Archive for category: Research

You are here: Home / Research

Social Market Analytics, Inc.

A Leader in Unstructured Financial Data

, , , , , ,

SMA Launches TSX Asset Class to Social Media Feeds

/by

Social Market Analytics converts textual data into quantitative signals for the investment community. To complement our U.S. Equity feed, we recently launch a Twitter based sentiment feed covering the largest TSX (Toronto Stock Exchange) equities. The initial universe included 233 TSX equities, and recently expanded to include an additional 200+ equities. The out-of-sample date for this dataset is January 20th, 2022 with history extended back to the beginning of 2020.

Similar to other asset classes in SMA’s database, the TSX asset class publishes Activity, S-Factor, Short Squeeze, and Hard to Borrow data feeds. To test the robustness of this new asset class, we conducted Daily Quintile and Threshold tests on the S-Factor feed.

TSX Equities Quintiles

In this Quintile test we look at SMA’s S-Factor feed at 3:40 pm ET (20 minutes prior to Market Close). Stocks are placed into different Quintile buckets based on the value of their S-Score. The S-Score is one of fifteen factors supplied in SMA’s S-Factor feed and provides a daily view of Social Sentiment from Twitter. S-Scores greater than 2 or less than -2 are considered extreme sentiment, while values closer to 0 indicate neutral sentiment. The lowest 20% of S-Score values are placed in Quintile 1 while the highest 20% of S-Score values are placed in Quintile 5. We look at subsequent Close-to-Close returns with each stock equally weighted within each quintile. Below is the daily cumulative return series of the TSX asset class quintiles.

After placing stocks in their proper Quintiles and taking the average return of each bucket, stocks with higher S-Scores tend to outperform stocks with lower S-Scores. The graph above shows a monotonic factor. This graph includes a large spread between Quintiles 1 and 5, which would result in a cumulative return of nearly 70% in 27 months. On average, the TSX asset class publishes 141 securities each market day. This number will only increase as SMA’s TSX universe expands.

We further explore the S-Score signal by filtering on securities that had 3 or more Tweets within the previous 24 hours (S-Volume >= 3), then conduct the same quintile analysis.

This new filter reduces the number of securities distributed in the graph from 141 to 95 securities each day. However, it also increases the Sharpe ratio and spread between Quintile 1 and Quintile 5.

TSX Thresholds

For this test, we look at all TSX securities at 3:40 pm ET and place securities in the ‘Long’ bucket if their S-Score >= 2 and place securities in the ‘Short’ bucket if their S-Score =< -2. Since 2020, there are an average of 5 stocks per day in ‘Short’ bucket and 17 stocks per day in the ‘Long’ bucket. We look at the subsequent Close-to-Close return for each security. Securities are equally weighted within each portfolio. In the Long/Short portfolio, the Long and Short baskets are equally weighted.

When only looking at securities with S-Score >= 2, the annualized return of this portfolio is greater than 45%. This portfolio also has a Sharpe Ratio of 1.57 which is well above the market benchmark (S&P/TSX Composite Index).

Many insights can be extracted using SMA’s TSX Social Media feeds. To find out more on this topic, email us at ContactUs@SocialMarketAnalytics.com or schedule a meeting using our 1 on 1 Meeting Signup.

/by
, , , , , , , , , ,

Ukraine War Impact on Corporate Filings

/by

Social Market Analytics released U.S. Machine Readable Filings (MRF) in partnership with S&P Global Market Intelligence in April 2020. Machine Readable Filings is the first product to provide parsed textual data of SEC Edgar Regulatory Filings at the Document, Part, Item, Section, and Notes level with historical baselines back to 2006. Global Machine Readable Filings (GMRF) was released in April 2021 in partnership with S&P Global Market Intelligence. Global Machine Readable Filings provides parsed textual data of international companies’ Annual, Semi-Annual, and Quarterly Reports and Financial Supplements across ~1.6M documents spanning over 200 countries. These magazine-style documents are parsed according to the structure of the underlying document with machine-readable text organized in multi-level JSON format.

In January of 2022, SMA released the Unstructured Data Terminal (UDT). This tool allows for real-time analysis of corporate filings in an easy-to-use interface. One feature of this product is ThemeX, which allows for real time custom topic searching of corporate filings. With the war in Ukraine just beginning, we were asked how quickly will companies react to this new conflict in their corporate filings?

Below we created a search looking at documents that must include the phrases ‘Russia’ and ‘war’ and has any mention of ‘Ukraine’. This searches all pieces of textual data throughout the entire document on all Global Filings going back 1 full year. The phrase ‘trade war’ is excluded to remove any information that is deemed irrelevant to the topic.

Below are the results of an executed search on the war in Ukraine. The search displays the trend of mentions month over month, a company breakdown, and country breakdown. ThemeX has ability to identify other related phrases and topics to the core theme as well, under the ‘Show Suggestions’ link.

 

The first chart shows the number of ThemeX hits on unique companies and filings by month.  The second table shows the number of documents hit on the ThemeX search for each company. The pie chart shows country exposure amongst the ThemeX hit companies.

If you want to explore the filing with ThemeX, you can read the section in each document where the reference is made. If you click PH Glatfelter, it shows the filing type, date, and section where the phrase was mentioned. From there you can read the entire document and browse SMA patented sentiment.

The invasion started Thursday, February 24th. In the first month, the mention of Ukraine in reference to war has already skyrocketed. Many companies have already added this topic to their Risk Factors section as there is significant worry the war will spread globally. Less than a month into the invasion, nearly 600 companies across the globe have referenced the risk of impact from the war into their corporate filings.

To learn more about this topic or explore other topics in ThemeX, email us at ContactUs@SocialMarketAnalytics.com or schedule a meeting using our 1 on 1 Meeting Signup

/by
, , , , , , ,

Alpha Around the World: Insights into Global Annual Reports

/by

Social Market Analytics (SMA) has created the Global Machine Readable Filings (GMRF) dataset in partnership with S&P Global Market Intelligence. GMRF provides parsed textual data broken down to sections along with metrics like word count and sentiment on ­­­­annual and interim reports for companies around the world. Extraneous information such as images, tables, and page numbers are removed for textual analysis purposes.

In this blog we began investigating word count and sentiment metrics related to Annual Reports (AR) in the GMRF dataset. Each document is linked with a company and each company can be linked to multiple securities. These securities could be different class shares on the same exchange, or the same company listed on multiple global exchanges.

Since these reports only come out once a year, there will only be one value of factors per company each year. To construct a calendar-time portfolio, metrics from the most recent Annual Report for each company is extended up to 12 months. Quintile portfolios are constructed at the end of each month based on values of the most recent factors from a company’s Annual Report. In our analysis below, Quintile 1 contains 20% of stocks with the lowest value of the factor. Quintile 5 contains the 20% of stocks with the highest value of the factor.

For time-series purposes and to avoid look-ahead bias, in this analysis we assume that Annual Reports are released publicly 90 days after the fiscal year end date for that company. For example, if a file had a fiscal year end date on 12/31/2011 the word count and sentiment factors associated with that Annual Report would be implemented into the quintile portfolios on 3/31/2012.

After cleaning the dataset by removing documents with low word count, extreme changes in word count, and filings within four months of each other, we map the documents to the securities’ monthly return. Stocks are placed in quintiles based on factor values. Returns are equally weighted within each quintile.

This chart illustrates the relationship between the length of Annual Reports and future price returns. Higher word count in an Annual Report indicates the company has more information to supply to investors. Quintiles 1 and 2 outperform the universe of stocks in the dataset while Quintile 5 severely underperforms. Quintile 5 contains companies that have produced extremely long Annual Reports over the past 14 years. Long Annual Reports typically contain unnecessary language or ‘fluff’ to ease concerns of investors and portray the company in a stable fashion. This text could be information addressing specific risks and business actions or could be new information on products and legal proceedings. The theory is shorter Annual Reports are more forthright with information regarding business affecting the company’s stock.  This is consistent with the classic Harvard Lazy Prices1 paper findings.

Next, we look at the Average Sentiment across an entire Annual Report. Average Sentiment is calculated by taking the mean of sentiment from each Part within an annual report. Each Part is a level one header, or the overarching categories in a Table of Contents. The sentiment of each Part is calculated by taking the sum of each sentence within each Part. Thus, Parts with more textual data (more sentences) can have a higher magnitude of sentiment. A Total Document’s Average Sentiment is slightly correlated to the documents word count.

Here we see an inverse relationship between a document’s word count and price returns. A low Average Sentiment outperforms the rest of the universe while high Average Sentiment underperforms. This can be traced to the company’s transparency with its investors. Theory is Annual Reports that state the company’s proceedings and risks, with no sugarcoating, tend to have a lower Average Sentiment in Annual Reports.

This analysis are preliminary findings on the GMRF dataset. SMA will explore further on how characteristics of Annual Reports can affect future stock prices by sector and region. To trial this product or learn more about other Social Market Analytics datasets please email us at ContactUs@SocialMarketAnalytics.com

 

  1. Harvard Lazy Prices, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1658471
/by
, , , , , , ,

Using Twitter Social Media for One-Month Holding Periods

/by

Social Market Analytics, Inc. (SMA) has been warehousing and normalizing social media signals, called S-Factors, since December 2011. SMA has collected nearly 10 years of out-of-sample data, which allows back tests of longer holding periods. SMA’s typical S-Factors are a daily view on social sentiment and is great at predicting subsequent price movement in securities 1-3 days out. By aggregating S-Factors over longer time horizons, testers / quantitative models can generate social sentiment signals that are more suitable for trading over weekly, monthly, or even quarterly holding periods.

This blog tests our hypothesis that heavily talked about securities with positive sentiment over the past month will outperform the market. We also expect heavily talked about securities with negative sentiment over the past month will underperform the market.

The strategy below is called the ‘Monthly Multifactor Strategy’ because the signal comprises of 3 S-Factors: Raw-S, S-Volume, and S-Buzz. These 3 factors are a view of sentiment and tweet volume over the previous 24-hour period. By taking these factors once a day, you can aggregate the data over any time horizon to create long-term factors at your desired length. Below are the Long-Term factors created for this strategy:

  • Long-term sentiment, ‘S50’, is calculated by taking the sum of Raw-S over the past 50 market days and standardizing it over a 200-market day baseline
  • Long-term volume, ‘SV50’, is calculated by taking the sum of S-Volume over the previous 50 market days and standardized to a 200-market day baseline
  • S-Buzz is a measure of unusual s-volume activity compared to the rest of the universe

 

 

This is a Monthly Long Only or Long/Short strategy.Trades are entered on the Market on Close price of the last market day of each month. Securities are selected 20 market days prior to trade execution. The Top 50 stocks with the highest Multifactor Score are Longs and the bottom 50 stocks with the lowest Multifactor Score are Shorts.

Stocks are weighted within each portfolio based on the rank of the Multifactor Score. Stronger scores have a higher rank numeric value (ranges from 50 to 1).

The combination of the two portfolios is the Long/Short. The Long and Short baskets have an equal dollar amount invested.

 

 

 

 

About Social Market Analytics:

SMA is a leader in unstructured financial data from Social Media to Machine Readable Filings (MRF). SMA is a preferred Twitter Finance partner and Stocktwits vendor. SMA partners with S&P Global Market Intelligence on Machine Readable Filings (MRF) across U.S. Edgar Filings and Global Annual Reports.

Founded in 2011, Social Market Analytics, Inc. goal is to harness the massive amount of unstructured financial data across both Alternative and Traditions sources into machine readable textual data feeds for Quantitative and Quantamental analysis. SMA’s DNA goes back to its predecessor firm Quantitative Analytics,

Inc. (QA Direct) which was acquired by Thomson Reuters (Refinitiv).

SMA’s Patented processes are unique in the emerging field of Machine Learning, Financial Sentiment Natural Language Processing Textual Parsing, and Source Accuracy.

Contact:Doug Hopkins at doug@socialmarketanalytics.com

 

/by
, , , , ,

Analysis of Subsequent returns on Short Squeeze Conversations

/by

In the wake of the recent GameStop (GME) short squeeze, we received a lot of inbound questions on social media’s predictive power for short squeezes. Social Market Analytics (SMA) classifies different types of conversations on Twitter with one topic being Short Squeezes. Looking at subsequent returns of stocks with short squeeze conversations leads to different return characteristics based on when the conversation occurred. This blog will explore subsequent Open-to-Close return characteristics when the short squeeze conversation occurs prior to market open and the subsequent Close-to-Open returns and when the short squeeze conversation occurs prior to the market close.

We generate our social media metrics on 4,000+ US Equities every minute 24/7. Our historical data is out-of-sample since 12/1/2011 making it suitable for quantitative back-testing. In this blog we used all equities in the SMA universe. Morning signals are generated at 9:10am US/Eastern, afternoon signals are at 3:40pm US/Eastern.

Tweets discussing short squeezes are identified and classified by our Natural Language Processing and Topic Modeling technology. The number of Tweets discussing a short squeeze in relation to a specific security are aggregated and ranked. In this analysis we are looking at the ‘squeeze-count’ metrics in our data feed, which is the Tweet volume that contains any ‘short squeeze’ keywords about a certain security. The Top 50 securities with the highest ‘squeeze-count’ are equally weighted in the ‘Top_Squeezes’ portfolio. If there is a tie on the 50th security, all securities tied are included in the portfolio. Therefore, portfolio counts could be greater than 50.

The first graph illustrates the subsequent Open-to-Close performance of securities with the most short squeeze terms prior to market open. The blue line represents the cumulative Open-to-Close performance of the top 50 stocks with the highest ‘squeeze-count’. For comparison, the black line represents the S&P 500. These securities in the portfolio significantly underperform intraday. It appears that conversations prior to market open are more about people wanting a squeeze versus it happening. The GME squeeze is marked in the graph.

Cumulative performance is -71.34% with a -2.62 T-Stat and a p-value of 0.0045. These returns are significant at the 1% level.  One caveat to the performance is there is no accounting for hard-to-borrow securities.

The next chart is more interesting. Short conversations occur during the day and are aggregated at 3:40pm US/Eastern. The average subsequent Close-to-Open returns of the top 50 stocks with the highest ‘squeeze-count’ are shown by the blue. As you can see from this chart, stocks with intraday short squeeze conversations move significantly higher overnight.

Cumulative return for this theoretical portfolio is 497% with a 2.89 Sharpe, 5.47 Sortino, 4.33 T-Stat, and an infinitesimally small P-Value. This proves the significance of this portfolio at the 1% level.

SMA Short Squeeze Data Feed is available through a RESTful JSON and XML API or as FTP files. The data can be packaged at different timestamps throughout the day.

Data Sample:

At SMA we aggregate these short squeeze conversations twice a day and make them available to customers. To access this data or many other SMA services, please ContactUS@SocialMarketAnalytics.com

/by

Application Of Sentiment Analytics To Fama-French Model

/by

In this paper we expand on the Liew (2016) paper exploring the predictive power of StockTwits based sentiment data. We postulate that using a fine grain text analytics approach on social media posts is a better representative of the sentiment factor. This new model shows improvement in both Fama French 5 Factor Model and the Liew paper.

/0 Comments/by