Business strategy and firm location decisions: testing Traditional and modern methods.
For nearly a century, economists have relied upon the neoclassical principle of a "profit-maximizing firm." Two modern challenges to this principle have arisen: the theory of the value-maximizing firm, and machine learning. In this article, we empirically compare the predictive power of both traditional and modern approaches to business decisions. To do so, we make use of an unusual natural experiment, and extensive data, as follows: (1) Outline competing models of business decision making from both traditional and modern approaches: Expert judgement; an income model of a profit-maximizing firm; a suite of machine learning models; and a recursive model of a value-maximizing firm. (2) Assemble data on costs, productivity, workforce, transit, and other factors for over 50 large North American cities. (3) Empirically compare these models to determine which best explains the selection of 20 cities by Amazon Inc. for its "HQ2." We observe first that expert judgement, of the type traditionally performed by business economists, outperformed all other approaches. Second, we observe that "supervised learning" machine learning models performed poorly, with results that were often worse than a coin flip. Third, we found that the model of a value-maximizing firm slightly outperformed an income model using the same underlying data, and handily outperformed machine learning. Based on these results, we conclude that expert human judgement remains superior over. machine learning methods, and warns against naive reliance on such models when the penalty for an incorrect decision is high. We also recommend that businesses economists consider value methods for business strategy decisions.
Keywords Machine learning * Recursive * Neoclassical * Managerial decisions * Artificial intelligence
JEL Classification B21 * C61 * D21 * J23 * K20 * L21
1.1 Traditional and modern theories of business decisions
For nearly a century, economists, managers, and lawmakers have relied upon the neoclassical principle of a "profit-maximizing firm." This principle underlies decades of business school teaching, numerous aspects of finance, and a panoply of models for inventory management, yield management, and pricing policies. (1)
Indeed, the principle of a firm choosing a location to maximize profits dates at least as far back as the landmark works of David Ricardo and Johann Heinrich von Thiinen, predating the emergence of the neoclassical school of economics by a half century. (2) Even the "new economic geography" that emerged in the last years of the 20th century relies upon an updated version of the profit-maximizing firm. (3)
However ubiquitous the principle, the practice appears quite different:
* Corporate managers commonly state that they focus on "shareholder value," and explain costly investment decisions, risky acquisitions, and speculative research expenditures on that basis. Excessive focus on maximizing profit has even been characterized negatively as "short-termism."
* American legal doctrine, which at one time was firmly supportive of the profit-maximizing goal of the corporation, has shifted toward both "shareholder value" and the concept of businesses pursuing larger societal goals. (4)
* Machine learning methods, often supported by "big data" collection and storage, increasingly drive advertising decisions, planned maintenance, yield management, and real-time pricing policies. The focus of these methods is often maximizing metrics other than profits, such as page views, purchases, number of users, and "clicks."
* The standard curriculum in graduate programs for business and economics now recognizes the value of managerial flexibility, the importance of "real options," and the use of hedging and other strategies to minimize risk. (5)
* The "world is flat" meme questions whether location decisions (and, implicitly, the costs associated with doing business in any one place) really matter in the 21st century. While not quite a theory, this idea has generated wide currency in the popular press and heated intellectual debate. (6)
* A number of widely recognized corporations have risen to extremely large market capitalizations without profits; some very large and highly profitable firms have made an explicit policy of not distributing dividends to their stockholders. (7)
None of these phenomena are consistent with the principle of maximizing profits. (8) What best explains important business decisions, if not profit maximization? We use a natural experiment, and a very large dataset covering multiple categories of information, to examine this question. We focus on the following important business decision: where do medium- and large-sized firms choose to locate?
1.2 Four approaches
We examine four approaches, two traditional and two modern, to model such an important business decision:
1. A traditional income model of the firm.
This approach is based on the profit-maximizing principle from neoclassical theory, and relies upon financial statement analysis and cost-minimization methods that are commonly taught in business and management courses.
Both professional judgement and analysis of financial statement data are used in this model, which requires only rudimentary calculations.
2. An index of key variables based on professional judgment.
This approach is representative of the notion that wisdom, experience, and judgment are superior to purely financial models. It uses the skills that business economists have traditionally employed.
We have examples from three different groups of economists for this approach.
3. A suite of machine learning models.
This approach represents the dramatic growth of machine learning and "big data" methods in the business world. It also represents a fundamental challenge to the notion of structured business decisions, suggesting that assembling a large amount of data, and "letting the data decide," is often superior to traditional managerial or economic methods.
Among these approaches, this makes the widest use of the large amount of data, albeit with a computational burden that is exponentially larger than with the traditional approaches. We used dozens of variations of such models for the comparison in this article.
4. A recursive decision model of a value-maximizing firm.
This novel approach, which has become practical only recently, makes use of advances in control theory and computational methods. It replaces the neoclassical principle of profit maximization with a new one of value maximizing. This approach relies upon the same financial statement analysis as a traditional income model, but takes into account management options such as the ability to change course in the future.
To solve such a model requires the use of functional equations and recursive methods. The amount of professional judgment on variables is somewhat higher than with the traditional approaches, as is the computational burden.
To establish a benchmark, we also developed "coin flip" and "educated guess" models. These represent the predictive results that could be expected without any expensive investment in quantitative methods.
Each of these is described in Sect. 1.2. Additional technical information is included in the "Methodology Appendix."
1.3 Large and varied data for use across models
The data needed to use four different approaches are both large and varied. We include data on all of the following:
1. Financial statement data on Amazon, Inc., as reported in their periodic financial disclosures. We supplement this company-specific data with industry and economy-wide data relevant to large retail and technology firms. These data informed our income model of the firm.
2. The population, location, and airport service characteristics of large cities and populous metropolitan areas in North America. These data were used to identify large cities that were candidates for a major investment for a large-sized firm.
3. A wide variety of quantitative variables, covering over 50 indicators across 56 cities. These data extend far beyond those traditionally used in business economics, and include the following:
* Traditional cost variables, such that of real estate, and wage and benefit costs for specific categories of workers.
* Size of the workforce, the number of graduates of higher education institutions in the area, and immigrants to the region with advanced degrees.
* Amenity indicators, such as "walkability," availability of parks, and number of sunny days each year.
* Sentiment variables, such as the share of votes for both major political parties in two different election years.
* Business tax burdens.
* Availability of an international airport with regular flights to certain cities.
* Transit and transportation use indicators, including hours lost to congestion, and number of trips on mass transit.
* Indices of economic freedom, representing policies related to worker and employer liberty.
* Indicators of disparity in income and racial inclusion.
* The share of households speaking English only.
* The number of major league sports teams.
It is important to note that these variables were not selected for a specific structural model. Furthermore, the indicators are of varying quality, overlap conceptually, and in some cases are produced by entities with explicit policy agendas. The use of a large number of variables covering multiple topics is characteristic of the machine learning approach, and allows for the data to partially determine both the parameters and the structure of the model.
4. Information on the competition among cities for the Amazon, Inc. "HQ2" that began in 2017. This includes the original RFP document outlining the intended HQ2 facility and its operation; the factors they considered important for their selection; the identity of many of the over 200 cities that submitted responses; and the 20 that were selected in January 2018.
5. A set of three predictions made in the Fall of 2017 by business economists, all of which were published after the Amazon HQ2 announcement in September 2017 but before their selection of candidate cities in January 2018. These included:
* The "HQ2 Index" assembled by a group of business economists at Anderson Economic Group, which included costs, workforce, higher education, tax burden, and transit variables for 35 metropolitan areas within the US. This dataset was later augmented by disaggregating the cities that were within the large metropolitan areas surrounding New York and Washington DC, and by adding cities (including Toronto, Canada, and Columbus, Ohio) that were originally excluded.
* A prediction of 20 MS As made by the Brookings Institution, based on a published rationale.
* A prediction of 10 MSAs made by Moody's Analytics, which included a list of factors and an economic rationale for their use.
This large and varied set of data allowed us to calculate predictions, and compare the accuracy of predictions, for a wide range of models that fit into four different approaches.
These data are further described in the "Data Appendix." Exploratory data analyses of variables used in the HQ2 Index are shown in Fig. 1 and Fig. 2.
2 Four approaches
We summarize below the four approaches we use in this comparison. Two of these arise from traditional business economics. Two have been recently introduced into the business world, and involve techniques that have not been part of traditional business economics methods.
2.1 First traditional approach: an index assembled by experts
Multiple economic consulting firms, experts, and pundits announced predictions for the Amazon, Inc. HQ2 selections. At least a handful of these arose from economic consultants that described both their general method and released a list of predicted selections well in advance of the announcement of candidate cities in January 2018. These included Moody's Analytics Inc., (9) the Brookings Metropolitan Policy Program, (10) and Anderson Economic Group LLC. (11) They were joined by a chorus of experts, journalists, editorialists, and interest groups that offered their own commentary.
All three of these experts described their selection process, with both Brookings and Anderson constructing an explicit index. All the rankings included explicit talent pool indicators. The Anderson and Moody's rankings included cost of doing business indicators. The Anderson HQ2 Index also included mass transit, congestion, and business tax burden variables. Moody's included an explicit "quality of life" indicator that the other two did not attempt to quantify.
The Anderson HQ2 Index included explicit indicators of variables and a ranking of 35 candidate MSAs, all in the United States. Brookings released an article that included 20 MSAs, including two in Canada. Brookings also included Seattle, Amazon's current headquarters. Moody's released results for 10 MSAs in the United States; one of these ("New York-Newark-White Plains") was arguably broad enough to capture at least four cities that were in contention. All three of these experts used metropolitan areas as the unit of analysis, which was consistent with Amazon's RFP
2.1.1 Similar results from expert predictions
A comparison of the reports from these three groups of experts reveals very similar thinking, and similar predictions:
All three included New York, Boston, Chicago, Atlanta, and Philadelphia. Brookings chose Detroit and San Francisco, while Anderson Economic Group chose Raleigh and Los Angeles as well. Moody's had similar, but not identical, selections, including Austin, Texas. (12)
All three of these expert predictions produced very good predictive results, as shown in Table 1. Furthermore, regardless of the precise "batting average" for each, they all seemed to have a good method for predicting the selections.
2.1.2 Note on the Comparison of Expert Predictions
A fair comparison required adjustment for two factors:
* First, Amazon selected one Canadian city (Toronto) that was not initially considered as a candidate by two of the three experts; it also selected two U.S. cities (Columbus, Ohio, and Pittsburgh, Pennsylvania) that, in a technical sense, did not meet the criteria stated in the RFP. (13)
* Second, Amazon selected multiple cities within two of the MSAs considered by all three experts, namely, Washington DC and New York.
In the comparison shown in Table 1, we disaggregate the cities within the selected MSAs and count them as correctly predicted. One expert later augmented the list of candidate cities to add Toronto, and found their index placed that city in the top twenty; we include that. One expert published a prediction for only ten MSAs; we use these ten in the table. As a result, a precisely comparable "batting average" cannot be calculated for all three.
2.2 Second traditional approach: a model of business income
Our second traditional approach is a model of income for a hypothetical Amazon HQ2 facility. For this approach, we examined Amazon's financial disclosures to understand the structure of their business. From this, and the indications from Amazon's RFP, we assembled a model income statement for the proposed HQ2.
2.2.1 Data on Amazon's business
Amazon's total revenue in 2016 was $135.99B, of which $79.8B was from North America and "AWS" (Amazon Web Services). They report 341,000 part-time and full-time employees. Their first risk factor is "intense competition," and the second is that "our expansion places a significant strain on management ..." (14)
Operating income in 2016 was positive for NA and AWS, and negative for International, for the 3 years 2015, 2016, and 2017. Total operating income was positive in 2015 and 2016; for calendar year 2016 it was $2.4B net income (after tax) out of $136.0B net sales.
2.2.2 Project scale and structure of proposed HQ2
We assumed the scale of the facility would be similar to its headquarters at Seattle. We also assumed:
* The planned second headquarters would be built in three stages (each lasting 5 years) over 15 years.
* The first-stage facility would have roughly one-third of 50,000 workers.
* Productivity of these workers would be similar to current Amazon employees in the headquarters; meaning that they would deliver the same level of revenues as current employees at the Seattle location.
* Fulfillment (warehouse and delivery) operations would take place largely from other facilities in North America and Overseas.
Based on these assumptions, and using historical Amazon financial information summarized above, we created a model income statement for the proposed HQ2 operation.
2.2.3 Data on costs and productivity in different cities
We used the large data assembled on the candidate cities, and the model income statement, to estimate profitability for the proposed facility in each city. These calculations included the projected profits at three stages of operation, beginning in the first phase where, as noted above, we assumed that about one-third of the planned workers would be hired. We included costs of expansion for each phase as if they were amortized over a relatively short time period.
The variables used for these calculations were those that the financial statement analysis, and traditional business economics professional judgement, would indicate were important: costs of labor for specific occupations, costs to rent office space, and business tax burdens. We did not consider lifestyle, sentiment, or other variables for which there was no immediate impact on a company's income statement.
2.2.4 Illustrating the income model calculations
An example of the income statements calculated for cities is in Table 2, "Income Statement Calculations for Selected Cities." Note in these results how the income model penalizes high-cost cities (such as New York) over lower-cost cities (such as Newark) that have similar access to skilled workers. In the Washington DC market, Northern Virginia also did well with access to a large workforce and relatively reasonable costs.
The intermediate results from these data and income model calculations are illustrated in Fig. 3, "Comparison of Cost and Productivity Metrics," and Fig. 4, "Distributable Profits for Standardized Facility, by City."
We provide additional information on this analysis in "Methodology Appendix."
2.3 Modern approach: value-maximizing firm
We created a model of value-maximizing firm following the method demonstrated in the (Anderson 2014) article in Business Economics. This method first required creating a model income statement for firms in specific industries, using industry and firm-specific information. These data and calculations were available from the income model described above.
The second step is the critical difference between an income and value model. In this step, the value of the firm is calculated as the result of an optimization problem. The quantity to be optimized is the value of the firm, in its current state. The "state" here is represented by the cost, productivity, and talent variables for each city. Given that state, the optimization occurs across a set of feasible actions or decisions of the firm, which in this case is to operate, expand, or shut down.
The value in each state, given that set of feasible actions, is the sum of two parts:
1. The current-period earnings, which are based on current market conditions and current management policies; and
2. The discounted expected value in the following period. This "continuation value" is defined as the sum of the earnings in the subsequent period and the expected discounted value in the following period.
Thus, the model is recursive, meaning the essence is replicated numerous times within itself. Recursive methods in theoretical economics have a distinguished history, having been pioneered by the Nobel laureates Robert Lucas Jr. and Thomas Sargent. (15) While commonly taught today in graduate courses, their use in practical business economics is still novel.
2.3.1 Formulating and solving the value functional equation
The optimization decision of the manager of a value-maximizing firm is captured in the value functional equation (sometimes called the "Bellman equation" after mathematician Richard Bellman) shown below:
[mathematical expression not reproducible]; (1)
where s is the state, x is the action, [GAMMA] is the feasible set of actions [beta] is a discount factor, t is a time index, and V([s.sub.t]) is the value in that state at that time when the manager follows the optimal policy.
Note that the maximization operator applies to the sum of current rewards and expected future discounted value. Thus, the manager explicitly considers the trade-off between current earnings and future value. The proponents of this approach argue that this mathematical formulation is much closer to the thinking of actual human managers than the static profit maximization equations of neoclassical economics.
This is a functional equation, meaning a function of functions. Not all functional equations are solvable. However, this one meets the conditions stated in (Anderson 2012) for a solvable value functional equation for a business, and we use an iterative method to solve it for the projected facility. Using the data outlined below, we solve it for the market conditions in each of the candidate cities, to arrive at a value-in-place estimate for a projected Amazon HQ2 for each. More information on this method can be found in the "Methodology Appendix."
2.3.2 Data for the value model
We use the same data in this model as for the income model, and use the current earnings calculated in the income model as an intermediate calculation for the value model. However, we take the additional steps of specifying transition probabilities for the individual cities, as well as income ratios that are slightly different for additional stages of operation.
We provide further information on this model in "Value Model," in the "Methodology Appendix." The software for this, and the income model, is also described there.
2.4 Modern approach: machine learning
Lastly, we consider the challenge posed by the explosion in the field amorphously called "artificial intelligence," and in particular on the better-defined field of machine learning. The reliance upon machine learning methods, and the collection of data for the purpose of using these methods, has been an undeniable change in business over the past decade. With this in mind, we make a tentative and well-qualified test of machine learning against the other methods.
Following the emerging practice of machine learning, we make use of a large amount of varied data. Unlike with traditional income models (and the value models that rely upon them), the standard practice in machine learning is not to select those variables that are known to directly affect costs or revenues.
Instead, machine learning methods allow the data itself to decide which variables are important. (16)
It is a conceit to attribute machine learning methods only to developments in the last two decades, or even the last half century. Many such techniques rely upon Bayes Rule, which dates from 1763. Linear discriminant analysis was pioneered by R.A. Fisher in 1936. Linear regression was used by Francis Galton in the late 1800s, although proper credit should be given to the least-squares methods of Carl Friedrich Gauss and Adrien-Marie Legendre in the prior century. (17) All these statistical methods were employed before the term "machine learning" gained currency, and now form the basis for several such techniques.
However, it was the advent of large datasets and fast computers that opened the door to machine learning on an industrial scale, and developments over the past two decades have been swift. One key methodological advance was the development by (Cover and Hart 1967) of "nearest neighbor pattern classification," which today forms the basis of one of the most commonly used methods, known as k nearest neighbor classification.
2.4.1 "Supervised learning" models
It is important to note that these are "supervised learning" models, meaning that the models were "trained" with the "supervision" of the outcome data. In this case, the cities selected by Amazon were known when these models were run, and the selections were an explicit part of the machine learning models. This means all the machine learning models had an advantage that the expert predictions lacked.
Of course, the author knew these data when assembling the income and value models, and for this reason we cannot completely ignore the fact that some human interaction occurred, and therefore these models (and nearly all others in the economics literature) are not completely ex ante.
Multiple machine learning models are described under "Machine Learning Models" in the Appendix. As noted there, we ran several dozen variations of different models, some with different subsets of variables.
2.5 Testing the methods, using a natural experiment
With these challenges in mind, we consider a fundamental question of economic geography: what drives the location decision for businesses? Is it profit maximization? Is it the newer theory of value maximizing? Is it an amorphous, constantly changing intelligence best revealed through machine learning? Or has the world "flattened" so much the entire question is moot?
To answer these questions, we make use of an unusual natural experiment, and proceed as follows:
1. We observe and document the Amazon HQ2 competition, including the RFP, for over 200 competing cities, the 50 or so cities that appear to be serious contenders, and the resulting selections. We use this as a natural experiment in business decisions.
2. To assess whether the natural experiment is indicative of business decisions by other firms, we compare the Amazon HQ2 selections with the metropolitan areas that have experienced new establishments over the past few years, and discuss the focus of this test.
3. We prepare competing economic models representing each of the four approaches to business decisions we identify above. For some of these approaches, we have multiple such models. We use well-documented mathematical software to program and run these models, as described in the "Methodology Appendix."
4. We use large sets of data from the multiple sources to populate each of the competing models. These data are described in the Appendix.
5. We estimate these models using the exact same data on the various cities and MSAs. For the expert predictions, the data were largely published before the Amazon selections. These same data can be used for the income and value models. The expert selections are therefore ex ante predictions, while the others are ex post. We train the machine learning models with these same data and the Amazon selections, thus undertaking what is known as "supervised learning."
6. We evaluate each of these against the actual selections by Amazon, using a consistent metric.
7. Based on these results, we provide conclusions important to business economists that use, or may consider using, these methods in their professional work. Many of these conclusions will also be important to business managers, as they contain serious insights into the risks of relying upon certain techniques.
2.6 The strength and limitation of the Amazon HQ2 experiment
The natural experiment of the Amazon HQ2 competition has clear strengths as a test of methods: it was widely publicized, over 200 cities entered the competition, the "rules" were published and the results released publicly, and a handful of contemporaneous predictions of experts are available. However, there are a few weaknesses:
* Amazon, regardless of its current status as an exemplar of the technology sector and a corporation of extraordinary capitalization and influence, is just one company. It has its idiosyncracies.
* We cannot rule out that Amazon actually relied upon one or more of the expert predictions we use.
* A number of cities and states publicly dangled incentives in front of Amazon.
We consider these worthy of noting, but not sufficient to undermine the findings we present below, for the following reasons:
* Using the same validation accuracy metric we use for the comparison methods, the Amazon selections appear to predict the cities with the largest growth in establishments among the candidate cities for which we had Census Business Dynamics data, and for both 100+and 500+ employee data series.
* If Amazon did not read the results of the experts that are cited here, they would have been avoiding free and useful advice. However, it appears they were inundated with such advice.
* The selections by Amazon of the 20 candidate cities had no apparent relationship to the publicly offered incentives.
3 Comparing model results
3.1 Comparison metric
Among the four approaches we considered, different models produced results in different forms. This was especially the case for the expert predictions, which arrived in index, top-ten and top-twenty fashion. The income and value models produced estimates that could be used to rank order the cities. The machine learning models all produced classification estimates.
This required some effort to select a comparison metric that summarized the relative accuracy of these models. As our main tool, we use a validation accuracy metric. This is the share of the data points that are correctly classified as either positive or negative. This validation accuracy is a common metric in machine learning classification models. However, it necessarily leaves out some useful information, such as the number of false positives and false negatives. It also ignores the ordinal ranking that appears in the expert judgement, income, and value models.
3.1.1 Limitations of this metric
There are limitations to this metric, as applied to this sample, that should be noted. These limitations are taken into account by the author's subjective ranking of the reliability of these methods, but are not captured in the calculations shown in the table below.
* The machine learning models were not constrained to classify exactly 20 cities as "selected." Some selected more, some less; some machine learning routines selected zero. On the other hand, we used exactly the top 20 for the expert index, income, and value models.
* The expert judgement predictions shown in Table 1 were done ex ante, without knowing the number of cities Amazon would select, or the cities they did select. The "supervised learning" models all made explicit use of the selection data that were not known until January 2018. (The HQ2 Index used in the numerical comparisons had the same structure as the one originally published in October 2017, but was augmented in the number of cities and contains some revised data.)
* There is no benefit for consistency or logical coherence in the metric. Thus, making reckless predictions carries the same penalty as making predictions with small errors. We include a coin flip and educated guess method to partially account for this.
As a result of these limitations, the calculated validation accuracies should be considered optimistic for the machine learning and "coin flip" methods, and conservative for the others.
We summarize in Table 3, "Summary ranking of approaches by accuracy and reliability," our ranking of approaches. In this ranking, we use both the validation accuracy metric (expressed as a range when we used multiple models) and our own observations about the strengths and weaknesses of each approach. We discuss each approach below.
3.2 Experts using professional judgment
In first place among competing approaches is expert judgement in the form of an index of selected variables. The validation accuracy of the expert index for which we had complete data from Anderson Economic Group, with 16 of the 20 cities selected by Amazon among its top 20 rank, was impressive. In addition, none of the cities in the bottom of the HQ2 Index were selected by Amazon.
Of course, it is possible that this group of economists just got lucky. However, comparing these results with those of two other expert predictions--from Brookings and Moody's--is informative. These two also achieve impressive results. They both bettered the coin flips and educated guesses, as well as many machine learning models; and they both included many of the same cities within their respective lists. Finally, the reasoning offered by Moody's and Brookings was similar to that expressed by Anderson when assembling the HQ2 Index.
Thus, we conclude the approach works well, and that luck alone cannot explain its superior results in this test.
Furthermore, all three of these experts published predictions before the selection of Amazon was made, and therefore before the data for the machine learning methods could even be assembled. It is impressive that the ex ante predictions of at least three different sets of experts were better than the ex post results from many of the machine learning methods.
See Table 1, "Predictions of experts."
3.3 Value-in-place model
The value-in-place method performed well. It favored large-workforce, transit-intense places such as those in the New York City metropolitan area to be a place conducive to business value. However, as suggested by the "value" terminology, the best value cities were those that had moderate costs but excellent workforce and transit metrics. Thus, Manhattan and Brooklyn within the metropolitan area did poorly; Newark did much better. Like the income model (and any model), it took its objective literally.
This, of course, matches closely what Amazon said it wanted. It also matches what other companies say they want, including those in other industries.
See Fig. 5, "Value in Place Results for All Cities."
3.4 Income-in-place results
The income-in-place model predicted Amazon's location decisions with less accuracy than the value-in-place model, but still explained much of the business reasoning behind the selections. High-cost locations such as New York City were ranked poorly on the income-in-place model, compared to lower-cost locations.
Selected income results are shown in Table 2, "Income Statement Calculations for Selected Cities." Income for all cities is illustrated in Fig. 4, "Distributable Profits for Standardized Facility, by City."
Using this model with a high-tech business such as Amazon did not change the results we would expect from a traditional service provider or light manufacturer: low-cost locations win, because they maximize profits compared to other location. On the whole, traditional income models fell behind both pure expert judgement, and the more sophisticated value maximization models, that took into account both income and opportunities. See Table 3, "Summary ranking of approaches."
3.5 Machine learning models, with expert tuning
We ran a large number of machine learning models, including techniques that date back decades such as k nearest neighbor classification, and classification based on regression trees. We also included relatively new techniques, such as support vector machines. The computer code necessary to run these models ran well over a thousand lines.
A handful of these models proved better than the benchmarks of coin flips and educated guesses, and also better than the income models. There are two aspects of these that are notable: first, they tend to be the simpler models, with more accessible intuition regarding the statistical process underlying them. Second, they benefited from being run, tuned with additional parameters, re-run, examined, adjusted to include a different set of variables, and then run again, all under the care of an economist familiar with the underlying business.
Among these two distinctions, we believe the latter is more important. The machine learning models that did reasonably well were those that could be understood, and were adjusted to improve their results as well as the understanding of the results.
3.6 Coin flips
It is always important when evaluating new methods to consider a common, low-cost alternative.
It turns out a slightly lucky coin flip would produce validation accuracy of just above 50% in this sample, and even a slightly unlucky coin flop produces accuracy of above 40%.
These are important benchmarks, and allow for an assessment of whether the extraordinary time and computing effort necessary to use sophisticated methods are worthwhile.
3.7 Machine learning models
Coming in last place are machine learning models used without expert tuning. Each of the over a dozen different machine learning methods that were attempted, some of which involved multiple variations, were initially provided the same extensive dataset on the same set of cities. They were also run using the same software, by the same author and set of collaborators.
However, the difference in the predictive accuracy between those run with modest intervention by a trained business economist, and those that were not, was stark. Some models failed miserably, achieving validation accuracy worse than an unlucky coin flip. Some did approximately as well as an educated guess.
Others just failed to run effectively, often due to multi-collinearity, categorical variables, or missing observations.
Based on these results of the natural experiment of the Amazon HQ2 selection, and the use of different models relying on the same data, we conclude as follows:
1. Attributes grounded in economic fundamentals are strong predictors of location decisions.
Among our competing approaches to explaining major business decisions, the most accurate was an index developed by expert business economists. The index used in our comparison included objective data on workforce, tax costs, facility costs, and transportation availability--all fundamental economic indicators. This index did not consider marketing efforts, political support, incentives, nor largely subjective considerations. Furthermore, two other expert predictions used similar reasoning and predicted a similar set of cities.
2. Machine learning models pose a serious risk when not used with professional judgment.
One striking result from this comparison is that naively used machine learning models can do worse than a coin flip or an educated guess. Furthermore, many machine models are opaque and provide little intuition regarding the underlying structure. As a result, it is difficult to decide whether to rely upon one rather than another. In this comparison, using the same data and set of cities, some models did well, others did poorly, and some did not run at all. Without the supervision of a person knowledgeable about the business, it would be difficult to choose which to rely upon. Thus, even though some machine learning models (using "supervised learning" procedures that required data on actual selections) did well on a simple validation metric, the approach itself was more risky than relying upon expert judgement, income models, or value models.
3. Relying upon profit maximization does not explain location decisions as well as available alternatives.
Another striking result is that traditional income models, based on financial statement analysis and consistent with neoclassical economic theory, do not outperform either expert judgement or value maximization models. All three of the expert predictions we utilize here took into account long-term objectives of Amazon that were not included in a standard financial model. The value maximizing model also considered at least one of these factors (the ability to hire additional trained workers in the future), and outperformed the income models.
Business managers have known for a long time that managerial flexibility, growth options, risk avoidance, and strategic considerations are critical to the value of a business. These results suggest, again, that the neoclassical premise should be re-examined when it comes to business decisions.
4. A value maximization model explains the location decision we examined quite well.
The value-in-place analysis performed better than an income-in-place analysis in predicting the 20 cities selected by Amazon, even though it relied upon exactly the same data on the candidate cities, and used exactly the same financial model for income under standard circumstances. The value model took into account the ability to expand (or not expand) in the future, and that the size of the talent pool in the area affected the cost of doing so. Incorporating the ability to choose in the future, and adding one additional cost factor, allowed this model to outperform a comparable income model.
These results suggest that value maximization, not machine learning or artificial intelligence, should be considered as an alternative to the neoclassical principle of profit maximization.
5. Human intelligence, not "artificial intelligence" or machine learning, still rules.
None of the simple financial models, or the sophisticated machine learning models, or even the novel value maximization models, beat the predictions of experts in this contest. The first lesson from this is that human intelligence still trumps "artificial intelligence" when it comes to major business decisions, even when sophisticated models and a large amount of data are available.
4.1 Postscript: final selections
Following completion of this research, Amazon announced that their final selections: a split between Queens, New York and Arlington, Virginia (Amazon 2018). (18) Although we focused our empirical test on the selection of 20 cities, it is interesting to note that the final selection of 2 cities corroborated our results. In particular:
* Expert judgement again was a strong predictor. The Anderson HQ2 Index ranked the selected MSAs 1 st and 6th among all contenders; Brookings' (unranked) list included both; and Moody's included one.
* The value model also performed quite well. Using a model of a value-maximizing business resulted in a high preference for a city within the New York MSA, but not high-cost Manhattan, such as Newark, NJ. Although we did not model Arlington (or Queens) specifically, in both cases the value model preferred these locations within the larger MSAs of Washington DC and New York City.
* Machine learning again severely under performed expert judgement and the value model. Only two of the machine learning models accurately predicted New York and Washington DC. Furthermore, Amazon's selections (of both 20 finalists out of over 200 candidates, and then 2 selections out of the 20 finalists) had an identifiable logic. In contrast, the results of the various machine learning models were-literally and figuratively-all over the map.
HQ2 index of economic attributes
The HQ2 Index was compiled by the consulting firm of Anderson Economic Group and originally published in advance of the selection date. (19) It is described by the firm as follows:
In Amazon's request for proposals (RFP), they emphasized the following items, among others:
--Metropolitan area with more than one million people
--A stable and business-friendly environment and tax structure
--Potential to attract and retain strong technical talent
--A highly educated labor pool and a strong university system
--Proximity to international airport, major highways, and mass transit
Recreational opportunities, educational opportunities, and high quality of life
Using measurable factors from the lists above, we compiled the AEG HQ2 index, which captured a city's measurable advantage in attracting Amazon's HQ2. For the 35 cities in the United States that met specific requirements from the RFP, we estimated their performance using 11 total metrics across three broad categories:
Access to Labor and Services, including four indicators: degrees granted in relevant fields of study by colleges; employment of workers in specific occupational categories; size of the business services industry; and the number of migrants with bachelor's degrees from other counties, states, and countries.
Ease of Transportation, including two indicators: hours of delay due to traffic congestion, and per-capita use of public transit systems.
Cost of Doing Business, including five indicators: state and local business taxes (using Anderson Economic Group's Business Tax Burden studies); rental costs for commercial real estate; and unit cost of labor in 3 occupations important to Amazon. Note that this measure of labor costs takes into account worker productivity, meaning that more productive workers can be ranked higher even if their wages are also higher.
The AEG HQ2 index is the average of the values for each category. For each category, a higher value is better than a lower value (e.g., a lower cost of doing business translates into a higher index).
We estimate the distributable profit for the operations of the proposed HQ2 facility, in each city or metro, and for each stage of operation. We use the following method:
1. To facilitate comparisons among cities and metropolitan areas, the output of the hypothetical facility is standardized across all areas, while the costs and productivity vary. This allows for direct comparisons of the likely profitability of a standardized facility located in multiple cities.
2. We use the assumptions described above regarding the underlying business, and our analysis of Amazon Inc. business data, to estimate the revenue. We use these data and assumptions, plus the data for costs and productivity for each metro and city, to estimate costs.
* COGS (cost of goods sold), which combines Amazon's "cost of sales" and "fulfillment" categories; we assume the COGS related to the products and services generated or otherwise arising from the HQ2 operations will be largely the same across all metro areas, and at all stages of operation.
* Operating expenses, which combines "marketing," "technology and content," and "general and admin" costs. We assume this stays stable (as a share of revenue) at all stages of operation.
* Other expenses, includes mainly amortization of intangible assets.
* Facility costs related to this specific facility, which includes cost of building, permitting, moving employees, training and hiring, and the rent or rent equivalents as well as property taxes, utility, and service costs for the facility itself. We assume this changes (increases as a share of revenue) by stage of operation, for two reasons: larger facilities require more land acquisition and longer construction lead time; and we expect incentives from state and local governments to be focused on the first stage (and possibly the second).
Because the operating costs (including wages and taxes) vary by place, and the facility costs (including rent and construction costs) vary by both place and stage of operations, each of these income statements will be different.
The value model uses the same input data as the income model, and the same income statement for the proposed facility at each stage of operation. However, it also includes a set of decisions (actions) that a manager of the firm could take, including expanding and shutting down the facility. There is a cost associated with expansion, and that cost is related to the size of the workforce in the respective city.
We compose and solve a decision model for the proposed facility in each of the respective cities, again using the same data as the income model.
The value model requires a "reward matrix" that captures the current income for each stage of production, and with each business decision. We create this matrix from the income statements for each city and stage, and costs for expansion or closure.
In addition to these elements, the model requires a time index, discount rate, growth factor, and transition matrix. We use a yearly time index and a small growth rate (g =.02) and reasonable discount rate for a corporation (d =. 15). The transition matrix is largely determined by the business decision itself.
The solution method can be expected to solve the resulting problem as it meets the conditions specified by (Anderson 2012) for a business value problem as composed in (Eq. 1). A value function iteration solution algorithm is used. The results of the optimization problem include, for each city, a value in that stage of operation, and a suggested business decision to achieve that value.
Machine learning models
The Machine Learning models used include:
* K nearest neighbors This classifier is a method for associating, clustering, or classifying data. This method selects the k "nearest" neighbors using a distance metric. The standard method is Euclidean distance. There are numerous variations among KNN routines, including on the distance metric, the number of neighbors, weighting of observations, and other factors. The KNN method is often a benchmark for other methods and has been used widely. For classification purposes, the method seeks a classification scheme that minimizes the distance among similarly classified data points.
* Naive Bayes classifier A classifier which applies Bayes' theorem on a collection of independent variables to classify features (or in this case, decisions).
* Ensemble trees This model creates an "ensemble" of classification learners, where each learner is a classification tree. The trees are "boosted" in this model by using a logistic function to evaluate the deviance (difference) between actual and predicted classifications. There are many alternative functions. To make more use of the data, this learner also "bags" trees by using a bootstrap (re-sampling) method.
* Fine classification tree This classifier uses Machine Learning algorithms to construct flexible, fast-to-estimate binary separations of the data into "decision trees." These "trees" may then be traversed with variable data to determine the corresponding prediction.
* Linear discriminant Discriminant classification analysis assumes that different classes generate data based on different Gaussian distributions. The fitting function estimates the parameters of a Gaussian distribution for each class, and then minimizes the classification error.
* Support vector machine classifier This method creates "support vectors" in hyperspace to attempt to separate the data. The use of a "kernel" that statistically transforms the data allows for it to be arrayed in manner that a separating hyperplane can be constructed from the support vectors.
* Logistic regression This classifier models probabilities as a function of the linear combination of predictors using a logarithmic cost function to determine the classifications.
Multiple versions of most models were run, and many such versions were run multiple times with different-sized datasets. The reported results contain the best-fitting models in each category where we had a least one version of the model that produced results (even if poor). Some models did not run at all.
A dimension-reduction technique (principle component analysis) was attempted with a number of models; it is reported for only one. In general, PCA did not improve the results, and made them even more difficult to interpret.
Software used for all models
The following software was used, which allowed for inputting the same data to all models, direct comparisons of the models, and for the use of the intermediate results from the income models in the value models. (20)
* Data collection and reporting: Microsoft Excel and Tableau were used to collect data from multiple sources, and to report and visualize the data. These same products were used for calculation in the HQ2 Index (including with augmented data).
* Exploratory Data Analysis: Matlab with Statistics and Machine Learning Toolbox.
* Income, Machine Learning, and Value Models: Matlab with the Statistics and Machine Learning toolbox were used for the income, machine learning, and value models.
* Value Models: The Rapid Recursive toolbox was used to compose and solve the value functional models.
Availability of extended results and base data
We intend to make available to all members of NABE (and subscribers to Business Economics) extended results in the following form:
* The base data used in the analysis, including all the data listed in the "Data Appendix," with the limitation noted below.
* A printout of the computer run for the income, value, and machine learning models, which include notes on all techniques, notes on the data, extensive intermediate results, additional EDA graphics, explicit parameter selections, and other information.
* Any written corrections or clarifications to the journal article and any revised versions of the dataset the author prepares for the purpose of documenting the work presented in the article.
This information will be made available at the Anderson Economic Group website (https://www.andersoneconomi cgroup.com) for at least 1 year after publication. The author cautions that, with over 50 variables from dozens of sources, some of the data used in this analysis will have been revised by the time of publication. The author wishes to acknowledge the extensive work by Brian Peterson of Anderson Economic Group, and Ervin Batka of Supported Intelligence, in collecting the data and running the machine learning and value maximization models.
Table 4 lists the data used in the machine learning, value, and income models presented in this article. The table also includes the variables used in the HQ2 Index that was one of the expert predictions.
The table shows the data short names, full names, and sources for each variable. These include outcome variables and three groups of explanatory variables: HQ2 Index variables, conventional economic indicators considered important for site selection, and quality of life and other variables. Additional data notes are included in "Special Data Notes."
Special data notes
If a region has an international airport, it was coded as a "2." If a region has an international airport with direct daily flights to Seattle, New York (JFK or LGA), San Francisco (SFO or OAK), and Washington DC (BWI or DCA), it was coded as a "3." This coding considers airports within 45 min of the MSA as per Amazon's RFP when determining access to direct daily flights.
New establishments 100+ and new establishments 500+
This variable shows the total number of new establishments emerging in a given MSA from 2010 to 2014 from new firms with greater than 100 employees. A derivative of this variable looks at the total number of new establishments emerging in a given MSA from 2010 to 2014 from new firms with greater than 500 employees.
Volunteer hours per capita, volunteer rate, and the share of population active in their neighborhood
These three metrics come from 2015 Volunteering and Civic Life in America, a dataset produced by the Corporation for National and Community Service (CNCS). CNCS is an independent federal agency that is dedicated to supporting the American culture of citizenship, service, and responsibility. The data were collected through two supplements to the U.S. Census Bureau's Current Population Survey (CPS)--the Volunteer Supplement (2015) and the Civic Supplement (2013). The data are reported by MSA before metro definitions were revised in 2015.
Volunteer rate is the percentage of individuals who responded on the Current Population Survey's Volunteer Supplement that they had performed unpaid volunteer activities at any point during the 12-month period that preceded the survey for or through an organization.
Walk score, transit score, and bike score
The scores are calculated by Walk Score for 141 largest core cities in the U.S. and Canada. Walk Score is a private tech company that originated in Seattle and is now owned by Redfin, a real estate agency.
Walk Score is designed to assess walkability in the area. The score analyzes walking routes to nearby amenities. Points are awarded based on the distance to amenities in each category. Amenities within a 5-min walk (.25 miles) are given maximum points. A decay function is used to give points to more distant amenities, with no points given after a 30 min walk. The score also measures pedestrian friendliness by analyzing population density and road metrics such as block length and intersection density.
Transit Score aims to reflect how well an area is served by public transport. Points are assigned to nearby transit routes based on the frequency, type of route, and distance to the nearest stop on the route.
Bike Score aims to reflect how convenient an area is for biking. Points are awarded based on availability of bike infrastructure (e.g., lanes, trails), hills, road connectivity, and the number of bike commuters. For each score, the points are summed and normalized to a score between 0 and 100. For details, see https://www.walkscore.com/.
Average hours of sunshine per year
The data for this metric come from the World Meteorological Organization Standard Normals dataset, accessed through United Nations Data portal. It measures the mean number of hours of sunshine per year for cities all over the world, including in the U.S. For most of the cities used in our analysis, the reported values are averages computed for the consecutive periods of 30 years, from 1961 to 1990. For details, see http://data.un.org.
Number of good air quality days per year
This metric is designed by U.S. Environmental Protection Agency and measures how clean or polluted air is by MSA, and whether the associated health effects might be a concern. The data are based on Air Quality Index (AQI), which focuses on measuring ground-level ozone and particle pollution. Days are evaluated as good, moderate, unhealthy for sensitive groups, unhealthy, very unhealthy, or hazardous. We use the number of days evaluated as 'good.' During 'good' days, air quality is considered satisfactory and poses little or no risk. For details, see https://www.epa.gov/outdo or-air-quality-data/air-quality-index-report.
Economic inclusion index and racial inclusion index
These indices are calculated by the Urban Institute for 274 largest cities across the U.S. The Urban Institute is a nonprofit research organization based in Washington, DC.
The Economic Inclusion Index measures the ability of residents with lower incomes to contribute to and benefit from economic prosperity. Among the indicators used to calculate the index are income segregation rank, share of renters who pay 35% or more of their income in rent, share of 16 to 19-year olds who are not in school and have not graduated, share of families that are below the poverty line with householder working full-time. Income segregation is computed through estimating the segregation between families above and below each income distribution bucket at the census tract level. The indicator values are then averaged (weighted by income comparative to the median income) to construct the city-level measure. The index is computed as an average of z-scores of these four indicators.
The Racial Inclusion Index measures the ability of residents of color to contribute to and benefit from economic prosperity. Among the indicators used to calculate the index are racial segregation, homeownership gap, educational attainment gap, poverty rate gap, and share of people of color. Racial segregation is calculated as (1/2) * ((# people of color in census tract/# people of color in city)--(# non-Hispanic white in census tract/# non-Hispanic white in city)). The homeownership gap is calculated as a difference between the share of white non-Hispanic households that own a home and the share of persons of color households that own a home. The education attainment gap is calculated as a difference between the share of white non-Hispanic population over 25 with a high school degree or more and the share of the person of color population over 25 with a high school degree or higher. The poverty gap is calculated as a difference between the poverty rate for white non-Hispanic population and the poverty rate for person of color population. The index is computed as an average of z-scores of these five indicators. For details, see https://apps.urban.org/ features/inclusion/.
Share of the foreign-born population
This metric is based on the 2012-2016 American Community Survey 5-Year Estimates, Table S0501, accessed through the American FactFinder. The share of foreign-born population is calculated by dividing the estimated number of foreign-born persons by the total population for each area in the analysis. For details, see https://factfinder.census.gov/.
Share of the population who speak only English at home
This metric is based on the 2012-2016 American Community Survey 5-Year Estimates, Table S0601, accessed through the American FactFinder. The share of the population that speaks only English at home is calculated by dividing the number of persons over 5 years old who speak only English at home by the total population over 5 years old. For details, see https:// factfinder.cen-sus.gov/.
The share of the population who speak English at home may be a more effective measure for ethnic and cultural diversity of the population than just the share of foreign-born residents. The share of the foreign-born population tells us only about the first-generation migrants, but the share of population who speak languages other than (in addition to) English at home also captures the children of the earlier generations of migrants who are likely to have preserved their cultural identity. The share of the population who speak only English at home is a proximate measure for the level of cultural homogeneity in a given area.
Rate of population growth 2010-2017
This metric is based on the data from Population Estimates Program by U.S. Census Bureau, accessed through American FactFinder. To find the rate of population growth, we divided the 2017 population estimate by the 2010 population estimate and subtracted one from the result. For details, see https:// fact-finder.census.gov/.
Violent and property crime rates
For this metric, we use 2014 crime rates as reported by Uniform Crime Reporting Statistics (UCR), U.S. Department of Justice. Crime rate is defined as the number of crimes per 100,000 residents. The crime rates are reported by UCR at the core city level and come from respective city agencies. Violent crime includes murder, rape, robbery, aggravated assault. Property crime includes burglary, larceny-theft, motor vehicle theft. For details, see https://www.bjs.gov/ucrdata/Searc h/Crime/Crime.cfm.
Median housing cost per month
This metric is based on the 2012-2016 American Community Survey 5-Year Estimates, Table B25105, accessed through American FactFinder.
Share of the population who reported mental distress and share of the population who reported bad physical health in the last 30 days
These metrics report the age-adjusted 2015 estimates from Local Data for Better Health dataset produced by Centers for Disease Control and Prevention, a U.S. federal agency under the Department of Health and Human Services. The dataset contains information for 500 largest core cities and was released in 2017.
The share of population who reported mental distress 14 or more days in the last 30 days was computed by dividing the number of respondents age 18 years or older who report 14 or more days during the past 30 days during which their mental health was not good, by the total number of respondents. The share of population who reported bad physical health 14 or more days in the last 30 days was computed by dividing the number of respondents aged 18 years or older who report 14 or more days during the past 30 days during which their physical health was not good by the total number of respondents. For details, see https://chronicdata.cdc.gov/.
Park Score is calculated by the Trust for Public Land, a U.S.-based Non-governmental Organization dedicated to creating and improving neighborhood parks. The score assesses quality and accessibility of parks in the 100 most populous core cities in the U.S.
The indicators behind the score are grouped into four areas: park acreages, investment, amenities, and access. For acreage, the indicators include median park size and parkland as a share of city area. For investment, the indicators include public spending, non-profit spending, and monetized volunteer hours worked any public parks and recreation agencies. For amenities, the indicators include the number of park amenities per capita, with amenities defined as playgrounds, rest rooms, dog parks, splash pads, recreation and senior centers, and basketball hoops. For access, the indicator is the share of population living within a 10-min walk of residence. Cities can earn a maximum score of 100. For details, see http://parkscore.tpl.org.
The data for this metric come front the Center for American Women and Politics (CAWP) at Rutgers Eagleton Institute of Politics. A value of "1" indicates the central city of the MSA in question had a female mayor as of March 2018, and "0" indicates that the central city had a male mayor. For details, see http://www.cawp.rutgers.edu/levels_of_office/ women-mayors-us-cities-2018.
Share of female members in the city council
The data for this metric were collected from individual city council websites for central cities of the MSAs. We divided the number of female members by the total number of members for that council. Mayors were excluded from the calculations.
Share of employees in arts, entertainment, and culture
This metric is based on the 2012-2016 American Community Survey 5-Year Estimates, accessed through American FactFinder. The share of the employees who work in arts, entertainment, and culture industries was calculated by dividing the number of persons who reported employment in these industries by the total population who reported employment. For details, see https://fact-finder.census.gov/.
Government spending, taxation, and labor market freedom scores by state
The three scores are the components of the Economic Freedom of North America (EFNA) Index reported by the Fraser Institute. The Fraser Institute is a think tank headquartered in Vancouver, British Columbia, that produces research about government actions in areas such as taxation, health care, aboriginal issues, education, economic freedom, energy, natural resources, and the environment.
Government Spending scores are designed to reflect the size of the government. Each score is calculated based on the following indicators: general consumption expenditures by government as a percentage of income, transfers and subsidies as a percentage of income, and insurance and retirement payments as a percentage of income. Taxation scores are aimed at assessing the tax burden. The score is calculated based on income and payroll tax revenue as a percentage of income, top marginal income tax rate and the income threshold at which it applies, property tax and other taxes as a percentage of income, and sales taxes as a percentage of income. Labor Market Freedom scores are based on minimum wage legislation, government employment as a percentage of total state/provincial employment, and union density. For each score, states/provinces in the U.S., Canada, and Mexico are included in the analysis, and are awarded points on a scale of 0-10. For details, see https://www.fraserinstitute.org/studies/ economic-freedom-of-north-america-2017.
Economic freedom index by MSA
This index comes from a 2013 article by Dean Stansel in Journal of Regional Analysis and Policy. To calculate each score, Stansel used the model of 2011 Economic Freedom of North America (EFNA) Index by the Fraser Institute. As in the EFNA scoring system, points are awarded to MSAs on a scale of 0-10.
While EFNA reports scores only for states/provinces, Stansel's study uses the model to assess economic freedom at a more granular level. In contrast to EFNA, this was a one-time study and it only includes areas within the United States. While most of the scores in the article are reported by Metropolitan Statistical Area, some are reported by metropolitan statistical division instead. We used the data for MSAs but, when information was available only for metropolitan statistical divisions, we selected the divisions where the central city of the relevant MSA was located. For details, see http://www.jrap-journal.org/pastvol-umes/2010/v43/ index431.html.
Rate of new entrepreneurs, opportunity share of new entrepreneurs, and startup density
The data for these three variables come from the 2016 Kauffman Index for Startup Activity produced by the Kauffman Foundation. The Kauffman Foundation focuses its work on education and entrepreneurship.
The rate of new entrepreneurs measures the share of adult population that became entrepreneurship in a given month. The opportunity share of new entrepreneurs measures the share of new entrepreneurs who were not unemployed or in school prior to becoming entrepreneurs. The startup density measures the number of startups per 1,000 firms, where startups are defined as businesses less than 1 year old that employ at least one person beside the owner. For details, see https://www.kauffman.org/kauffman-index.
Number of days with pleasant temperatures per year
For this metric, we used the Global Surface Summary of the Day (GSOD) database of the U.S. National Oceanic and Atmospheric Administration, accessing it through NCEI Climate Data Online Data Search. The days were counted as pleasant if the daily mean temperature was between 59 and 77 F, the maximum temperature did not exceed 85 F, and the minimum temperature did not fall below 45 F. We looked at the time period from 1/1/2007 to 12/31/2017 or smaller periods if data from a single station were unavailable. We then divided the total number of pleasant days by the number of years within the time period considered.
The values recorded by weather stations in or close to the central cities of relevant MSAs were chosen for analysis. Most of the stations chosen were located in international airports to maximize completeness and reliability of data. For details, see https://www7.ncdc.noaa.gov/CDO/cdose lect.cmd.
Republican and Democrat votes in the 2016 presidential election
This metric is based on the final official 2016 presidential election result data reported by the state and county authorities (e.g., boards of elections, county clerks) on their websites. The data are reported by county. For each MSA, we only include the county where the central is located. We calculated vote percentages by dividing the counts of Republican and Democrat votes by the number of the total votes cast. In cases where the necessary data were not easily accessible on a government entity website, we used information from NPR Election 2016 Results special series.
This index is produced by Gallup-Sharecare annually since 2008, based on a survey of 175,000-(-respondents. The scores and ranks are reported by MSA.
The survey questions used to calculate the index are associated with one of the five elements of well-being. Among the five elements of well-being that Gallup-Sharecare chose to include are purpose (liking what one does, being motivated to achieve their goals), social (having supportive relationships, love), financial (managing one's economic life to minimize stress and increase security), community (liking where one lives, feeling safe and proud of one's community), and physical condition. Gallup categorizes the respondents as thriving, struggling, or suffering for each of the five elements. For details, see https://wellbeingin-dex.sharecare. com/.
Number of major professional sports league teams
This metric reflects the number of professional football, basketball, baseball, and hockey teams headquartered in each MSA included in the analysis that belong to NFL, NBA, MLB, and NHL, respectively. We found the number of teams by consulting NFL, NBA, MLB, and NHL websites.
Data estimates for Toronto
Since Toronto is not included in most data, we made manual estimates. We describe our methodology for the estimates below.
Average hours of sunshine per year The measure is reported as "Total Hours of Bright Sunshine." It is calculated by the Government of Canada using 1981-2010 station data from Toronto. This source was used in lieu of missing data from the World Meteorological Organization Standard Normals dataset for hours of sunshine. Both sources were last updated in 2010.
Number of good air quality days per year This variable counts the number of "low risk" days (values of 1-3) on the Ontario Ministry of the Environment and Climate Change's Air Quality Health Index as reported at the Toronto Downtown station. All values are from 2017 (like data from the EPA).
Rate of population growth This measure reflects the population percentage change from 2011-2016. This estimate is slightly distinct from data for MSAs in the U.S., which cover 2010-2017. That said, Statistics Canada's population by census subdivision exists for census years (2016, 2011, etc.). Quarterly data that could be used to find the 2010-2017 rate of growth are on the country, territory, and providence level.
Share of foreign-bom population Data on foreign-born individuals come from Statistics Canada's 2016 Census that looks at the total number of immigrants born outside of Canada in Toronto. Toronto's population is from Statistics Canada's 2016 Census.
Share of employees in arts, entertainment, and culture This measure is calculated by dividing the number of employees in arts, entertainment, and culture by total number of employees in Toronto. The former value comes from Statistics Canada's 2016 census, which counts the number of workers in "Occupations in art, culture, recreation and sport" with its "National Occupational Classification." The total number of employees in Toronto comes from their count of the "Total labor force population aged 15 years and over."
Median housing costs per month This value is the "median monthly shelter costs for rented dwellings" as reported in Dollars from Statistics Canada's 2016 Census. It is for Toronto only. Rented dwellings were taken because data did not exist for both rented and owned dwellings.
Share of population who speak only English at home This percentage is calculated by dividing the number of individuals with knowledge of only English in Toronto by the total population of Toronto excluding institutional residents (as data on their knowledge of languages were not collected). Data were from Statistic Canada's 2016 census.
Alphabet, Inc. Form 10-K, 2017. 2018. Alphabet investor relations website.
Amazon. 2018. "Amazon Announces Candidate Cities for HQ2." January 18, retrieved from www.amazon.com.
Amazon, Inc. 2017. Amazon HQ2 RFP. (September 7) www.images-na. ssl-images-amazon.com.
Amazon, Inc. 2016-2017. Form 10-K (annual results). Securities and Exchange Commission.
Amazon, Inc. 2016-2017. Form 10-Q (quarterly results). Securities and Exchange Commission.
Anderson Economic Group. 2017. 2017. The HQ2 Index (October); and Updated: The Anderson Economic Group HQ2 Index (February 2018). Both retrieved from www.andersoneconomicgrou p.com.
Anderson Economic Group. 2017. 2017 Slate Business Tax Burden Rankings Report, www.andersoneconomicgroup.com.
Anderson Economic Group. 2018. 2018 State Business Tax Burden Rankings Report. Retrieved from www.andersoneconomicgrou p.com.
Anderson Economic Group. 2018. Supplemental Information on Professional Sports Teams Acquired from Websites of NFL, NBA, MLB, and NHL Professional Sports Teams; Supplemental Information on Female Mayors and Female City Council Membership Acquired from Websites of Relevant Cities; County Government Election Official Websites of the Central Cities for HQ2 Analysis Regions; Latitude and Longitude for Cities.
Anderson, Patrick. 2012. Economics of Business Valuation. Stanford: Stanford University Press.
Anderson, Patrick. 2014. Policy Uncertainty and Persistent Unemployment: Numerical Evidence from a New Approach. Business Economics 49(1): 2-20.
Burwell v. Hobby Lobby, et al. 2014. No. 13-354, 573 U.S. 723 F. 3d 1114.
Corporation for National and Community Service. 2015. Volunteering and Civic Life in America, www.nationalservice.gov/vcla.
Cover, T.M., and P.E. Hart. 1967. Nearest Neighbor Pattern Classification. IEEE Transactions on Information Theory 13 (1): 21-27.
Dixit, Avanish, and Robert S. Pindyck. 1994. Investment Under Uncertainty. Princeton: Princeton University Press.
Dodge v. Ford Motor Company. 1919. 204 Mich. 459, 170 N.W. 668.
Federal Transit Administration. 2017. National Transit Database Monthly Module, August 2016-July 2017. www.transit.dot.gov.
Flightview. 2018. Airport information for HQ2 analysis regions, acquired from Flight- view.com, 2018.
Fraser Institute. 2017. Economic Freedom of North America Index, www.fraserinstitute.org/studies/economic-freedom-ofnorth-america-2017.
Friedman, Milton. 1953. The Methodology of Positive Economics, in Essays in Positive Economics, 3--13. Chicago: Chicago University Press.
Friedman, Thomas. 2005. The World is Flat. New York: Farar, Straus & Giroux.
Fujita, Masahisa, Paul Krugman, and Anthony Venables. 1999. The Spatial Economy: Cities, Regions and International Trade. Cambridge: MIT Press.
Gallup Organization. 2017 Gallup-Sharecare Wei I-Being Index, www. wellbeingindex.sharecare.com.
Galton, Francis. 1886. Anthropological Miscellanea: Regression towards Mediocrity in Hereditary Stature. The Journal of the Anthropological Institute of Great Britain and Ireland 15: 246-263.
Ghemawat, Pakaj. 2007. Why the World Isn't Flat. Foreign Affairs.
Harvard Business School Institute for Strategy and Competitiveness. 2015. Business Services Cluster Employment. U.S. Cluster Mapping Project 2015. www.clustermapping.us.
Hotelling, Harold. 1929. Stability in Competition. Economic Journal 39 (153): 41-57.
Jones Lang LaSalle. 2017. Office Outlook Q4 2016. www.us.jll.com/ united-states/en-us.
Jones Lang LaSalle. New Jersey Office Statistics Q4 2016. www.us.jll. com/united-states/en-us.
Jones Lang LaSalle. 2017. Washington, D.C. Office Statistics Q4 2016" www.us.jll.com/united-states/en-us.
Kaldor, Nicholas C. 1966. Marginal Productivity and the Macro-Economic Theories of Distribution: Comment on Samuelson and Modigliani. The Review of Economic Studies 33 (4): 309-319.
Kauffman Foundation. 2016. Index of Startup Activity Metropolitan Area Rankings, www.kauffman.org/kauifman-index/reporting/ startup-activity.
Krugman, Paul. 1991. Increasing Returns and Economic Geography. Journal of Political Economy 99 (3): 483-499.
Ljungvist, Lars, and Thomas J. Sargent. 2012. Recursive Macroeconomic Theory, 3rd ed. Cambridge: MIT Press.
Lucas, Roberts. 1976. Econometric Policy Evaluation: A Critique. In The Phillips Curve and Labor Markets (ed. K. Brunner and A. Meitzer), 19-46. New York: American Elsevier.
Modigliani, Franco, and M. Miller. 1958. The Cost of Capital, Corporation Finance and the Theory of Investment. American Economic Review 48 (3): 261-297.
Marshall, Alfred. 1890. Principles of Economics. London: MacMillan.
National Center for Education Statistics. Integrated Postsecondary Education Data System: Degrees Granted 2015-2016. www.nces. ed.gov/ipeds.
Parilla, Joseph. 2017. Who is best positioned to land Amazon's HQ2? Brookings Metropolitan Policy Council, www.brook-ings. edu/blog/the-avenue/2017/09/08/which-cities-are-well-positioned -to-land-amazons-hq2. Accessed July 2018.
Ricardo, David. 1817. Principles of Political Economy and Taxation. London: J.M. Dent & Sons.
Rubenstein, Mark. 2006. A History of the Theory of Investments: My Annotated Bibliography. New Jersey: Wiley.
Rutgers University Center for American Women and Politics. 2018. Women Mayors in U.S. Cities, 2018. www.cawp.rutgers.edu/level s_of_office/women-mayors-us-cities-2018.
Samuelson, Paul A. 1983. Thiinen at Two Hundred. Journal of Economic Literature 21: 1468-1488.
Stansel, Dean. 2013. An Economic Freedom Index for U.S. Metropolitan Areas. Journal of Regional Analysis and Policy 43 (1): 3-20.
Stokey, Nancy, Robert Lucas, and Edward Prescott. 1989. Recursive Methods in Economic Dynamics. Cambridge: Harvard University Press.
Texas Transportation Institute. 2015. 2015 Urban Mobility Scorecard. www.mobility.tamu.edu/ums.
Tobin, James, and W.C. Brainard. 1977. Asset Markets and the Cost of Capital. Cowles Foundation Discussion paper no. 427. www. cowles.yale.edu.
Thunen, Johan Heinrich. 1826. Von Thiinen's Isolated State. Oxford: Pergamon Press.
Trust for Public Land. 2018. Park Score, www.tpl.org/10minutewalk.
United Nations World Meteorological Organization. 2010. Standard Normals. www.wmo.int/pages/prog/wcp/wcdmp/GCDS_l.php.
U.S. Bureau of Labor Statistics. 2016. Occupational Employment Statistics. www.bls.gov.
U.S. Bureau of Economic Analysis. 2015. GDP & Personal Income data by State, Retrieved from www.bls.gov.
U.S. Census Bureau. 2010-2014 Business Dynamics Statistics, www. cen-sus.gov/ces/dataproducts/bds.
U.S. Census Bureau. 2016. American Community Survey 1-Year Table B01003: Total Population, www.factfinder.census.gov.
U.S. Census Bureau. 2016. American Community Survey 5-Year Table B01003: Total Population, 2012-2016. www.actfinder. census.gov.
U.S. Census Bureau. 2016. American Community Survey 1-Year Table S0701: Geographic Mobility by Selected Characteristics. www.fact-finder.census.gov/.
U.S. Census Bureau. 2016. American Community Survey 5-Year Table B25105: Median Monthly Housing Costs, www.factfinder.census.gov.
U.S. Census Bureau. 2016. American Community Survey. 5-Year Table S0501 : Selected Characteristics of the Native and Foreign-Born Populations, www.factfinder.census.gov.
U.S. Census Bureau. 2016. American Community Survey 5-Year Table S1601 : Language Spoken at Home, www.factfinder.censu s.gov.
U.S. Census Bureau. 2016. American Community Survey 5-Year Table S2403: Industry by Sex for the Civilian Employed Population 16 Years and Over, www.factfinder.census.gov.
U.S. Census Bureau. Table PEPANNRES: Population Estimates, 2010-2017. www.factfinder.census.gov.
U.S. Center for Disease Control and Prevention. 2017. Local Data for Better Health, www.cdc.gov/500cities.
U.S. Department of Justice. 2014. Uniform Crime Reporting Statistics. www.ucrdatatool.gov.
U.S. Environmental Protection Agency. 2017. Air Quality Index Report. www.epa.gov/outdoor-air-quality-data/air-quality-index-report.
U.S. National Oceanic and Atmospheric Administration. 2018. Temperature Data for HQ2 Analysis Regions, www.ncdc.noaa.gov.
Urban Institute. 2015. Economic Inclusion Index, www.apps.urban.org/ features/inclusion.
Urban Institute. 2015. Racial Inclusion Index, www.apps.urban.org/ features/inclusion.
Walkscore. 2018. Walk Score, Transit Score, and Bike Score for Various Cities and Regions, www.walkscore.com.
Zandi, Mark. 2017. Where Amazon's Next Headquarters Should Go, and Metro Analysts on Amazon's Top Cities, www.economy.com/ dismal/analysis/commentary/298321 /Where-Amazons-NextHeadquarters-Should-Go. Accessed 12 Oct 2017.
(1) In economics, at least since the first edition of Alfred Marshall's Principles of Economics (Marshall 1890), the neoclassical school has been the dominant approach to Microeconomics. One should not overlook the same in related disciplines. In finance, the famous Modigliani-Miller principles (Modigliani and Miller 1958) rely upon the supposition of profit-maximizing firms. The wide-ranging intellectual history authored by (Rubenstein 2006) identifies Milton Friedman's influential "survivor" argument in (Friedman 1953), which asserts that any other approach to business management is doomed to fail.
Friedman's argument finds much support in standard business management methods. Inventory management methods (such as "economic order quantity" models), advertising management methods (such as the classic "customer lifetime value" models), and production management methods (particularly in the objective function for the enterprise) commonly rely upon the presumption that businesses maximize profits.
(2) The seminal works include Principles of Political Economy and Taxation (Ricardo 1817) and Von Thiinen's Isolated State (von Thiinen 1826; translated 1966). Both Ricardo and von Thiinen modeled the decision of a landowner in an agricultural society as a profit maximization problem. No less a luminary than Paul Samuelson observed the debt modern economics owes to von Thiinen, in his bicentennial essay (Samuelson 1983).
(3) The new economic geography is associated with Krugman (1991) and Fujita et al. (1999). Before the "new" economic geography, Hotelling (1929) described the classic location selection method for a profit-maximizing firm.
(4) The seminal case supporting the principle of the profit-maximizing firm is Dodge v Ford Motor Company, Michigan Supreme Court, 1919, in which Henry Ford faced a shareholder revolt led by the Dodge brothers. That court concluded as follows:
A business corporation is organized and carried on primarily for the profit of the stockholders. The powers of the directors are to be employed for that end. The discretion of directors is to be exercised in the choice of means to attain that end, and does not extend to a change in the end itself, to the reduction of profits, or to the non-distribution of profits among stockholders in order to devote them to other purposes.
Almost a century later, in the Burwood v. Hobby Lobby (2014) case, the US Supreme Court confirmed that a shift had taken place:
While it is certainly true that a central objective of for-profit corporations is to make money, modern corporate law does not require for-profit corporations to pursue profit at the expense of everything else, and many do not do so.
(5) The term "real options" began with Myers (1977); the intellectual groundwork for their importance was laid by Dixit and Pindyck (1994).
(6) This controversy is epitomized by the differences between Thomas Friedman's book The World is Flat (Friedman, 2005) and Pakaj Ghemawat's rejoinder journal article in Foreign Affairs "Why the World Isn't Flat" (Ghemawat 2007).
(7) Examples include, at various times, Microsoft, Google, and Amazon. Alphabet (now the parent company of Google), states their rationale for such a policy explicitly in the annual report for the year 2017:
We have never declared or paid any cash dividend on our common or capital stock. We intend to retain any future earnings and do not expect to pay any cash dividends in the future (Alphabet 2018, p. 21, emphasis in original).
(8) While profitability is obviously directly related to shareholder value, it is not identical--as will be demonstrated in the differing results for income and value models in this article, even though they use the same underlying data.
In addition, as is now recognized in both the "real options" literature and in formal guidance for banks in most countries, prudent managers will often deliberately, prudently, and rationally reduce profits to achieve other objectives.
(9) Mark Zandi, "Where Amazon's Next Headquarters Should Go," and "Metro Analysts on Amazon's Top Cities," Moody's Analytics, October 12, 2017.
(10) Joseph Parilla, "Who is best positioned to land Amazon's HQ2?" Brookings Metropolitan Policy Council, September 2017.
(11) Anderson Economic Group, "The HQ2 Index," October 2017; and "Updated: The Anderson Economic Group HQ2 Index," February 2018.
(12) Among the top ten in all three experts' rankings were the New York, Boston, Philadelphia, and Atlanta MSAs. There were differences, of course; Moody's picked Rochester NY as fourth on their list, while that city did not make the top 20 for Anderson or Brook ings; only Brookings initially considered Canadian cities, and included both Vancouver and Toronto; the Anderson rankings when augmented later included Toronto in the top 20, but not Vancouver.
(13) The RFP stated "an international airport with daily direct flights to Seattle" and other specific cities were "an important consideration." Apparently, metropolitan economic development officials considered the scheduling of such flights to be an easy decision should Amazon locate their HQ2 in their areas.
(14) These are derived from annual and quarterly reports (Amazon 2016, 2017).
(15) The "Lucas critique" of Keynesian macroeconomics (Lucas 1976) is one motivation for the use of recursive techniques in modeling economic decisions.
The textbook Recursive Methods in Macroeconomics (Ljungqvist and Sargent 2009) describes both the theoretical underpinnings and numerous uses in academic economics.
(16) We will note here that, for some of the machine learning models for which results are presented below, the economist tuned the model by adjusting parameters and selecting variables.
These we call "machine learning with professional judgement" in Sect. 3.
(17) The essay by Kopf (2015) collects some of the letters of Gauss and notes the controversy over who invented the method commonly called regression. It appears Gauss and Legendre both used the method of least squares, but Galton was the one who publicized the term "regression," including in his 1886 article "Regression towards Mediocrity in Hereditary Stature."
The term "mediocrity" is often deleted from contemporary discussions of this technique, in favor of "mean." As Kopf notes, "unfortunately," Galton sometimes used statistical techniques in support of the "science" of "eugenics," which is a term he coined.
(18) Amazon blog, "Amazon selects New York City and Northern Virginia for new headquarters," Amazon, November 13, 2018.
(19) See Anderson Economic Group (2017, 2018).
(20) These software packages are available from Microsoft, Tableau, The MathWorks, and Supported Intelligence.
Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Patrick L. Anderson is principal & CEO of Anderson Economic Group LLC, a business consulting firm that serves clients in multiple countries. He is the author of Economics of Business Valuation (Stanford University Press, 2012) and many other works. This and two prior articles in Business Economics received the annual Edmund Mennis award from the National Association for Business Economics.
Patrick L. Anderson [1, (iD)]
Published online: 6 February 2019
[email] Patrick L. Anderson
 Anderson Economic Group LLC, East Lansing, MI, USA
Caption: Fig. 1 Exploratory Data Analysis of Cost Data
Caption: Fig. 2 Exploratory Data Analysis of Workforce and Transportation Data
Caption: Fig. 4 Distributable profits for standardized facility, by city
Caption: Fig. 5 Value in place results for all cities
Table 1 Predictions of experts in advance of Amazon's selection. Sources: (a) Business Wire, "Amazon Announces Candidates for HQ2," January, 2018. (b) Anderson Economic Group, "HQ2 Index," October 2017; updated in January 2018. (c) Moody's Analytics, "Where Amazon's Next Headquarters Should Go," October, 2017. (d) Brookings Institution, "Which Cities are well Positioned to Land Amazon's HQ2?" September, 2017 City/region Amazon HQ2 Anderson Economic finalist (a) Group HQ2 index (top twenty) (b) Atlanta X X Austin X Baltimore X Birmingham Boston X X Buffalo Charlotte X Chicago X X Cincinnati X Cleveland X Columbus X (A) Dallas X X Denver X X Detroit Grand Rapids Hartford Houston Indianapolis X Jacksonville Kansas City Las Vegas Los Angeles X X Louisville Memphis Miami X Milwaukee Minneapolis X Montgomery County, MD X X Nashville X X New Orleans New York X X Newark X X Northern Virginia X X Oklahoma City Orlando Philadelphia X X Phoenix Pittsburgh X (A) Portland Providence Raleigh X X Richmond Rochester Sacramento Saint Louis Salt Lake City X San Antonio X San Diego X San Francisco San Jose Seattle Tampa X Toronto X Tucson Vancouver Virginia Beach Washington, DC X X Number predicted that match 14 Amazon selections: 16 with augmented data Memo: Number of cities included 20 23 as candidates (with MSAs desegregated): (A) denotes predictions if airport limitations were removed City/region Moody Analytics Brookings (top ten) (c) Institution (top twenty) (d) Atlanta X X Austin X X Baltimore X Birmingham Boston X X Buffalo Charlotte Chicago X Cincinnati Cleveland Columbus Dallas X Denver X Detroit X Grand Rapids Hartford Houston X Indianapolis Jacksonville Kansas City Las Vegas Los Angeles X Louisville Memphis Miami X Milwaukee Minneapolis X Montgomery County, MD X Nashville New Orleans New York X X Newark X X Northern Virginia X Oklahoma City Orlando Philadelphia X X Phoenix Pittsburgh X Portland X Providence Raleigh Richmond Rochester X Sacramento Saint Louis Salt Lake City X San Antonio San Diego X San Francisco X San Jose X Seattle Tampa Toronto X Tucson Vancouver X Virginia Beach Washington, DC X Number predicted that match 8 14 Amazon selections: Memo: Number of cities included 11 23 as candidates (with MSAs desegregated): (A) denotes predictions if airport limitations were removed Note on cities within MSAs: We include Newark within the "New York City" MSA, and Montgomery County and Northern Virginia in the "Washington, DC" MSA for this list Table 2 Income Statement calculations for selected cities. Sources Author's calculations, using data and income method described in text. All income statements are for stage 1 of the proposed HQ2 facility City Major income Projection statement concept ($millions) New York Revenue 4600 COGS 3344.2 Gross Profit 1255.8 OpExpense 1004.7 Facility Cost 165.67 OpProfit 85.405 OtherExp 2.3 Pretax Profit 83.105 dividend 62.329 Newark Revenue 4600 COGS 3344.2 Gross Profit 1255.8 OpExpense 1004.7 Facility Cost 85.438 OpProfit 165.64 Pretax Profit 163.34 dividend 122.51 Northern Virginia Revenue 4600 COGS 3344.2 Gross Profit 1255.8 OpExpense 986.32 Facility Cost 100.22 OpProfit 169.26 OtherExp 2.3 Pretax Profit 166.96 dividend 125.22 Table 3 Summary ranking of approaches by accuracy and reliability. Sources Data and methods described in text Method name Ex ante Ex post prediction validation accuracy accuracy Professional judgment 82% -- Value-in-place decision 64% -- model Income-in-place comparison 57% -- Machine learning with -- 60-80% expert tuning: ensemble trees (various boosting methods), fine classification tree Coin flips and Educated 46-54% 46-54% Guesses Machine Learning: Logistic -- 0-46% Regression, Linear Discriminant, Support Vector Machine, Models using PCA Method name Comment Professional judgment Professional judgments also avoided predicting most of the non-selected cities. This method also had the highest number of correct positive predictions. Value-in-place decision Does not select very model high-cost cities, but does select high value cities where costs are reasonable and expansion opportunities exist. Income-in-place comparison Uses standard financial statement analysis; consistently picks the lower-cost cities. Machine learning with These routines benefitted expert tuning: ensemble from an expert tuning the trees (various boosting model and selecting the methods), fine variables; they involve substantial professional classification tree judgement as well as machine learning. They also produced widely-varying predictions among the cities, and in some cases, mostly negative predictions. Coin flips and Educated This benchmark indicates the Guesses baseline below which it makes no sense to invest time in quantitative methods Machine Learning: Logistic These routines often produced Regression, Linear opaque results, and could not Discriminant, Support be expected to outperform a Vector Machine, Models coin flip. Some did not even using PCA run. Some predicted all negative results. Validation accuracy for Machine Learning models is done using k-fold cross-validation with k = 5 Machine Learning models are "supervised learning" done ex post; the Income and Value models are done without explicit use of the selection data Table 4 Data variables and sources Outcome variables HQ2_FINAL Amazon HQ2 finalist BDS_EST100 New establishments 100+ BDS_EST500 New establishments 500+ Explanatory variables Group 1 : AEG HQ2 index variables AEG_BTB 18 AEG Business Tax Burden 2018 AEG_BTB17 AEG Business Tax Burden 2017 UCOSTLABOR_INFO Unit cost of labor (Information) UCOSTLABOR_DATA Unit cost of labor (Data Processing) UCOSTLABOR_MGMT Unit cost of labor (Management of Businesses) COMM_RENT Commercial rent per square foot MASS_TRANSI T_TRIPS Transit trips per capita CONGESTION_HOURS Congestion (annual hours of delay per capita) DEGREES Completion of related degrees EMPL_RELATE D_OCC Employment in related occupations EMPL_BIZ_SER VICES Business services cluster employment IMM_W_BA Immigrants with bachelor's degree or higher AIRPORTS Airport qualified Group 2: Conventional economic indicators HOUSING_COST Median housing costs per month FOR_BORN Share of the foreign-born population ENG_ONLY Share of the population who speak only English at home EFNA_MSA Economic Freedom Index by MSA, EFNA_LABOR Labor Market Freedom score by state EFNA_GOV Government spending score by state EFNA_TAX Taxation score by state, LAT Latitude LONG Longitude Group 3: Quality of life and other variables VOL_HOURS Volunteer hours per capita VOL_SHARE Share of the population who volunteer ACTIV_SHARE Share of the population active in the neighborhood WALK_SCORE Walk Score TRAN_SCORE Transit Score BIKE_SCORE Bike Score SUN_HOURS Average hours of sunshine per year GOODAIR_DAYS Number of good air quality days per year ECON_INCL Economic inclusion index RACJNCL Racial inclusion index POP_GROWTH Rate of population growth 2010-2017 VIOL_CRIME Violent crime rate PROP_CRIME Property crime rate MENT_HEALTH Share of population who reported mental distress > = 14 days in the past 30 days PHYS_HEALTH Share of population who reported bad physical health > = 14 days in the past 30 days PARK_SCORE Park Score FEM_MAYOR Female mayor FEM_COUNCIL Share of female members in the city council EMPL_ART Share of employees working in arts, culture and entertainment NEW_ENTR_R ATE Rate of new entrepreneurs, Kauffman Foundation NEW_ENTR_0 PP Opportunity share of new entrepreneurs, Kauffman Foundation NEW_ENTR_S HARE Startup density, Kauffman Foundation GOODTEMP_D AYS Number of days with pleasant temperature per year, U.S. National Oceanic and Atmospheric Administration REPVOTE16 Share of Republican vote in 2016 presidential election DEMVOTE16 Share of Democrat vote in 2016 presidential election WELL-INDEX Well-Being Index, Gallup-Sharecare B4_TEAMS Number of major professional sports league teams Outcome variables HQ2_FINAL Amazon Chosen Candidates for HQ2, Amazon Press Room BDS_EST100 U.S. Census Bureau Business Dynamics Statistics, 2010-2014 BDS_EST500 U.S. Census Bureau Business Dynamics Statistics, 2010-2014 Explanatory variables Group 1 : AEG HQ2 index variables AEG_BTB 18 2018 State Business Tax Burden Rankings Report, Anderson Economic Group AEG_BTB17 2017 State Business Tax Burden Rankings Report, Anderson Economic Group UCOSTLABOR_INFO U.S. Bureau of Economic Analysis UCOSTLABOR_DATA U.S. Bureau of Economic Analysis UCOSTLABOR_MGMT U.S. Bureau of Economic Analysis COMM_RENT Office Outlook Q4 2016, Jones Lang LaSalle MASS_TRANSI T_TRIPS Calculated using the number of transit trips from Monthly Module, National Transit Database, Federal Transit Authority, August 2016-JuIy 2017, and population totals from 2016 American Community Survey 1 -Year Estimates CONGESTION_HOURS 2015 Urban Mobility Scorecard, Texas Transportation Institute DEGREES 2015-2016 academic year records, Integrated Postsecondary Education Data System (IPEDS) EMPL_RELATE D_OCC Occupational Employment Statistics, U.S. Bureau of Labor Statistics EMPL_BIZ_SER VICES U.S. Cluster Mapping Project, Institute for Strategy and Competitiveness, Harvard Business School IMM_W_BA 2016 American Community Survey 1-Year Estimates, U.S. Census Bureau AIRPORTS FlightView and the websites of relevant airports Group 2: Conventional economic indicators HOUSING_COST 2012-2016 American Community Survey 5-Year Estimates, Table B25105, U.S. Census Bureau FOR_BORN Calculated using estimates of foreign-born population and total population from 2012-2016 American Community Survey 5-Year Estimates, Table S0501, U.S Census Bureau ENG_ONLY Calculated using estimates of population over 5 years old who speak only English at home and total population over 5 years old from 2012-2016 American Community Survey 5-Year Estimates, Table S1601, U.S. Census Bureau EFNA_MSA Stansel, Dean. An Economic Freedom Index for U.S. Metropolitan Areas, Journal of Regional Analysis and Policy 43(1): 3-20, 2013 EFNA_LABOR Economic Freedom of North America Index, 2017, Fraser Institute EFNA_GOV Economic Freedom of North America Index, 2017 Fraser Institute EFNA_TAX Economic Freedom of North America Index, 2017, Fraser Institute LAT Google Maps LONG Google Maps Group 3: Quality of life and other variables VOL_HOURS Volunteering and Civic Life in America, 2015, Corporation for National and Community Service VOL_SHARE Volunteering and Civic Life in America, 2015, Corporation for National and Community Service ACTIV_SHARE Volunteering and Civic Life in America, 2015, Corporation for National and Community Service WALK_SCORE Walk Score, 2018 TRAN_SCORE Walk Score, 2018 BIKE_SCORE Walk Score, 2018 SUN_HOURS UN Data, World Meteorological Organization Standard Normals, 2010 GOODAIR_DAYS Air Quality Index Report, Environmental Protection Agency, 2017 ECON_INCL Urban Institute (2015) RACJNCL Urban Institute (2015) POP_GROWTH Calculated using annual resident population estimates 2010 and 2017 from 2017 Population Estimates, Table PEPANNRES, U.S. Census Bureau VIOL_CRIME Uniform Crime Reporting Statistics, 2014, U.S. Department of Justice PROP_CRIME Uniform Crime Reporting Statistics, 2014, U.S. Department of Justice MENT_HEALTH Age-adjusted 2015 estimates from Local Data for Better Health, 2017, Centers for Disease Control and Prevention PHYS_HEALTH Age-adjusted 2015 estimates from Local Data for Better Health, 2017, Centers for Disease Control and Prevention PARK_SCORE Park Score, The Trust for Public Land, 2018 FEM_MAYOR Women Mayors in 2018 U.S. Cities, Rutgers Center for American Women and Politics, March 2018 FEM_COUNCIL Websites of city councils, accessed June 2018 EMPL_ART 2012-2016 American Community Survey 5-Year Estimates, Table S2403, U.S. Census Bureau NEW_ENTR_R ATE Metropolitan Areas Rankings, Kauffman Index of Startup Activity, Ewing Marion Kauffman Foundation, 2016 NEW_ENTR_0 PP Metropolitan Areas Rankings, Kauffman Index of Startup Activity, Ewing Marion Kauffman Foundation, 2016 NEW_ENTR_S HARE Metropolitan Areas Rankings, Kauffman Index of Startup Activity, Ewing Marion Kauffman Foundation, 2016 GOODTEMP_D AYS Global Summary of the Day, U.S. National Oceanic and Atmospheric Administration REPVOTE16 Websites of county governments of the central cities for MSAs included in the analysis DEMVOTE16 Websites of county governments of the central cities for MSAs included in the analysis WELL-INDEX Gallup-Sharecare Well-Being Index, 2017 B4_TEAMS Websites of NFL, NBA, MLB, and NHL Fig. 3 Comparison of cost and productivity metrics Cost and Productivity Metrics for US Metro Areas Anderson Economic Group HQ2 Data, July 2018 Metro BTB Cost Labor Cost Labor Data Info Cost Factor Atlanta 0.06971 0.4383 0.4876 Austin 0.08313 0.4219 0.5471 Baltimore 0.08607 0.4457 0.5193 Birmingham 0.08428 0.4432 0.5193 Boston 0.08889 0.4801 0.5193 Buffalo 0.1126 0.4183 0.5193 Charlotte 0.06913 0.5075 0.6962 Chicago 0.09736 0.4432 0.3356 Cincinnati 0.07792 0.3699 0.4694 Cleveland 0.07792 0.3867 0.4331 Columbus 0.07792 0.4287 0.4384 Dallas 0.08313 0.3738 0.7105 Denver 0.08206 0.409 0.6173 Detroit 0.07516 0.4422 0.4904 Grand Rapids 0.07516 0.4322 0.5193 Hartford 0.09492 0.3211 0.4446 Houston 0.08313 0.4432 0.6256 Indianapolis 0.06996 0.4432 0.4308 Jacksonville 0.09571 0.5117 0.5193 Kansas City 0.07007 0.4432 0.5193 Las Vegas 0.101 0.407 0.2698 Los Angeles 0.08242 0.395 0.3507 Louisville 0.0953 0.391 0.634 Memphis 0.07947 0.3731 0.5193 Miami 0.09571 0.4358 0.521 Milwaukee 0.08016 0.4718 0.5193 Minneapolis 0.1034 0.4385 0.5566 Montgomery County 0.08607 0.4536 0.5193 Nashville 0.07947 0.3904 0.6636 New Orleans 0.08328 0.2943 0.5369 New York 0.1126 0.4131 0.5193 Newark 0.1126 0.4131 0.5193 Northern Virginia 0.08549 0.4536 0.5193 Oklahoma City 0.06622 0.4641 0.6055 Orlando 0.09571 0.4557 0.5193 Philadelphia 0.09106 0.5146 0.5193 Phoenix 0.08628 0.4829 0.6197 Pittsburgh 0.09106 0.4882 0.2417 Portland 0.06798 0.4903 0.6099 Providence 0.1235 0.4661 0.5193 Raleigh 0.06913 0.4506 0.5193 Richmond 0.08549 0.4055 0.4539 Rochester 0.1126 0.419 0.5193 Sacramento 0.08242 0.4699 0.5193 Saint Louis 0.07007 0.4781 0.7605 Salt Lake City 0.06987 0.4709 0.5193 San Antonio 0.08313 0.4281 0.7527 San Diego 0.08242 0.4727 0.4928 San Francisco 0.08242 0.5589 0.5193 San Jose 0.08242 0.6816 0.09339 Seattle 0.09188 0.4733 0.5193 Tampa 0.09571 0.4512 0.5193 Toronto 0.08685 0.44 0.49 Tucson 0.08628 0.4477 0.7007 Virginia Beach 0.08549 0.4242 0.5367 Washington DC 0.1247 0.4536 0.5193 Metro Cost Labor Rent SqFt Mgmt Atlanta 0.8162 23.91 Austin 0.8321 34.08 Baltimore 0.8321 23.34 Birmingham 0.8321 28.06 Boston 0.8073 33.99 Buffalo 0.9011 28.06 Charlotte 0.8321 24.47 Chicago 0.8433 30 Cincinnati 0.8321 19.35 Cleveland 0.8349 18.98 Columbus 0.8397 19.37 Dallas 0.8679 25.94 Denver 0.8672 26.99 Detroit 0.8321 19.88 Grand Rapids 0.8321 28.06 Hartford 0.8666 21.1 Houston 0.8321 30.78 Indianapolis 0.8712 19.78 Jacksonville 0.8238 19.66 Kansas City 0.85 28.06 Las Vegas 0.826 28.06 Los Angeles 0.878 38.27 Louisville 0.831 17.44 Memphis 0.8193 28.06 Miami 0.8137 36.94 Milwaukee 0.8459 19.85 Minneapolis 0.8321 25.87 Montgomery County 0.8321 29.02 Nashville 0.5869 23.59 New Orleans 0.8821 28.06 New York 0.8388 73.01 Newark 0.8388 24.06 Northern Virginia 0.8321 33.08 Oklahoma City 0.8321 28.06 Orlando 0.8689 21.01 Philadelphia 0.8321 26.25 Phoenix 0.8389 24.48 Pittsburgh 0.8345 23.05 Portland 0.8321 27.56 Providence 0.8464 28.06 Raleigh 0.8321 22.27 Richmond 0.857 19.17 Rochester 0.8613 28.06 Sacramento 0.9079 23.76 Saint Louis 0.8496 19.05 Salt Lake City 0.8321 22.99 San Antonio 0.4351 23.68 San Diego 0.8321 31.44 San Francisco 0.8404 73.65 San Jose 0.8321 58.08 Seattle 0.8565 34.9 Tampa 0.8566 23.01 Toronto 0.8321 19.88 Tucson 0.8871 28.06 Virginia Beach 0.8652 18.62 Washington DC 0.8321 37.25
|Printer friendly Cite/link Email Feedback|
|Comment:||Business strategy and firm location decisions: testing Traditional and modern methods.|
|Author:||Anderson, Patrick L.|
|Date:||Jan 1, 2019|
|Previous Article:||Heigh Ho, Heigh Ho: flexible labor contracts with real option characteristics.|
|Next Article:||Are yield-curve/monetary cycles' approaches enough to predict recessions?|