NBERWORKINGPAPERSERIES FROMMANVS.MACHINETOMAN+MACHINE: THEARTANDAIOFSTOCKANALYSES SeanCao WeiJiang JunboL.Wang BaozhongYang WorkingPaper28800 nber/papers/w28800 NATIONALBUREAUOFECONOMICRESEARCH 1050MassachusettsAvenue Cambridge,MA02138 May2021 The authors have benefited from discussions with Svetlana Bryzgalova, Will Cong, Jillian Grennan, GerryHoberg,MarkusPelger,SiewHongTeoh,andChristinaZhu,andcommentsand suggestions from seminar/conference participants at CKGSB and Stanford Engineering the AI BigDatainFinance ResearchForum(ABFR)webinar.Theviewsexpressedhereinarethoseof the authors and do not necessarily reflect the views of the National Bureau of Economic Research. NBERworkingpapersarecirculatedfordiscussionandcommentpurposes.Theyhavenotbeen peerreviewedorbeensubjecttothereviewbytheNBERBoardofDirectorsthataccompanies official NBERpublications. 2021bySeanCao,WeiJiang,JunboL.Wang,andBaozhongYang.Allrightsreserved.Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission providedthatfull credit,includingnotice,isgiventothesourceFromManvs.MachinetoMan+Machine:TheArtandAIofStockAnalyses SeanCao,WeiJiang,JunboL.Wang,andBaozhongYang NBERWorkingPaperNo.28800 May2021 JELNo.G11,G12,G14,G31,M41 ABSTRACT An AI analyst we build to digest corporate financial information, qualitative disclosure and macroeconomic indicatorsis abletobeatthemajorityof humananalystsinstockpriceforecasts and generate excess returns compared to following human analyst. In the contest of “man vs machine,” the relative advantage of the AI Analyst is stronger when the firm is complex, and when information is highdimensional, transparent and voluminous. Human analysts remain competitive when critical information requires institutional knowledge (such as the nature of intangibleassets).TheedgeoftheAIoverhumananalysts declinesovertimewhenanalystsgain accesstoalternativedataandtoinhouseAIresourcesbining AIscomputationalpowerand the human art of understanding soft information produces the highest potential in generatingaccurate forecasts. Our paper portraits a future of “machine plus human” (instead of humandisplacement)inhighskillprofessions. SeanCao J. MackRobinsonCollegeofBusiness GeorgiaStateUniversity 35BroadStreet,Suite1243 Atlanta,GA303023992 scaogsu.edu WeiJiang GraduateSchoolofBusiness ColumbiaUniversity 3022Broadway,UrisHall803 NewYork,NY10027 andNBER wj2006columbia.edu JunboL.Wang SchoolofBusiness LouisianaStateUniversity BatonRouge,LA70803 junbowanglsu.edu BaozhongYang GeorgiaStateUniversity J. MackRobinsonCollegeofBusiness 35BroadStreet,Suite1243 Atlanta,GA30303 bzyanggsu.edu1. Introduction Since its inception and as it rises, articial intelligence (AI) constantly makes human beings to rethink their own roles. While AI is meant to be intelligence augmentation for humans, concerns abound that AI could replace human tasks and increasingly skilled ones, and thus displace jobs by those currently performed by the better-paid and better-educated workers (Muro, Maxim, and Whiton, 2019). Such a concern and the associated debates have motivated a quickly growing literature. Recent work by Webb (2020), Acemoglu, Autor, Hazell, and Restrepo (2020), Babina, Fedyk, He, and Hodson (2020), and Jiang, Tang, Xiao, and Yao (2021) have all conducted large-sample analyses on the extent of job exposure and vulnerability to AI-related technology as well as the consequences on employment and productivity. The existing literature has been mostly focusing on characterizing the type of jobs that are vulnerable to disruption by, as well as those that could be created due to, AI evolution. In other words, the sentiment of the existent studies mostly involves a theme of “man- versus-machine,” i.e., to characterize the contest between human and AI, to explore ways human adapts, and to predict the resulting job redeployment. In such settings, human beings are often rendered passive or reactive dealing with disruptions and looking for new opportunities dened by the AI landscape. There has been relatively little research devoted to prescribing how skilled human workers could tap into a higher potential with enhancement from AI technology, presumably the primary goal for human beings to design and develop AI in the rst place. This study aims to connect the contest of “man-versus-machine” (“man v. machine” hereafter) to a potential equilibrium of “man-plus-machine” (“man + machine” hereafter). Our study could be motivated by the experience of chess grand master Garry Kasparov. The story that IBMs Deep Blue beat the then reigning grand master Kasparov in 1997 was well-known. Multiple contests repeated in a similar setting afterwards killed any remaining suspense for the outcome of man v. machine in chess playing. What is far less known is that 1humans, despite having lost interest in man-versus-machine chess contests, have not lost interest in either the game or the machine. In fact, the encounter with the Deep Blue was a catalyst for people like Kasparov to pioneer the concept of man + machine matches, in which a chess player equipped with AI assistance (a “centaur” player) competes against AI. Up to today the centaur keeps an upper hand against machines; and even more encouragingly, there have been more and better human chess players with the advent of aordable AI- engineered chess programs. 1 If AI can help more humans become better chess players, it stands to reason that it can help more of us become better at many skilled jobs, from pilots, medical doctors, to investment advisors. In this study, we zoom into the profession of stock analysis, whose data availability allows us to calibrate both man vs. machine and man + machine. Stock analysts are among the most important information intermediaries in the market place (e.g., Brav and Lehavy, 2003; Jegadeesh, Kim, Krische, and Lee, 2005; Crane and Crotty, 2020). Their job, which require both institutional knowledge and data analytics, has not been spared by AI as more and more investors start to heed to recommendations about stock picking and portfolio formation made by AI-powered tools. 2 To trace out the path from “man v. machine” to “man + machine,” we decided to build our own AI model for year-end stock predictions so that we have a consistent and time-adapted benchmark for AI performance which we understand and are able to explain. 3 Target prices and earnings are the two primary subjects of analyst forecasts, we choose the former as the latter are subject to managerial discretion, which a large body of accounting literature on earnings management manifests. Our “AI analyst” is built on training a com- 1 Source of information: The Inevitable, by Keven Kelly, Penguin Publishing Group, 2016. See also “Defeated Chess Champ Garry Kasparov Has Made Peace With AI,” Wired, February 2020. 2 Sources: “What Machine Learning Will Mean for Asset Managers,” Robert C. Pozen and Jonathan Ruane, Havard Business Review, December 3, 2019. “How Startup Investors Can Utilize AI To Make Smarter Investments,” Jia Wertz, Forbes, January 18, 2019. 3 We focus on year-end prices because these are a typical focal point for investors and corporations for tax and reporting reasons. For example, “Credit Suisse Raised Its SP 500 Target Earnings Are Too Good to Ignore,” Jacob Sonenshine, Barrons, April 30, 2021. 2bination of current machine-learning (ML) tool kits 4 using timely publicly available data and information. More specically, we collect rm-level, industry-level, and macro-economic variables, as well as textual information from rms disclosure (updated to right before the time of an analyst forecast) as inputs or predictors, but deliberately exclude information from analyst forecasts (past and current) themselves. We resort to machine learning models, instead of traditional economics models (such as regressions) due to the advantages to the former in managing high dimensional unstructured data, and in their exibility in optimizing and tting unspecied functional forms. More recent development in the area also allows us to mitigate over-tting and to improve out-of-sample performance. We kept training and improving the model until we were condent that our AI analyst is able to beat human analysts as a whole: The AI analyst based on the nal “ensemble” model outperforms 53.7% of the target price predictions made by all IBES analysts during the sample period of 2001-2016. 5 Moreover, a monthly rebalanced long-short portfolio based on the dierences in the opinions of AI and human analysts 6 is able to generate a monthly alpha of 0.84% to 0.92% using the Fama-French-Carhart four-factor model. Though building anAIanalystisnottheultimategoalofthisstudyandthoughwedonotclaimourAIanalyst to be the best of the kind, its performance already suggests that the profession of nancial analysts is subject to technology disruption as our model is a lower bound of the state-of- the-art. To the extent that we have, at our disposal, an AI analyst that beats the average of its human counterparts, we are able to explore the relative advantages of, and potential synergies between, the two sides. First, we examine the circumstances under which human analysts retain their advantage, in that a forecast made by an analyst beats the concurrent AI forecast in terms of lower 4 We start with two versatile quasi-linear ML models, Elastic-Net and Support Vector Regressions, that are adept at tasks with a large number of variables. We then add on three highly nonlinear ML models, Random Forest, Gradient Boosting, and Long Short-Term Memory (LSTM) Neural Networks. Random Forest and Gradient Boosting can both capture complex and hierarchical interactions among the input variables while the LSTM model is designed to model time-series patterns in the data. 5 In comparison, the predictions by an OLS model only outperform 19.3% of analyst forecasts. 6 Such a portfolio would long a stock if the AI forecasts a higher target price than the median analyst forecast over a time horizon and short a stock if otherwise. 3absolute forecast error relative to the ex post realization (i.e., the actual year-end stock price). We nd human analysts perform better for more illiquid, smaller rms, and rm with asset-light business models (i.e., higher intangible assets), consistent with the notion that such rms are subject to higher information asymmetry and require better institutional knowledgeorindustryexperiencetodecipher. Analystsaliatedwithlargebrokeragehouses also stand a higher chance of beating the machine, a combination of their abilities and the research resources available to them. Moreover, analysts are more likely to have the upper- hand when the associated industry is experiencing distress, suggesting that the AI has yet to catch up on relatively infrequent changes such as an industry recession. This is consistent withthelimitationofcurrentmachinelearningandAImodelswhichlackreasoningfunctions and thus cannot learn eectively from infrequent events. 7 As expected, AI enjoys a clear advantage in its capacity to process information, and is more likely to out-smart analysts when the volume of public information is larger. Just like the “centaur” chess player which Kasparov pioneered, the superior performance of an AI analyst does not rule out the value of human inputs. If human and machine have relative advantage in information processing and decision making, then human analysts may still contribute critically to a “centaur” analyst, i.e., an analyst who makes forecasts that combine their own knowledge and the outputs/recommendations from AI models. After we addanalystforecaststotheinformationsetofthemachinelearningmodelsunderlyingourAI analyst, the resulting “man + machine” model outperforms 57.3% of the forecasts made by analysts, and outperforms the AI-only model in all years. Thus, AI analyst does not displace human analysts yet; and in fact an investor or analyst who combines AIs computational power and the human art of understanding soft information can attain the best performance. We are thus interested in knowing when the incremental value of human to a man + machinemodelisthehighest,asmanifestedintherelativeperformanceoftheman+machine model versus the pure AI model. Similar to previous ndings, we nd inputs from analysts 7 Source: “What AI still cant do,” Brian Bergstein, MIT Technology Review, February 19, 2020 4are more valuable when covering rms that are more illiquid and rms with more tangible assets. Moreover, analyst inputs have more incremental value in long-horizon forecasts, and during the time period an industry is experiencing diculty. Importantly, the incremental value of human does not decrease as the volume of information (hence demand for processing capacity) increases, though this constitutes a human disadvantage when alone. Similarly, analystsfromsmallbrokeragehousesmakesimilarlevelofcontributiontotheman+machine modelcomparedtotheircounterpartsfromlargerbanks, suggestingthatAIcouldpotentially help level the disparity in institutional resources. Finally, we resort to an event study to sharpen the inference of the impact of integrating man and machine in stock analyses. In recent years, the infrastructure of “big data” has cre- atedanewclassofinformationaboutcompaniesthatiscollectedandpublishedoutsideofthe rms, and such information provides unique and timely clues into investment opportunities. An important and popular type of alternative data captures “consumer footprints,” often times in the literal sense such as satelliteimages on retail parking lots. Such data, which have to be processed by machine learning models, have been shown to contain incremental infor- mation for stock prices (Zhu, 2019; Katona, Painter, Patatoukas, and Zeng, 2020). We build on data from Katona, Painter, Patatoukas, and Zeng (2020) on the staggered introduction of several important alternative data bases, and conduct a dierence-in-dierences test of analysts performance versus our own AI model before and after the availability of the alter- native data. The underlying premise is that analysts who cover rms that are served by the alternative data could be in the situation of man + machine, as they have the opportunity to use the additional, AI-processed, information. Indeed, we nd that post alternative data, analysts covering aected rms improve their performance relative to the AI-only forecast model we build. Furthermore, such improvement concentrates in the subset of analysts who are aliated with brokerage rms with strong AI capabilities, measured by AI-related hiring 5using the Burning Glass U.S. job posting data 8 and the classication algorithm developed in Babina, Fedyk, He, and Hodson (2020). Overall, results support the hypothesis that analyst capabilities could be augmented by AI, and moreover, analysts work possesses incremental value such that they, with the assistance of AI, can still beat a machine model without human inputs, analo