts. Kolbe et al. assert that, provided you have access to high quality market transaction data, you can perform accurate and efficient mass appraisals of land values. They chose Berlin because it "has a very effective system of property transaction data collection and storage," in contrast to other parts of Germany. They cite some prior work by Almy (2014) studying Canada, the Netherlands, and the United States, suggesting that the assessment cost per property can be brought down to 20 Euros–25 times cheaper than what some other people (Fuest, et al. (2018)) assert. Given an average tax receipt of 2,000 Euros per property, this means that the assessment cost should represent only about 1% of the funds raised. Is that good? Let's take this assertion at face value for the moment and compare it to the cost of the IRS. Federal tax receipts in 2020 were $3.42 trillion, and operation costs for the IRS were $12.3 billion, or 0.36%. However, the IRS outsources most of the labor of tax preparation to the taxpayers themselves, with compliance costs estimated between $200 billion and $400 billion a year, to the delight of Intuit. Add that up and the total cost of federal tax collection to the economy is anywhere between 6-12% of the amount it raises. And what about sales tax? According to a 2006 report by PriceWaterHouseCoopers: The study finds that the national average annual state and local retail sales tax compliance cost in 2003 was 3.09 percent of sales tax collected for all retailers, 13.47 percent for small retailers, 5.20 percent for medium retailers, and 2.17 percent for large retailers So a compliance cost of 1% would be way more efficient in terms of cost collection than the other two most common forms of taxation, and taxpayers don't even have to do anything themselves, other than pay the bill. Alrighty, how about the accuracy? The authors cite two international examples, Australia and Lithuania, as among the few countries in the world that have both a Land Value Tax and statistical methods for mass appraisals. Hefferan and Boyd (2010) assert that objections to assessments from property owners in Australia are less than 1%. I'm willing to buy the improved efficiency claims just by taking a look at some methodologies. It seems reasonable that computerized records and algorithms can cut costs significantly; the real question is if you're trading off accuracy. The other papers I found on the subject are Bencure, et al (2019) in BayBay City, Philippines, Kilić, et al (2019) in Croatia, Yalpir & Unel (2017) in Konya, Turkey, and Raslanas et al. (2014) in Vilnius, Lithuania. Let's dive in and examine some methods. 5. Mass Appraisal Methods Here are some of the latest mass appraisal methods cribbed from the research papers listed above. All of these are based on taking market transaction data, plotting them out on a map, and running computations over them to estimate valuations for the properties you don't have known values for. Furthermore, all of these methods are able to value land and building values separately. Multiple Regression Analysis This paper by Yalpir and Unel out of Turkey gives a straightforward example of using Multiple Regression Analysis for land valuation. For those of you who didn't study math, let me explain regression analysis. This is a family of mathematical models where you basically take a data set, ask the question "what mathematical formula would best fit this data," choose a basic equation model, and then have a computer search for a set of coefficients that "best fit" that curve to the data with the least amount of error. The simplest example is using linear regression on a scatterplot of observed data points to fit a trend line. This is a common exercise in freshman physics and statistics classes. You can use more complicated versions of this numerical method to take a big bag of observations (real estate sales) and use "multiple regression" to tease out dependent variables (land value and improvements value) based on the independent variables (size, location, age, number of bedrooms) of your observations. In this case the team identified about a hundred different factors that can affect the price of a property: Then you create an entry for each property, fill in the values for each of those characteristics, and run it through the regressor. Take note of how many of these factors start with the words "proximity to." Each of these can be calculated automatically just by knowing where the property is on a map, and each of them is an independent contributor to the value of the property's location. The next step is to generate individual "index maps" that combine various related features into combined heat maps. Then you run everything through and see if it works. You can get the land share of the final value by combining the contributions of all the individual factors that you associate with "land," such as proximity to important things. In the verification section the authors say: As a result of the analysis, since the significance level (0.000) p <.05, corresponding to the F values in the ANOVA test, indicates that the regression analysis is appropriate and the models are significant. The criteria that make up the model account for about 85% of the market value and 15% cannot be explained for reasons such as economic, non-existent data and unearned income. Unfortunately, they don't say anything about how accurate their model is for assessing land values specifically. Otherwise, this is a pretty good example of using the Multiple Regression method for estimating the individual contributions of various factors to overall property values. Gwartney says Multiple Regression Analysis was a standard method he typically used, of which this specific paper is just one example. Nonparametric kernel regression This will be a method familiar to the programmers in the audience who have any experience with image processing algorithms. Here's an example from this old Gamasutra article: The basic idea here is to take a matrix of numbers, called a "kernel", and run that over every pixel in a source image. The kernel tells you how strongly to weight all of the source pixel's neighbors to compute a final result for that position. A simple "box blur" is a kernel where every value is 1 (meaning it averages the values of all neighboring pixels within a range). The more subtle gaussian blur illustrated above uses a two-dimensional normal distribution of values so that each pixel is most affected by those nearest to it. So let's apply the same principle to land valuations. If you have a map with lots of transaction data of pure land sales–defined as sales of either vacant land or teardown properties (where the building value is essentially zero)–then you can use a special kernel filter to smoothly interpolate land values across the region. So you basically have a smooth curve that mostly favors close-by points, tapers off a bit, and then disregards anything outside a certain distance entirely. The big assumption here is that land values change smoothly and do not change suddenly across very short distances. There are, in fact, locations with sharp jumps in value (any town with an "other side of the tracks," for instance). But for cases where we know a priori that land values change smoothly, this method is appropriate. No other prior restriction is placed on the form of the land value map, however, and this is why it's called "nonparametric." Here's an illustration. The outer box is the entire search distance that the kernel considers, and the circles represent the falloff of the curve itself. The size of the box is called the "bandwidth" and is set by the user. Everything outside of it will have zero influence on the kernel's output at any given location. This method operates on the same basic logic that I used when I hand-estimated the land value of that San Francisco house in Part I based on the value of the empty lot next door. However, it makes the whole procedure systematic. It can easily and accurately estimate the land value of a property with a big fat building on it simply by smoothly interpolating the known values of the nearby parking lots. Of course, it has limitations. First and foremost, it's a highly local operation, so if you have properties you're trying to value that don't have nearby pure land sales data, you can't really do much with this. Also, most people assume that city centers have less market transactions for undeveloped land than the countryside, as did I until I read that paper by Albouy in Part I. But in any case, this is just one method in your toolbox and might not be sufficient by itself. Its key advantage is that it works directly from true market data for land and doesn't need or want any other subjective data. In the end, basic kernel estimation just fills in the land value of unmeasured locations with a local weighted average of known locations. Nonparametric adaptive regression Kolbe, et al. build on the kernel regression method with a technique called Adaptive Weights Smoothing (AWS), which runs in several iterations and adds additional weight to any observed data points that are sufficiently close to the point being estimated. I'm not 100% sure about what all the math means, but it seems like it's basically a "smarter" version of the basic kernel method. Left: Nonparametric kernel regression, Right: Adaptive Weights Smoothing. I think the authors goofed and printed the same figure twice with different headings because they're identical if you overlay them in Photoshop. Semiparametric regression Now, the above two methods assume you have plenty of "pure" land sale records to work with. But if you're trying to work out prices in the city center, you've probably mostly got land and buildings mixed together. To do this effectively, we need more data, and this is where the "parameter" in "semiparametric" comes in. The model described in Kolbe et al. seems like a flavor of multiple regression analysis that takes the price, the location, and various characteristics of the building and feeds it into a regressor. But we've got "semi" parametric here. What does that mean? Well, if you already know how certain relationships between the data work a priori, it's better to enforce those relationships yourself rather than leave it to the computer. Here, we enforce the assumption that if two properties are right next to each other, then the value due to location is going to be essentially identical. This algorithm starts by ordering things geographically and then working out the differences in observed price by regressing on the difference between remaining property characteristics. In this method, the power of "location, location, location" is not something we're leaving to the regressor to discover by itself. Results of the Semiparametric regression method, we can see some significant differences from the simple kernel-based model. As you can see above, this gives you more detailed and likely more accurate results, and you're better able to assess the values of properties with buildings on them, even in the absence of pure land sales. This technique is more complicated and bakes in assumptions about the power of location, but otherwise doesn't assign subjective human weights to the various property characteristics. The chief human bias comes in the form of deciding which property characteristics are measured and made legible to the model in the first place. Okay great, but how accurate are the above three methods? Their main point of comparison is this thing called the "Bodenrichtwerte," or BRW. I think that means "ground-level-values" in English, and it's an expert-assessed map of land values for Berlin done the traditional way. The nonparametric kernel regression method has a correlation of 0.704 with the traditional method and has the added disadvantage that it's not able to produce estimates for the city center, only the outlying areas. Furthermore, the BRW map does show sharp discontinuities, which is another knock against the kernel method, at least for the city center. What about the iterative method? Kolbe et al. find that "the agreement between [Adaptive Weights Smoothing] land value estimates and, both, land prices and BRW land values is fairly good for all values of λ." Doing some quick checks, their values seem to be within about 85% of the BRW values. A different Kolbe et al. paper called Identifying Berlin's land value map using adaptive weights smoothing goes into more detail and claims to give "similar" values to that of the BRW. For the semiparametric method, they "found a strong positive correlation of 0.845" between their numbers and a previously expert-assessed set done using the traditional method. That sounds pretty good. It seems their margin for error is about plus or minus 15% compared to the traditional expert method. I'd like to see more direct comparisons against market transactions themselves, though, because if the prior expert assessments are wrong, then the main achievement here is improved efficiency, not accuracy. However, this method doesn't seem to be dramatically less accurate than the old way of doing things. The last three models came from the Berlin case study, where you have excellent market transaction data in an extremely wealthy and high-trust society. But what if you're trying to assess land in a developing nation with poor market transaction records, weak institutions, and widespread poverty? Innovative Land Valuation Model (iLVM) This is the particular name of the method described in Development of an Innovative Land Valuation Model (iLVM) for Mass Appraisal Application in Sub-Urban Areas Using AHP: An Integration of Theoretical and Practical Approaches by Bencure, Tripathi, Miyazaki, Ninsawat, and Kim. They used BayBay City, Philippines as their case study. Whereas the previous models are very "hands-off" and let the computer work out the relationships between prices and property characteristics, here you get expert human opinion directly involved in building the model, baking in weights that directly embody judgments like "properties next to major roads are more valuable." These judgments are based on expert opinions that presumably come from observed experience but are a priori judgments nonetheless. Here, look at this big complicated flowchart. The "Analytic Hierarchy Process" in the box on the left is a particular kind of method for getting experts to set weights. The authors give this reason for using it: Despite criticism pinpointed by other scholars, the AHP remains the commonly used in many research fields and practical applications. This is because the AHP: (1) overcomes human difficulty in making simultaneous judgment among factors to be considered in the model; (2) is relatively simple as compared to other MCDA [multi-criteria decision analysis] methods; (3) is flexible to be integrated in various techniques such as programming, fuzzy logic, etc.; and (4) has the ability to check consistency in judgment After identifying a list of "factors" that can affect land value, they group them into taxonomical buckets: Note that certain factors like "Coastline" appear in multiple buckets; this captures the various influences a characteristic can have. For instance, land on the coast tends to be more economically valuable because of tourism, shipping, fishing, etc., so that goes under "economic." But land that's next to the coast is also more likely to flood, so it also goes under "environmental." And then there are various land use restrictions that apply specifically to coastal areas, so it goes under "legal" as well. In this way, a single factor like "the property is on the coastline" can have both positive and negative effects on land value (e.g., it's more economically valuable but it also might flood, and there are certain things you aren't allowed to do there). The next step is to set down some rules for how sensitive each factor is to location and distance. So here we can see that the economic benefit of being on the coast is most strongly felt if you're within half a kilometer of the ocean, but the environmental effect (e.g., risk of flooding) is most strongly felt when you're within 0.03 kilometers. And so on and so forth. Your experts help you work out all these rules. Note that for a few of these factors (such as land use and slope), you use metrics other than distance (e.g. land use classification and grade). Then you take all that stuff and assign everything a value between 0 and 5. Your team of experts then uses this table to come up with a set of weights for everything. What essentially comes out of this is a big linear equation with a bunch of coefficients for every one of your factors, which is then broadly fit to the observed market prices. When you're done, you can take any property on your list, multiply each of its characteristics by its respective weight, run that through your equation, and calculate the predicted price of the land. So how accurate is it? The authors compare it to standard Multiple Regression Analysis and claim it fares better. The Root Mean Square Error is quite a bit less than MRA. In addition, I think it's also saying that the MRA algorithm decided that only four of the factors were significant and basically ignored all the rest. By contrast, iLVM was able to maintain contributions from all the factors, because it doesn't leave that decision to the computer. I'm not 100% sure; it's not clear from the paper. The authors claim that about 67% of the variability is explained by their model, but they note that there are some areas where the model can be off by more than a factor of 1.0 in either the positive or negative direction. One thing that's kind of fun about this model is that you can make neat graphs like this that show the individual contribution of each factor: The main downside to this model is that it relies on a whole lot of subjective expert opinion and can be questioned on that basis. That said, it can be cheaply deployed in a transparent and consistent way across a large area. You can see why that's attractive for a developing nation with weak institutions and poor market transaction records; the argument is that this is a significant improvement over the former status quo. I wonder how well this model performs when you feed it better market transaction data, and how that would compare against all the others methods under identical conditions. More research is needed. Rather than drag you through a bunch more research papers, I'll just leave these others I found cited in the above studies: Killić et al. (2019) - Fuzzy expert system for land valuation in land consolidation processes
Inline links: J. J. Pastoriza's effort in setting up a Georgist tax regime in Houston, Texas in 1911, Land Value Appraisal Using Statistical Methods, Almy (2014), Fuest, et al. (2018), $12.3 billion, $200 billion, $400 billion, delight of Intuit, 2006 report by PriceWaterHouseCoopers, Hefferan and Boyd (2010), Bencure, et al (2019), Kilić, et al (2019), Yalpir & Unel (2017), Raslanas et al. (2014), https://substackcdn.com/image/fetch/$s_!stkG!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fef4c64b9-38bc-43c8-aa83-05eae3576e03_923x600.png, https://substackcdn.com/image/fetch/$s_!_9z0!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fdd98e114-7eb2-4566-a979-1f2f2dd27c22_701x867.png, https://substackcdn.com/image/fetch/$s_!8HN7!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1232522-51e0-4438-998a-b0be4615df6b_534x806.png, https://substackcdn.com/image/fetch/$s_!jFqw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fd5836457-7642-4235-9410-00906f043428_662x357.png, Gamasutra article, https://substackcdn.com/image/fetch/$s_!foLQ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Ff74d378c-39b7-49f9-9655-8cbbf7c89ff5_592x270.png, https://substackcdn.com/image/fetch/$s_!AjnN!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F458458cc-614e-4ab3-a57b-3f28b70db6c3_458x317.png, https://substackcdn.com/image/fetch/$s_!lVcf!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F77ea4f72-0e51-4c63-b14b-15c603ac2500_901x418.png, https://substackcdn.com/image/fetch/$s_!zoSx!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fed78916c-12c5-42ca-b581-7b59aa25bbd5_757x718.png, https://substackcdn.com/image/fetch/$s_!7Wm9!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F8fe15dbb-181f-46ae-bdde-840bdd6a2064_752x735.png, https://substackcdn.com/image/fetch/$s_!zig5!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F958762f1-a425-4017-86cf-058cb3eb4d59_713x389.png, https://substackcdn.com/image/fetch/$s_!8ZGT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Ff37696ef-7a6a-48ae-8169-734b875b0b57_800x319.png, Identifying Berlin's land value map using adaptive weights smoothing, https://substackcdn.com/image/fetch/$s_!3CR3!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F0489e086-69ae-4840-b658-59fee6b3af44_2000x1672.png, https://substackcdn.com/image/fetch/$s_!fA0K!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F5d39769d-46e2-4891-92aa-cb3766068204_2000x978.png, https://substackcdn.com/image/fetch/$s_!phFK!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F5d1a519c-93d7-4bed-9577-7478fb239bca_1968x3548.png, https://substackcdn.com/image/fetch/$s_!XtLN!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Ff59cb148-e0da-456b-b205-973e04239be7_587x647.png, https://substackcdn.com/image/fetch/$s_!My3b!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fee7b484b-3be8-4363-bcb8-1cb4fb4a7c01_661x655.png, https://substackcdn.com/image/fetch/$s_!SqQA!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F8a84b431-1250-427e-a67a-b3e2b8a3c0dd_896x623.png, https://substackcdn.com/image/fetch/$s_!qPOz!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fb71526d0-736c-45d7-ad14-36e5670f78ab_1153x881.png
Motivated Reasoning As Mis-applied Reinforcement Learning
February 01, 2022 · Original source
No direct inline source block was recovered for this mention.
The Passage Of Polymarket
February 07, 2022 · Original source
No direct inline source block was recovered for this mention.
Highlights From The Comments On Motivated Reasoning And Reinforcement Learning
February 11, 2022 · Original source
No direct inline source block was recovered for this mention.
Links For November 2023
December 01, 2023 · Original source
No direct inline source block was recovered for this mention.
Highlights From The Comments On Hanson And Health Care
May 10, 2024 · Original source
No direct inline source block was recovered for this mention.
Contra Stone On EA
May 30, 2024 · Original source
No direct inline source block was recovered for this mention.
Your Book Review: Real Raw News
July 26, 2024 · Original source
No direct inline source block was recovered for this mention.
Your Book Review: The Pale King
September 06, 2024 · Original source
No direct inline source block was recovered for this mention.
Your Review: Alpha School
June 27, 2025 · Original source
No direct inline source block was recovered for this mention.
Your Review: Project Xanadu - The Internet That Might Have Been
September 19, 2025 · Original source
No direct inline source block was recovered for this mention.