The objective of Manzama Insights Health Scores (MIHS) is to distill the news about a specific company into numbers that accurately and succinctly represent their corporate health. Right now, the single estimates for the overall news-based health for Boeing (NYSE: BA), Deutsche Bank (NYSE: DB), and Bristol-Myers Squibb (NYSE: BMY) are -4, 0, and +2, respectively. Manzama further breaks down these numbers into six Factor Health Scores (FHS’s) that categorize the news under the following “Factors”: Financials, Operations, Products & Services, Management, Government, and Partners & Competitors (Table 1).
The purpose of this blog post is to build confidence on the validity of Health Scores as a measure for corporate health by (1) illustrating how companies’ health scores reflect the headlines that mention them, and (2) presenting specific examples of companies’ health scores moving in sync with their stock prices. So what do these numbers mean? Where did these numbers come from?
The following main headlines explain Boeing’s unfavorable Management, Financials, and Government news: the executive in charge of the 737 MAX crisis was fired; Boeing’s Q3 profit fell short; and the FAA confronted Boeing over misleading messages regarding the 737 MAX back in 2016. In other words, Boeing’s -4 MIHS really represents the 737 MAX catastrophe. Note that Boeing’s very positive Partners & Competitors score of +8 (caused by a US$20 million investment in space tourism company Virgin Galactic) was dwarfed by the other Factor scores because there were only 218 articles categorized under this Factor, while there were 833 under Government, 903 under Products & Services, and 1,250 under Operations. In other words, the influence of each Factor score on the MIHS is proportional to the fraction of all articles that are classified under that Factor.
Deutsche Bank’s balanced MIHS is the outcome of adverse Operations headlines, mostly informing of their plans to cut jobs, and their positive Government stories, the most impactful being the company’s claim that they do not have Donald Trump’s tax returns.
Bristol-Myers Squibb’s positive score is primarily driven by the fact that we obtained a larger quantity of financial articles about stockholders increasing their position in BMY stock than decreasing it over the last 30 days. Additionally, news about their immunotherapy drugs getting positive results in lung cancer trials and announcements of their collaboration with Fitbit on enhancing atrial fibrillation diagnostics further boosted their news-based Health Score from +1 to +2.
Firing executives, low profits, schisms with regulatory agencies, and job cuts are almost always considered negative for a company. Partnering with other companies and investing in new lines of business is almost always considered positive for a company. However, it’s not clear whether distancing from Donald Trump’s congressional probe is good or bad for Deutsche Bank. The machine learning algorithm was trained to classify headlines as positive, negative, or neutral according to tens of thousands of manually-labelled examples, but this is the first time in United States history that a company has tried distancing themselves from a sitting President’s tax return scandal, so we cannot expect the neural network to know if this is a good thing or a bad thing. Moreover, sometimes there’s disagreement on whether a complicated headline is good or bad for a company. We needed an external mechanism that validated the decisions that we were making when creating training data and, ultimately, validated the network’s classifications and Health Scores. We decided to use an established, wisdom-of-crowds corporate health metric to provide support for our scores: the stock price.
Stock prices reflect the corporate health of a company through the aggregate behavior of investors. If a company is healthy, then investors are likely to purchase stock expecting future returns, thereby increasing the stock price; if a company is unhealthy, then investors will sell stock, causing a drop in the price. Similarly, the Manzama Insights Health Scores reflect corporate health through the aggregate behavior of news editors. If a company is healthy, then good things should be happening to them, which would translate into favorable headlines written by news editors and a rise in the Health Scores; if a company is unhealthy, then bad things should be happening, which would cause news editors to publish unfavorable headlines and thereby cause a drop in the Health Scores. The point is that the stock price and the Health Scores quantify corporate health by compiling the actions of different groups of people. Therefore, we studied the relationship between the stock prices and Health Scores of 99 companies (Appendix I) by computing correlations during a 90-day window between June 30th and September 27th to see whether the actions of investors and news editors were related.
We standardized the stock prices by computing the percent change in the closing price since the last trading day and standardized the MIHS and FHS by using the absolute change since the last trading day. This resulted in 61 unique data points (the number of trading days in a 90-day interval) for each Health Score per company, equivalent to 42,273 Health Score representations from an aggregate count of 173,327 headlines over that 90-day time period.
One of the main findings of the study is that the absolute change of the MIHS and the percent change of the stock price on the same day are correlated. In other words, when the stock price changes, the MIHS usually moves in the same direction. Three FHS’s drove this same-day correlation during this 90-day time frame: (1) Financials; (2) Operations; (3) Products & Services. This suggests that the actions of investors and news editors are related.
To make this finding more concrete, consider Boeing as an example. Chart 1A presents Boeing’s stock price and their MIHS during the 90-day interval studied. Chart 1B presents a linear regression between Boeing’s percent change in stock price and absolute change in the MIHS on the same day. As can be observed, the linear fit exhibits a slight positive trend, indicating that these two values are weakly correlated. Although this linear regression’s coefficients are not statistically significant, aggregating statistics for the 99 companies in Appendix I yields a statistically significant positive average correlation.
The most important finding of the report is that there are statistically significant correlations between news-based Health Scores and stock prices that transcend short time intervals. For instance, on average, changes in the Financials FHS the day before are correlated to the stock price the day after.
This finding can be exemplified by observing Bristol Myers Squibb’s stock price and their Financials FHS in Chart 2A, and the subsequent linear regression between the change in the Financials FHS and the stock price the day after, presented in Chart 2B. This positive trend is greatly affected by the increase in the Financials Health Score on July 24th, 2019: one day before BMY posted strong second quarter earnings.
The day before, the Financials Health Score had increased because there were more headlines about stockholders growing their positions (Iowa State Bank, Elm Advisors LLC, Surevest Inc.) than stockholders lowering their positions (Beese Fulmer Investment Management Inc, First National Bank of Mount Dora Trust Investment Services) in BMY stock. The following day, their stock pricejumped 5%. In future research, we hope to confirm the consistency and reliability of these patterns for other companies using outlier-resistant correlation metrics and more data.
Although the MIHS’s were never intended to converse well with stock prices, we were able to find statistically significant correlations with changes in stock prices on the same day and on the day after. This finding lends credence to the notion that both statistics track the health of a company. This research supports the exciting idea that Manzama’s News-Based Health Scores are a numerical representation of corporate health. Read the full report here.
Appendix I: Companies
|Company Name||Stock Symbol||90-Day Article Count|
|6||The Walt Disney Company||DIS||5717|
|13||Ford Motor Company||F||2512|
|19||JP Morgan Chase||JPM||2171|
|30||Bank of America Merrill Lynch||BAC||1633|
|31||Johnson & Johnson||JNJ||1600|
|35||Delta Air Lines||DAL||1446|
|40||The Procter & Gamble Company||PG||1364|
|41||Advanced Micro Devices||AMD||1329|
|57||United Parcel Service||UPS||1028|
|58||Hilton Worldwide Holdings Inc.||HLT||1016|
|59||Kraft Heinz Mondelez||KHC||981|
|60||Costco Wholesale Corporation||COST||978|
|62||Gilead Sciences, Inc.||GILD||960|
|68||American Express Company||AXP||848|
|69||Brookfield Asset Management Inc||BAM||830|
|70||The Kroger Co||KR||827|
|80||Union Pacific Corporation||UNP||740|
|85||NextEra Energy, Inc.||NEE||668|
|86||Dominion Energy, Inc.||D||668|
|88||Motorola Solutions, Inc.||MSI||631|
|89||Philip Morris International||PM||621|
|90||General Mills, Inc.||GIS||613|
|92||The Williams Companies Inc||WMB||592|
|94||Credit Suisse Group||CS||581|
|95||Tyson Foods, Inc.||TSN||580|
|97||Lloyds Banking Group||LYG||512|
|98||Dollar General Corporation||DG||483|
 To learn more about Manzama Insights Health Scores, check out the blog post “Assessing Corporate Health with Machine Learning”
 At time of writing: October 25th, 2019 at 11:00 AM PT
 Article counts are presented in parentheses next to the Factor scores in Table 1
 Sources include thelincolnianonline.com, watchlistnews.com, dailypolitical.com, marketbeat.com, amongst others
 Is news about a price hike for a product good, bad or neutral for a company? Some could argue that it is a good thing because it signals high demand for the product; others could argue that it is a bad thing because it will likely decrease future sales. I argue that it is neutral because we do not have enough elasticity information to anticipate an increase or decrease in profits. To learn more about disagreement rates, check out our report on Krippendorff Alpha Reliability metrics (link). The idea here is that if the stock price moves up after a price hike then that suggests investors think the price hike was a good thing, which can act as a sanity check for our own beliefs of what is good, bad or neutral for a company.
 The trend lines superimposed in Charts 1A and 2A are Locally Estimated Scatterplot Smoothing (LOESS) lines of best fit, which are constructed using both the past and future neighboring points.
 Although the linear regression is being biased by this outlier, the average correlation and the correlation of the median remain positive.