9 – 1 1 8 -0 6 8

R E V : J U N E 2 5 , 2 0 2 1

Professor Charles C.Y. Wang and Research Associate Kyle Thomas prepared this case. It was reviewed and approved before publication by a company designate. Funding for the development of this case was provided by Harvard Business School and not by the company. HBS cases are developed solely as the basis for class discussion. Cases are not intended to serve as endorsements, sources of primary data, or illustrations of effective or ineffective management. Copyright © 2018, 2021 President and Fellows of Harvard College. To order copies or request permission to reproduce materials, call 1-800-545- 7685, write Harvard Business School Publishing, Boston, MA 02163, or go to www.hbsp.harvard.edu. This publication may not be digitized, photocopied, or otherwise reproduced, posted, or transmitted, without the permission of Harvard Business School.

C H A R L E S C . Y . W A N G

K Y L E T H O M A S

New Constructs: Disrupting Fundamental Analysis with Robo-Analysts

CEO David Trainer and COO Lee Moneta-Koehler of New Constructs had just met with a potential client. Their pitch was simple: New Constructs “leverages the only parsing technology capable of reviewing every detail of every 10-K and 10-Q” to deliver quality fundamental analysis at scale. After the presentation, the client responded, “You know, you might be right. Your data probably is better. But, as long as everybody’s using the same bad data, I'm OK with that.” This was a familiar response to Trainer and Moneta-Koehler: they were frustrated by investors who did not see the value of New Constructs’ data or technology. They were concerned about the role of quality fundamental analysis in a market increasingly focused on more technical and other short-term trading strategies.

New Constructs Trainer began working on Wall Street at Credit Suisse First Boston (CSFB) as a stock analyst in 1996,

where he honed his skills in financial modeling and fundamental analysis. At CSFB, he spearheaded an effort to develop a consistent framework for measuring, comparing, and analyzing the economic earnings and profitability across all firms and industries globally. After reading through thousands of corporate filings, he realized that “the complexities of what’s going on in modern day business are so much greater than what the current accounting standards can capture in the income statement and balance sheet.” To construct a more accurate economic picture of the firm and to facilitate more meaningful comparisons of performance, his financial models incorporated quantitative details hidden in footnotes and the management’s discussion and analysis (MD&A) section, such as operating lease obligations or components of income or expenses that are transitory in nature.

Although these adjustments were often meaningful to his overall assessment of a firm’s operating performance and valuation, integrating these details into financial models was not the norm among many sell-side analysts. Trainer believed this was due to a few reasons. For one, the increasing length and complexity of corporate filings and the differences in the application of accounting rules across firms for similar transactions made the execution of such detailed financial models impractical. Even for Trainer, this mode of analysis was difficult to scale.

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

118-068 New Constructs: Disrupting Fundamental Analysis with Robo-Analysts

2

Another factor was the conflict of interest in the industry. “Sell-side analysts were not necessarily incentivized to provide the most accurate research for the stocks they cover,” said Trainer, “Research departments are cost centers in investment banks. They had incentives to promote certain stocks, which helps to generate investment banking revenues.” During the IPO tech bubble in the late-1990s, Trainer witnessed first-hand huge sums of investment capital being allocated to companies that his (as well as his colleagues’) fundamental analyses showed had no substantive economic value. Nevertheless, these firms posted massive post-IPO returns and “people appeared to care less and less about 10-K’s and 10- Q’s.” Even his colleagues began buying these same “worthless” stocks. “’How could you buy these stocks when you know there’s no real viable business here?’” Trainer recalled asking one coworker. “He told me, ‘it doesn’t matter David, it’s called the greater fool theory, as long as someone else out there is willing to pay a higher price down the road, I’m buying it now.’”

These experiences deeply troubled Trainer, who believed that sound fundamental analysis and information efficiency in the capital markets is not just a matter of investment profits. “Improvements in the standards of living around the world,” he said, “are driven by the market’s ability to efficiently allocate capital.” He believed that a deteriorating quality in the analyses of firm performance exacerbated information asymmetries between market participants and threatened the trust that formed the bedrock of capital markets.

After the internet bubble burst in the early 2000s, the ensuing regulatory response brought to light the conflicts of interests between the research and investment departments of investment banks, particularly where “financial analysts [tailored] their research reports and stock ratings to win investment-banking business.”1 Ten of the largest U.S. investment banks—including Credit Suisse— agreed with the regulators in 2003 to a $10+ billion global settlement regarding biased equity research, including, most notably, a mandate to establish a barrier between the research and investment banking units. However, Trainer was skeptical.

In 2002, Trainer founded New Constructs with the goal of “providing transparent valuation models based on more systematic extraction of data from corporate filings that filters out the noise of accounting distortions.” His vision was to disrupt the financial data and research industry by culling and organizing all relevant information hidden in corporate filings that would allow financial analysts to better understand the true economic picture of any firm and easily scale fundamental analyses that required detailed accounting adjustments. Clean and comprehensive accounting data would be collected by combining his knowledge of the complexities of financial statements with state-of-the-art technology for automatically processing and parsing filings. The end goal was to usher in a new era “where highly-scalable and high quality fundamental analysis could finally meet and bring more integrity to capital markets.”

Economic Earnings and Return on Invested Capital

As a fundamental analyst, Trainer believed that to truly understand a firm’s operating performance and intrinsic value, one must understand the firm’s sustainable operating profitability as well as the amount of net operating assets (or “invested capital”) deployed to generate those operating profits. The intrinsic value of a firm boils down to the discounted future economic earnings that can be expected from the firm. Conceptually, economic earnings represent the operating earnings generated by the firm in excess of the economic cost of (or the required rate of return on) the total operating capital deployed. (See Exhibit 1 for the relation between intrinsic value of a firm and economic earnings.)

However, performance and valuation metrics based on these concepts never achieved mainstream popularity. “Many firms built consulting businesses around these ideas, like Stern Stewart’s Economic Value Added or HOLT’s Cash Flow Return on Investment. These types of metrics were all the craze in

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

New Constructs: Disrupting Fundamental Analysis with Robo-Analysts 118-068

3

the 90s,” Trainer recalled. “But the problem was that they were black boxes that could not scale. Nobody really understood what went into them and everyone had a different formula. As a result, these important concepts never caught on in the mainstream.”

New Constructs sought to solve this problem in two ways. First, it would collect comprehensive quantitative data from annual and quarterly reports, including information that is disclosed only in the MD&A or the footnotes to the financial statements. Second, it would use these data to apply a set of transparent and clearly-defined adjustments to firms’ net income and total assets in order to measure their sustainable operating profits and invested capital.

To capture a firm’s total invested capital (“IC”) that is used to generate operating profits, New Constructs measures the “sum of all cash that has been invested in a company’s operations over its life with no regard to financing form or accounting name.” To do so, New Constructs first adjusts a firm’s total assets for any accounting distortions that can arise from off-balance-sheet accounting treatments (e.g., capitalization of operating leases). (See Exhibit 2 for the major income statement and balance sheet adjustments made by New Constructs.) In addition, New Constructs reverses write-downs of operating assets in place (e.g., reversal of operating asset write offs).

For debt investors, which GAAP was primarily designed for, write-downs are analytically helpful. They provide a more accurate assessment of the liquidation value of a company’s assets. For equity investors, on the other hand, write-downs are not helpful because they distort the return on invested capital (ROIC) of a company. Write-downs allow management to erase equity from the balance sheet, which inflates any return on capital metric. So, we add back asset write-downs (after tax) to our measure of invested capital. Keeping those write-downs in invested capital holds companies accountable for all the capital invested in their business over their operating lives.

Having normalized the balance sheet, assets and liabilities are split into operating versus non- operating accounts. Doing so often requires data from footnote disclosures, which disaggregate operating and non-operating items that have been combined in balance sheet accounts (e.g., long-term asset accounts that can include non-operating assets like prepaid pension assets and other equity investments). A firm’s invested capital is computed by subtracting operating liabilities (e.g., accounts payable, accrued salaries, deferred revenue, and income taxes payable), that is total liabilities minus non-operating liabilities (e.g., debt or debt equivalent liabilities, including capital leases or capitalized operating lease liabilities, unfunded pension liabilities, unfunded postretirement medical costs, and restructuring reserves), from its operating assets (e.g., accounts receivable, fixed assets, and other long- term operating assets), that is total assets minus non-operating assets (e.g., non-operating cash balances and marketable securities). In principle, IC represents the net operating assets that support the operations of the business.

To measure a firm’s sustainable operating profits, revenues or expenses that are identified as non- operating (e.g., interest income and expenses) or are non-recurring in nature (e.g., one-time gains or losses and expenses associated with write-downs of invested capital) are removed or added back to net income (net of tax effects). This adjusted income measure, called the net operating profit after taxes (NOPAT), represents the normal, operating, and unlevered (i.e., available to both debt and equity investors of the firm) after-tax operating profit generated by the business.

Trainer believed that, ultimately, the key to a firm’s success boils down to one metric: “return on invested capital” (ROIC). ROIC is defined as the ratio of NOPAT to IC and represents how much sustainable operating profit a company generates for every dollar invested in the company’s operations. In his eyes, ROIC “drives economic earnings” and is “by far the most important driver of

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

118-068 New Constructs: Disrupting Fundamental Analysis with Robo-Analysts

4

firm valuation.”2 (See Exhibit 3 on the relation between the enterprise value to IC multiple and ROIC.) From a market perspective, the intuitive appeal of ROIC is based on a simple premise: “If the purpose of capital markets is to promote the most efficient use of capital, it makes sense that the market would reward companies that earn the most profit per dollar of capital invested with the highest valuations.”

Adjusting Income Statement and Balance Sheet: Two Examples

NVIDIA and Symantec are two S&P 500 technology firms with similar market capitalizations and market valuations as of 2015. Founded in 1993, NVIDIA is a manufacturer of graphic processing units used in computers. (See Exhibit 4 for NVIDIA’s financial statements.) In 2015, it had a market capitalization of $11 billion and a price-to-book multiple of 2.55x. Symantec, founded in 1983, is a software publisher most notable for its Norton antivirus product. (See Exhibit 6 for Symantec’s financial statements.) In 2015, it had a market capitalization of $16 billion and price-to-book multiple of 2.71x.

New Constructs makes a number of adjustments to each company’s balance sheet and the income statement before calculating their ROICs. In calculating the NOPAT of NVIDIA in 2015, New Constructs identified three revenue or expense items hidden in footnote disclosures it deemed to be transitory in nature: sales of inventory that was previously written off, net gain on sale and disposal of long-lived assets and investments, and net acquisition-related costs. To compute IC for NVIDIA in 2015, New Constructs capitalizes the firm’s minimum operating lease payments, as disclosed in the footnotes, as operating assets and liabilities. Across all firms, this was one of the biggest adjustments made in terms of its impact on the balance sheet. The implied interest expense from capitalizing operating leases is not considered an operating expense in the computation of NOPAT. In addition, New Constructs generally considered 5% of the company’s operating revenues as cash balance necessary for operations, as was the case for NVIDIA. However, this percentage can differ across firms or for the same firm but across years based on the firm’s proprietary methodology for estimating operating cash. Cash in excess of this amount is classified as non-operating assets. Finally, accumulated write-downs since 1998 are capitalized as operating assets. (See Exhibit 5.)

Symantec has relatively few but similar adjustments. For example, in calculating the NOPAT of Symantec in 2014, New Constructs considers acquisition-related expenses as non-recurring. In computing Symantec’s IC, New Constructs capitalizes the firm’s minimum operating lease payments and adds back its accumulated goodwill amortization as well as accumulated operating asset write-downs. In estimating Symantec’s net financial assets, New Constructs considers 5% of the company’s operating revenues in 2014 as operating cash, but this proportion is reduced to 2% in 2015. (See Exhibit 7.)

Unleashing the Robo-Analyst The central challenge in implementing New Constructs’ framework for NOPAT and invested

capital is that achieving consistent measurement at scale for every firm was a “nearly impossible task” for a traditional analyst covering more than a handful of firms. Information underlying the identification of operating or non-operating accounts were reported in varying locations across 200+ page filings, often hidden deep within the footnotes and disclosed using different terminology by different firms. For a human analyst to make the appropriate balance sheet or income statement adjustments to facilitate apples-to-apples comparisons across firms, she had to attentively read hundreds of pages of filings for multiple firms and multiple years and accurately collect all the data. The result was a tradeoff: “the more companies a person covered, the lower the quality of research.”

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

New Constructs: Disrupting Fundamental Analysis with Robo-Analysts 118-068

5

As an analyst himself, Trainer faced this dilemma—particularly when filings were only available in paper form. This process became less tedious with the advent of digital filings in the late 1990s. “That was awesome,” Trainer recalled, “but then I thought this is terrible because I’ve got 100 things I need to control+f and it’s still going to take forever.” The new digital filings, however, led him to conceive of a robo-analyst: a human-analyst-assisted computer program that could perform all the necessary data collection from corporate filings. The ultimate goal would be a machine that could intelligently assimilate the patterns of data processing from the human analysts for any given firm’s filings, and eventually process all the data in every filing without human intervention.

Perfecting this data collection process involved three steps. The initial step was an internal application that allowed human analysts to quickly and easily locate, tag, and mark up all relevant quantitative data points from the footnotes and the MD&A. This tool facilitated the collection of “information about every data point, including original text, data value, units, and specific location in the filing,” that were necessary for the computation of NOPAT and IC. The second step was the development of a taxonomy that organized each collected quantitative number into the appropriate category (e.g., operating or financing revenues and expenses) for the automated computation of NOPAT and IC. The last component was automation through a machine-learning algorithm that parsed the entire corporate filing (i.e., an algorithm that could read, locate, tag, extract, and categorize every data point). The idea was “if a human expert parsed certain data points into the same bucket enough times, then the machine could take over.”

Although the initial manual mark up of filings was tedious, the database of correctly- and fully- parsed filings allowed New Constructs to gradually automate parsing through an intelligent algorithm that could “confidently collect and process as much information as it can, and if it's not 100% sure, push it up to the human analyst and then learn from her actions.” For example, if a firm reported a “fracking waste disposal” charge that the machine was unsure about because it had never seen it before, the algorithm would push it to the analyst for further review. Once the analyst provided the appropriate metadata on the data point and classified it in the taxonomy, the machine would then be able to automatically parse it in the future. (See Exhibit 8.)

By 2015, New Constructs was able to automatically parse about 35% of the data in 10-Ks and 10-Qs without human intervention and had amassed a dataset of over 120 thousand fully-parsed filings. In 2016, New Constructs made a major advancement when it introduced natural language processing (NLP) to its data collection technology. NLP allowed New Constructs to leverage all of is previously collected knowledge and construct a smarter data collection algorithm that could process more of the unstructured data (i.e., data not listed in a table format) found in the filing text without human intervention. Within six months of its introduction, virtually 100% of all the data in 10-Qs was automatically processed and most 10-Ks could be completely processed within minutes. On average, 10-Ks (roughly around 200 pages on average) require the parsing of 70 tables and 256 data points embedded in the text.

Over the years, New Constructs has implemented various organizational structures to promote continual improvements in the technology and the data’s quality. Trainer and Moneta-Koehler fostered a company culture that was obsessive about attention to detail and perfection. It began with identifying and training suitable analysts. “We try to screen hard on the front end for people who want to be here, who believe in what we are doing, and who are dedicated,” said Moneta-Koehler. “All of our incoming analysts go through a rigorous training program in accounting and finance, and if they pass the final exam after six months of training, they’re allowed to touch the live data under close supervision.” A dedicated team reviews all analysts’ work. If an analyst had persistent issues with accurately processing a particular data point, their past work would be flagged for further review and the analyst

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

118-068 New Constructs: Disrupting Fundamental Analysis with Robo-Analysts

6

would be coached on improving their work. The quality of an analyst’s work (i.e., error rates) also entered into her bonuses. Commenting on their approach to analysts, Moneta-Koehler said, “I would rather have an analyst validate nothing but total assets and do it correctly 100% of the time than an analyst that knows how to do everything 50% accurately.”

The workforce was internally organized to promote constant refinement in the data-collection technology, especially its data classification system. The engineers and analysts worked in close proximity, and were encouraged to be in constant communication with each other. Senior Software Engineer Chesley Lovejoy explained,

The key to perfecting our data-collection process is the cooperation between the engineers and the analysts. Analysts know enough about the tech and engineers know enough about the finance so that we can talk about this in a shared language. For example, when accounting rules change, an analyst is going to come talk to us about what we should do on the technology end and we will make small changes together and get immediate feedback. This collaboration helps us be flexible and adaptive.

Finally, New Constructs designed automated quality control mechanisms to ensure the accuracy of the robo-analyst processed data. These included computer programs that performed basic sanity checks (e.g., checking sums of line items against totals and that the numbers in a document are internally consistent), checked the data for unusual or extreme values (relative to both a firm’s time series and the universe of collected data), and validated the data against external data sources. Lovejoy said, “We realize we can’t be perfect, but if you aren’t striving for perfection then what is the point.” The importance of attention to detail pervaded the office space. Moneta-Koehler pointed out his two favorite office posters: they read “when you are 90% finished, you are half way there” and “slow down to speed up.”

Business Challenges From the beginning, New Constructs explored various ways to monetize its unique data and

research, “This data is the trunk of a tree that you can build many branches off of: investment ratings, wealth advising, auditing, corporate governance, academic research.” Certain investors, such as hedge funds, purchased direct access to the raw data. For other investors, such as institutional portfolio managers, New Constructs offered varying ways of analyzing or accessing investment insights from the data. Through its web-based research platform, New Constructs provided valuation tools that leveraged its unique fundamental data to compute estimates of a stock’s intrinsic value. The platform’s tools were built to be flexible and transparent. Users were able to scrutinize all corporate filings from which every data point used in the valuation model came. Users could also manipulate the assumptions in the models. “Almost all of the adjustments can be over-ridden,” said Trainer, “I don’t want to be in the business of selling analytical religion. I want to give you a better dataset with which you can make a more informed decision.”

New Constructs also generated a variety of investment research insights that were made available through the platform and other distribution channels. For example, it provided an overall rating for each covered stock, a 5-point scale ranging from “Very Attractive” to “Very Unattractive,” based on quantitative assessments of the firm’s business strength and the attractiveness of the stock’s current market valuation. (See Exhibit 9.) Mutual funds, ETFs, and REITs were also rated in a similar fashion. These ratings were embedded in the roughly 10,000 robo-analyst reports (approximately 3,000 companies and 7,000 mutual funds and ETFs) that were automatically generated and updated daily, which detailed New Constructs’ main accounting adjustments and their impact on the company’s

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

New Constructs: Disrupting Fundamental Analysis with Robo-Analysts 118-068

7

income statement and balance sheet. Clients could choose varying levels of access to New Constructs’ services, with fees ranging from $49 per month to $3,600 per month.

In 2017, 45% of New Constructs’ revenue came from hedge funds and institutional investors, 30% was from wealth advisors, and 25% was from retail investors. 70% of their customers, however, were from retail, 20% from advisors, and 10% from institutions. In broadening its revenue stream, New Constructs was working on adding several new products and customer pools. One obvious service was forensic accounting and auditing: New Constructs’ technology could help to detect anomalies in financial reporting on a large scale. Another service was providing ROIC measurements to compensation consultants interested in implementing ROIC-based metrics in incentive contracts. One Ernst and Young consultant commented,

It is difficult to compute a standardized measure of ROIC to measure value creation apples to apples across firms and industries, because firms publish numbers in different ways in financial statements. With New Constructs, we can quickly compute comparable ROIC for a firm’s industry peers, enabling more insightful analysis into the strategies firms use to create economic value. To-date we have not seen a platform with the capability that New Constructs provides.3

New Constructs faced two primary challenges: alternative providers of parsed financial data and insufficient demand for clean fundamental data. Trainer recognized that New Constructs’ competitive edge lay in producing parsed data that competitors could not match in breadth and quality. “My value add is not that I have a better model. It’s that I have better inputs, and I have them with scale. I’m bringing a bulldozer to what everyone else is trying to do with a shovel. And that’s why the technology is so important.” Since its inception, New Constructs had sought to disrupt the much larger, established financial data collection efforts at firms like Thomson Reuters and FactSet. While these firms typically relied on traditional data collection methods (e.g., hiring thousands of data-entry people who focused mostly on the data found in tables of 10-Ks and 10-Qs), they were beginning to aggressively apply more advanced algorithms like NLP to extract the same kinds of unstructured data that New Constructs had specialized in. However, Moneta-Koehler believed that New Constructs had a crucial advantage over its competitors in the use of machine learning algorithms to parse corporate disclosures: “our advantage is the massive, clean, and battle-tested dataset of fully-parsed financials that has been used by many investors for over a decade.” This unique database, or training data set, allows the machines to efficiently learn how to correctly parse a new filing and to “automatically get fundamental data from the K’s and Q’s that’s almost 100% accurate at scale.”

However, convincing others of the value of New Constructs’ data proved to be a challenge. Early on, potential customers frequently questioned the incremental value of the data, asking, “If our data is 70% correct, how much more value is there in getting the remaining 30%?” Others pushed harder, “Your data may be better, but as long as everyone is using the same bad data as I am, then I’m sticking with the bad data.” In response, Trainer would highlight the growing regulatory emphasis on due diligence as well as the duty of care when acting as a fiduciary: “If you know the data could be wrong, how can you in good faith use it as the basis for your investment?”

Another frequent question to Trainer is, “If New Constructs’ data and system are so great, then why don’t you manage money with it?” In 2008, Trainer launched a hedge fund. Despite some initial success, it had difficulty raising additional funds due to the financial crisis. After closing the fund and refocusing on research in 2012, New Constructs continued to face questions regarding the relevance and usefulness of their data. However, interest in ROIC as a performance metric has risen in recent years. In 2016, Goldman Sachs reported that ROIC was near the top of the list of “quantitative metrics

This document is authorized for use only by Zac Higgs in ACC-345-H7078 Fin Stmt Analysis/Bus Value 22EW5 at Southern New Hampshire University, 2022.

118-068 New Constructs: Disrupting Fundamental Analysis with Robo-Analysts

8

investors care about” and stated that it was “the best way to distill what activists view as the most critical skill of management: how they allocate capital.”4

Critics also pointed out some notable misses in New Constructs’ stock recommendations. Amazon, for example, never received a New Constructs’ rating above “Ne