Topic tones of analyst reports and stock returns: A deep learning approach

• Published date: 01 December 2023
• Author: Hitoshi Iwasaki,Ying Chen,Jun Tu
• Date: 01 December 2023
• DOI: http://doi.org/10.1111/irfi.12425

ORIGINAL ARTICLE

Topic tones of analyst reports and stock returns:

A deep learning approach

Hitoshi Iwasaki

1

| Ying Chen

2,3

| Jun Tu

4

1

Department of Statistics & Data Science,

National University of Singapore,

Singapore, Singapore

2

Department of Mathematics, National

University of Singapore, Singapore

3

Risk Management Institute, National

University of Singapore, Singapore

4

Lee Kong Chian School of Business,

Singapore Management University, Singapore

Correspondence

Hitoshi Iwasaki, Department of Statistics &

Data Science, National University of

Singapore, Singapore.

Email: iwasaki@u.nus.edu

Funding information

National University of Singapore,

Grant/Award Number: A-8000014-00-00;

Singapore Management University,

Grant/Award Number: G17C20428

Abstract

We present a novel approach that analyzes topics and tones

of analyst reports using a deep neural network in a super-

vised learning approach. By letting trainedclassifiers evaluate

topics and tones of the reports, we find thatincorporation of

topic tones significantly enhances the accuracy of predicting

cumulative abnormal returns, increasing adjusted R2from

6.1% without considering textual information to 17.9% with

detailed topic tones. This improvement is primarily driven

by the inclusion of opinion and corporate fact type of

topics. Our findings highlight importance of topic assess-

ment to make the most use of analyst reports for informed

investment decisions.

KEYWORDS

DNN approach, information content, textual analysis, topic tones

JEL CLASSIFICATION

G11, G12, C89

1|INTRODUCTION

Financial analysts are recognized as vital intermediaries in the capital market, and previous research has

mainly focused on the analyst's information role based on quantitative forecasts such as stock recommendations,

target prices, and earnings forecasts in analyst reports (e.g., Abdel-Khalik & Ajinkya, 1982; Alon & Reuven, 2003;

Hitoshi Iwasaki is an employee of ACG Management Pte, Ltd.

Received: 10 May 2022 Revised: 20 June 2023 Accepted: 5 July 2023

DOI: 10.1111/irfi.12425

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which

permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no

modifications or adaptations are made.

© 2023 The Authors. International Review of Finance published by John Wiley & Sons Australia, Ltd on behalf of International Review

of Finance Ltd.

International Review of Finance. 2023;23:831–858. wileyonlinelibrary.com/journal/irfi 831

Barber et al., 2010; Bradley et al., 2014; Kothari et al., 2016; Womack, 1996). However, besides the quantitative

information, analyst reports also contain a wealth of qualitative analysis on various topics, such as accounting perfor-

mance, corporate governance, competitive landscape, risk, and macroeconomic conditions. Previous studies have

documented that the aggregate textual tone is a crucial contributor for stock returns drift; for example, see the stud-

ies of corporate disclosures (Li, 2010), analyst reports (Huang et al., 2014), news articles (Tetlock, 2007; Zhang

et al., 2016), conference calls (Matsumoto et al., 2011) and Twitter posts (Sul et al., 2017). Nonetheless, the complex-

ity of the text in analyst reports, which encompasses different topics, may obscure the important information from

the nonessential ones. Moreover, the aggregated tone of the text could be skewed by less insignificant topics, poten-

tially resulting in misinterpretations of the report.

Our study involvesbreaking down the potentially intricatetext of an analyst report into multiple sections that cor-

respond to various topics. These topics follow a hierarchical structure within the text. Firstly, analysts commonly pro-

vide supportive arguments and justifications for their numerical forecasts, such as stock recommendation ratings,

target prices,and earnings forecasts. Thisportion of the text contains additionalinformation related to analysts' quanti-

tative forecastsand is defined as the Justificationtopic. The remaining text contains broader perspectivesderived from

the analyst's qualitative insights, including their views onobjective news and information about corporate events. We

define this as the Qualitative topic and split it into the text for the Opinion topic and the text for the Corporate Fact

topic. We further split these three topics into 10 more detailed topics, namely three Justification type topics: recom-

mendation rating, target price,earnings forecast; three Opiniontype topics: conclusion, generalargument, discussion on

risk; and, four Corporate Fact type topics: profit items, non-profit items, officialguidance, and corporate action.

Our study analyzes the topic tone of a unique dataset consisting of 113,043 Japanese analyst reports. This

dataset pertains to the Japanese stock market, which is the third largest stock market in the world based on market

capitalization and traded stock values, following the US and Chinese market.

1

As of the end of 2022, there are 3869

listed companies in Japan, which is comparable to the 2243 firms listed on the New York Stock Exchange,

2

the 3300

firms on NASDAQ,

3

the 2137 firms on the Shanghai Stock Exchange,

4

and the 2696 firms on the Shenzhen Stock

Exchange.

5

We choose to focus on analyst reports written in Japanese, as we believe that little is known about

Japanese financial text and there is a need for further textual study. Researchers have made efforts to apply the

results and findings of financial textual studies in the US market to other markets. For instance, the most commonly

used sentiment word lexicon, Loughran and McDonald (LM) lexicon (Loughran & McDonald, 2011), has been trans-

lated and adapted to local markets in foreign languages such as German (Bannier et al., 2019) and Chinese (Du

et al., 2021). In a comprehensive survey of textual analysis for the Chinese financial market, Huang et al. (2019) out-

line recent advancements in textual studies in China and highlights differences from existing works in the US market.

Additionally, researchers have explored similarities and differences between financial texts in the US and those in

other countries, investigating topics of interest in international contexts, such as the impact of IFRS adoption on

annual report disclosure quality (Lang & Stice-Lawrence, 2015).

In contrast, there has been relatively little research on textual studies in the Japanese market, with no systematic

a review of such studies having been undertaken to date. This is particularly true of research that employs analyst

reports (Miwa, 2021; Miwa, 2022). The scarcity of research in this area can be attributed to several distinctive fea-

tures of the Japanese market and language. Firstly, the Japanese market is highly localized, with a majority of

investment-related information being intermediated by local analysts. In fact, Japan has the highest percentage

of local analysts in the world, at 88%, compared to China's 9% (Bae et al., 2008). Secondly, the Japanese language is

regarded as the most linguistically distant from English (Chiswick & Miller, 2005; Hart-Gonzalez &

Lindemann, 1993).

6

As a result, the cost of learning Japanese and the difficulty of translation are high for English

speakers. This may explain why only a quarter of analyst reports written in Japanese are accompanied by simulta-

neous English translations (Miwa, 2021). These factors make it challenging for non-Japanese researchers to conduct

precise textual studies in the Japanese market.

While the difficulties of text analysis in the Japanese market are undeniable, it would be erroneous to view these

challenges as diminishing the market's relevance. Given its considerable size and the abundance of historical data,

832 IWASAKI ET AL.