About relationship between business text patterns and financial performance in corporate data

Korea’s small and medium-sized enterprises’ (SMEs) global competitiveness is relatively weak due to a national economic structure favoring larger corporations. To strengthen the competitiveness of SMEs, policy discourages large companies from entering industries that are appropriate for SMEs. In most industries, large companies and SMEs collaborate and compete at the same time; however there may be industries to which SMEs are more suited. Park et al. studied the selection of industries suitable for SMEs in Korea (Park et al., 2016). However, for SMEs to be able to compete internationally, they need to be competitive with global companies. To strengthen their competitiveness, SMEs that are not rich in human resources and financial status must be innovative. Open innovation can be a way of achieving this. Open innovation benefits from both internal and external knowledge. Research on innovation, or open innovation, has been actively pursued within academia. Witt has studied innovation that is necessary for our future (Witt, 2016). Yun et al. have studied dynamics that lead from open innovation to evolutionary change (Yun et al., 2016). Svirina, et al. conducted a study applying the concept of open innovation to social businesses (Svirina et al., 2016). By utilizing both internal knowledge and external knowledge, open innovation can strengthen SME competitiveness. This study analyzes the annual reports of US listed companies, with the aim of helping Korean SMEs absorb foreign knowledge and open innovation.

Bibliometric analysis studies using patent or paper data in the field of technology intelligence research are being studied for the purpose of discovering new opportunities from the technical planning point of view. In bibliometric research, the areas where many papers and patents are produced or cited are defined as promising technologies or areas. According to Jeong et al., patent analysis can identify new free technology and utilize it as a seed technology for SMEs and as a means of growth engine (Jeong et al., 2014). In another study, Jeong et al. extracted future promising research areas by using bibliographic coupling and clustering technique for the top 1% of cited papers in SCOPUS data (Jeong et al., 2008). These techniques can be used as a way to shorten the effort and time of R & D and technology development in countries with limited resources.

On the other hand, the analysis of the correlation between the business text and the financial analysis, which is the subject of this study, is interested in exploring the characteristics of the business documents of the companies with good financial performance and exploring the companies and the areas with the good financial performance. As a result, the bibliometric technique for papers and patents focuses on the area of people’s interest. On the other hand, this study is interested in finding companies or business area that is expected to perform well in the future in terms of financial or operating performance.

The purpose of this study is to find out whether there is a correlation between business sales patterns and business text patterns. This is based on the research that other researchers in the past have found that estimating future financial performance with a time-series pattern of financial performance is limited and business texts better represent future prospects (Shirata et al. 2011). In this paper, we use text mining techniques to create business text patterns and explore how they relate to corporate financial performance. In other words, if the financial performance of a company is good, we want to find out what patterns appear in the business text.

The study is a field of business intelligence that predicts the future financial performance of a company through business text analysis. In this field, robots that predict future stock prices through various data analysis have recently appeared, but there is not much research in academic aspect. This may be because new areas such as Big Data, as mentioned by Celia, SatikoIshikiriyama et al. (2015), have attracted much more research interest among researchers. In addition, traditional financial information researchers seem to be lacking knowledge about methodology that utilizes text mining techniques to forecast company financial performance. This field is a fusion field where text mining knowledge and financial knowledge are combined. However, as the interest of academia is concentrated elsewhere, it is considered that there is not much research in this field and it is sporadic. Although this study does not comprehensively deal with text mining and sales performance, it is expected to be a good alternative to predict future business performance by analyzing the correlation between business text pattern and business performance.

The annual reports of US listed companies is publicly available in 10-K format. These reports can contain information that illustrate the companies’ global economic conditions. Furthermore, these reports contain not only the financial information of a company but also the main business content, competitiveness and risk factors within each industry. This basic information is useful for grasping global industry trends. Therefore, in addition to traditional financial analysis, text analysis on annual reports has been attempted using text-mining techniques. Lee et al. attempted to analyze a company’s business model by applying text mining to an annual report (Lee et al., 2014). The study analyzed the report’s business descriptions and future strategies. Specifically, business keywords were extracted, and the average frequency and growth rate of each keyword were calculated. The position of the business keywords in the “business model evolution map” based on these two values was shown. Additionally, it is thought that annual reports have richer information in the text than in the numerical values. Consequently, there have been attempts to apply text-mining techniques to annual reports (Kloptchenko et al., 2002). Kloptchenko et al. study the implications for future financial performance that can be drawn from the textual portion of quarterly reports using self-organizing maps and text clustering techniques. Self-organizing maps and text clustering were used to analyze financial quantitative information and qualitative data analysis respectively. In the text-clustering process, word histograms were created from the text documents and made into distribution functions. Subsequently, Euclidean distances between histograms were obtained to measure the similarity between reports. A comparison of the quantitative result of financial information and the qualitative results of text showed inconsistencies. Thus, they argued that the style of writing changes before a dramatic change occurs in financial performance. For example, if the firm’s position is expected to deteriorate in the next quarter, the current quarter’s report will be more pessimistic, even though the financial performance is, as yet, unchanged. Another study aimed to predict the possibility of a firm defaulting, by applying text-mining techniques to annual reports (Shirata et al. 2011). Shirata et al. used textual information from Japanese companies’ annual reports to establish key phrases that could be used to predict corporate default. Prediction of default is one of the most important research topics among accounting researchers, with most of them using financial ratios. However, a change in the business performance of a company is actually articulated earlier in the nonfinancial information. Their study argued that “dividend” or “retained earnings” appearing in the same paragraph is effective for distinguishing between a defaulting company and a non-defaulting company. Lee et al. conducted a text- mining technique on 10-K reports to extract and visualize industrial service portfolios (Lee et al., 2016). In this study, they used a self-organizing map (SOM) technique to visualize the service portfolio and check word- usage patterns. Lee. et al.’s study predicted stock price by analyzing financial events of 8-K documents (Lee et al., 2014). In this study, the accuracy of the stock price prediction for the following day is improved by 10% when the text is considered. Of course, they emphasized that the analysis should not be considered a comprehensive trading strategy, but stressed that text analysis has predictive power in terms of stock price volatility. SatikoIshikiriyama et al., 2015attempted to ascertain topics of interest in business intelligence using sample analysis on the top 35 read papers in the field. Text-mining techniques determining the frequency of words, or word clouds, were created to identify the main topic of discussion in the paper. The research showed a decreasing interest in business intelligence in academia, suggesting that new areas, such as Big Data, may have received more attention. Pulliza’s study used sentiment analysis in the modeling of speculation in 10-K documents (Pulliza, 2015). This model was applied to the MPQA corpus, to extract features with high correlation with speculative sentences in the 10-K document set. As part of the results, they argued, that "regulation, fund, and supplier were ranked higher than in the documents with the highest amount of speculative sentences".

The studies reviewed indicate that applying text-mining techniques to annual reports is more effective for predicting the future value of business or business trends than analyzing financial information alone. The methods are diverse: ‘Term frequency-inverse document frequency’ (TF-IDF) is calculated to extract important words from text, followed by a clustering method; a SOM or word cloud can be produced as a visualization technique; and techniques such as sentiment analysis were also used. Sentiment analysis is often referred to as opinion mining, which is generally defined as being “aimed at determining the attitude of a speaker, writer, or other subject with respect to some topic or the overall contextual polarity or emotional reaction to a document, interaction, or event." (Wikipedia 2018). In this study, we apply various text or data-mining techniques to text referring to risk factors and sales information in the annual reports of US listed companies. Sentence and word counts, sentiment analysis, keyword extraction and clustering are applied. In addition, we examine the correlation between these text patterns and sales.

This study uses the annual reports of US listed companies (10-K) retrieved from the EDGAR (2018) operated by the SEC (U.S. Securities and Exchange Commission). In the EDGAR’s company filings search, we use the Standard Industry Classification to search the list of companies particular to that industry [SEC EDGAR website] (Fig. 1). This study analyzes the annual reports of the companies belonging to two categories of US SIC (Standard Industry Classification) in the IT sector: 7370 and 7373 (Table 1). This provides a list of all companies that have published annual reports, in the past and present, in these categories. Specifically, it provides company names, state/country identification codes, and company identification numbers and central index keys (CIK), which are linked to a company’s filing report. However, the list also includes those companies that have published reports but no longer exist, or companies that have diminished in size and are consequently no longer obliged to submit reports. In this study, among those companies that submitted a 10-K report after 2015, only those companies that provide information on sales for the last three consecutive years were analyzed. While many companies that submitted 10-K reports provide sales information for the last five years, some only provide information for the last three. Additionally, some smaller companies withhold sales information, specifically sales in “Item 6. Selected Financial Data”, on the grounds of their size. Most of the sales information is shown as “revenue”, but is sometimes marked as “sales”. We do not distinguish between the two notations. Companies that did not provide sales information because of their small size used the following expressions: “As a smaller reporting company, we are not required to provide the information in response to this Item.” This study excludes these firms from the analysis. As a result, we analyze 54 companies, accounting for 7.5% of 717 companies within categories 7370 and 7373, broken down further in Table 2. We use the entire text of “Item 1A. Risk Factors” and the revenue information of “Item 6 Selected Financial Data” (both highlighted in green) for the analysis.

This study uses text and data-mining on items describing risk factors in annual reports to investigate the relationship between the item’s wording and the Compound Annual Growth Rate (CAGR) of revenue in the last three to five years. This study establishes several hypotheses and examines the validity of the hypothesis through data analysis to examine the aforementioned relationships. The hypotheses used in this experiment are shown in Table 3.

Hypothesis 1 can be verified by simply comparing two CAGR values of sales (Fig. 2).

Hypothesis 2 is to verify whether there is a correlation between sales performance and text length of ‘Risk Factors’ of 10-K report. The text length was defined as the number of sentences, the number of words, and the number of words per sentence. This is a direct way to see how each of the three variables representing text length correlates with sales performance. In addition, we examined whether the clustered firms are correlated with the sales performance by clustering firms with three variables such as the number of sentences, the number of words, and the number of words per sentence. That is, it is a method to check whether the clustering result considering the three variables representing the text length is correlated with the sales performance.

To test hypothesis 2, we calculate the number of sentences, words and words per sentence in text regarding “Risk Factors” and compare it with the CAGR values. After obtaining the four values, plotting a ‘scatter plot’ can be used to ascertain the likelihood of a correlation. Subsequently, in the indication of a correlation, the correlation coefficient can be calculated. In addition, using the statistical values from the texts, we use clustering to check whether clusters are formed among companies with similar sales results (Fig. 3).

Hypothesis 3 is a method of examining whether the positive or negative tone in the business text correlates with sales performance. The technique of analyzing the tone of affirmation or negation of text is called ‘sentiment analysis’. According to Taboada et al., this is defined as follows: "Sentiment analysis refers to the general method to extract subjectivity and polarity from text (potentially also speech)" (Taboada et al., 2011). In sentiment analysis, for example, if the sentence contains a positive expression such as ‘good’, the tone of the sentence is evaluated as a positive sentence, and a sentence containing an expression such as ‘bad’ is determined as a negative sentence. In this study, we examine whether the number of positive statements or negative statements correlates with the sales performance of individual companies. For example, if companies with a large number of affirmative statements have good sales performance, companies with a lot of positive sentences will be expected to have good sales performance in the future.

To test hypothesis 3, we conduct a ‘sentiment analysis’ on each sentence in the “Risk Factors” item. Sentiment analysis, or opinion mining, is a method for determining whether a sentence’s tone is positive or negative. We use the ‘RSentiment’ package in R [Package ‘RSentiment manual]. Using the ‘calculate_total_presence_sentiment’ function in this package, the entire sentence is classified as positive, negative, very positive, very negative, neutral, or sarcasm. The number of sentences in each of these six categories is calculated from each company’s “Risk Factors” item. A scatter plot is used to determine the existence of a correlation between the CAGR values and the sentiment analysis. In addition, we apply a clustering technique to the sentiment analysis result to determine whether the companies are grouped by sales performance.

Hypothesis 4 assumes that the words in the business text will appear differently depending on sales performance. In other words, it is assumed that words that have a positive meaning in a group with good sales performance are shown, whereas in the opposite case, a lot of words with a negative meaning are expected to appear. However, this method differs from the sentiment analysis used in the hypothesis 3. The sentiment analysis analyzes the tone of the sentence, while the analysis focuses on what words are used directly. For example, a company with a good sales performance is expected to talk more about future prospects and plans, while a company with a poor sales performance is expected to talk more about immediate sales and risk reduction. Based on these assumptions, we check whether word patterns are correlated with sales performance. In the meantime, we extracted words using TF-IDF, which is widely used in the field of information retrieval, in order to remove unnecessary words and select good keywords that reflect the core of contents.

To test hypothesis 4, texts from the highest and lowest three ranking companies, based on sales’ annual growth rate, are grouped into two files, respectively, and the two text files subsequently compared. Prior to comparison, textual refinement is performed using natural language processing techniques. Refining removes whitespace, punctuation, numbers, and stop-words, and changes upper-case letters to lower case. After that, term frequency-inverse document frequency (TF-IDF) is calculated to select meaningful words, which are then used to create a term-document matrix (TDM). Note that in this process, the two groups of text files are placed in corpus, and the unit of ‘document’ in TF-IDF calculation is a paragraph. In other words, since, from the viewpoint of CAGR of sales, the upper group is represented as one document and the lower group is also represented as one document, in this study, the unit of ‘document’ was set to the paragraph in the text instead of the text per company for TF-IDF calculation. R language was used for data analysis. Specifically, after loading text with ‘readLines’ function of ‘tm’ package, we made it into a corpus by using ‘VectorSource’ function. After that, the TF-IDF is calculated and the TF-IDF values are added for each term and finally the top 200 words with high total values were selected. 200 words are selected for each of the upper and lower groups, so if the word sets in each group are completely different, 400 are selected. We checked the characteristics of the selected words that do not overlap between the two matrices. In addition, a word cloud is created using the TF-IDFs of the words in the two text files (Fig. 4).

In this study, we apply text mining to the annual reports of US companies. The aim was to investigate whether word patterns found in selected texts were related to the business performance of the company. We test four hypotheses: hypothesis 1 postulates that category 7370 companies, which include a large number of companies engaged in SNS activities, will have a better business performance than category 7373 companies, such as Yahoo. Hypothesis 1 is verified. Hypotheses 2 through 4 are analyzed by applying text and data-mining techniques to the risk factors of annual reports. Hypothesis 2 postulates that sales performance affects text statistics such as number of sentences. There is some evidence of correlation between sales performance and text statistics, however, further research is required. Hypothesis 3 postulates that the tone of the text correlates with sales performance. Applying sentiment analysis, no correlation was found and, thus, hypothesis 4 is rejected. Hypothesis 4 postulates that word usage in the text is correlated with the sales performance, and the hypothesis is temporarily adopted.

In summary, we identify a number of correlations between sales performance and the text pattern of company reports by applying text-mining technology. We expect to have more themes to be studied in the future. For hypothesis 2, better results can be expected by changing the data-processing method. Hypothesis 3 is rejected in this instance, but it is expected that better results can be obtained if the classification method of sentiment analysis is optimized for the text of annual reports. The analysis framework for hypothesis 4 needs to be designed to cover all the data, not only the highest and lowest three ranking companies in terms of sales performance. Also, if key phrases are extracted rather than the number of words, results may be more meaningful.

This study provides the following conclusions. Companies with good financial performance and bad companies often use different words. Therefore, it is very meaningful to analyze the words that appear predominantly in the business text in predicting a company’s future sales performance. However, the positive or negative tone of the business text is not relevant to forecasting the future financial performance of the company, as it appears to be lacking in correlation with sales performance. As a result, we hope that this study will serve as a stepping stone to develop the research contents by predicting future financial performance of companies and finding promising business areas.