Just recently released into production, the StarMine M&A Target Model is the newest predictive model from LSEG StarMine. This model has been designed to predict the likelihood of a publicly traded company becoming the target of a takeover or merger offer. Like all StarMine models, it ranks companies on a 1-100 scale. In the case of this model, companies with high scores are, on average, nine times more likely to become a target than those with low scores (Exhibit 1).
The model has two major components: A Fundamental Component and a Text Component. The Fundamental Component uses LSEG content from data sources such as merger & acquisition (M&A) deals, proxy fights, financial statements, valuation metrics, the StarMine Combined Credit Risk (CCR) model and other structured data. While the model has global coverage, StarMine research found regional differences – U.S. companies are more likely to be acquired than those in other regions.
There are also sector trends – M&A activity tends to move in waves across sectors. A recent wave of M&A activity in a certain sector suggests that sector will attract further interest. Companies in financial trouble, measured by the probability of default, are more likely to be seen as an attractive turnaround opportunity. Companies that have been engaged in a proxy fight, perhaps by an activist investor, are more likely to be acquired, as are those that have been a target before. Small companies make more viable candidates than the largest companies, which may be too expensive to acquire. Companies that pay a dividend are less likely to be acquired. And finally, cheaper companies are more enticing targets than those that appear overvalued.
In its Text Component, the M&A Target Model uses a large language model named BERT-RNA from LSEG. BERT-RNA is a financial domain specific adaptation of BERT (short for Bidirectional Encoder Representations from Transformers), an open source foundation model from Google. BERT is similar to, but distinct from, Chat GPT. One difference is that BERT is better suited for tasks like classification (event prediction), entity recognition, and sentiment, whereas GPT is better suited to tasks like creating new text and chat (generative AI). BERT-RNA is based on the case-sensitive version of BERT, with further language model pre-training using large amounts of in-domain data from business and financial news in the Reuters News Archive. Pre-training BERT on the Reuters News Archive makes it better able to understand the financial and business domain that Reuters covers, whereas GPT is trained on data across the internet, not finance-specific data. BERT-RNA is then fine-tuned with labelled news articles for M&A target predictions.
Exhibit 1: Global M&A Frequency by Decile January 2000 – June 2023
Source: StarMine White Paper
We can illustrate the model and its predictive power with a recent example. On Sept. 21, Cisco Systems announced it would be buying Splunk Inc. (SPLK.O) for $28 billion. At the time of the announcement, Splunk had a North America regional rank of 98 and a Technology sector-relative rank of 99 (Exhibit 2), indicating a high likelihood it would become the target of a deal.
Looking at its history (Exhibit 3), Splunk had a regional score of 100 as far back as early March 2022, roughly a year and a half ago. The M&A Target Model’s detail page can be accessed by entering the symbol, followed by MATM in the Workspace search bar or under the Event pulldown on the Company Overview page.
Exhibit 2: Splunk M&A Target Model Region, Country, Sector and Industry Relative Ranks
Source: LSEG Workspace, StarMine
Exhibit 3: Splunk’s M&A Target Model Regional Score vs. Price 2-Yr History Chart
Source: LSEG Workspace, StarMine
Finally, Exhibits 4-5 shows Splunk’s Text Mining and Fundamental Component scores of 96 and 97, respectively, along with news excerpts and fundamental subcomponents. Quants, who take this model as a feed, also receive access to current and historical component and subcomponent data, which they can use in their entirety or in part, if they so choose.
Exhibit 4: Text Mining Component Score and Details
Source: LSEG Workspace, StarMine
Exhibit 5: Fundamental Component and Subcomponent Details
Source: LSEG Workspace, StarMine
Unrelated to this model, we find that this deal announcement comes on the heels of a blow-out quarter. Spunk reported Q2 EPS results of $0.71, 58.3% above its consensus (I/B/E/S Mean) estimate of $0.45. On the day of its earnings release, the StarMine Predicted Surprise % was 6.13%, the difference between consensus and the StarMine SmartEstimate® of $0.48. When the Predicted Surprise is 2% or more in either direction, it gets roughly 70% of subsequent earnings surprises directionally correct and revenue surprises about 78% directionally correct. StarMine is often best known for its ability to rate and rank sell-side analysts’ accuracy and, from that, create SmartEstimates.
Exhibit 6: Splunk Historical Surprise Table – EPS Annual and Interim Results (HSUP in Workspace)
Source: LSEG Workspace, StarMine
As the first commercially available model of its kind, we anticipate the new StarMine M&A Target Model will be used to help hedge fund managers reduce risk on the short side of their portfolios, by avoiding shorting companies with the greatest likelihood of receiving a buy-out offer, triggering a price jump and short squeeze. The buy-side may also value being able to add this event-prediction model to their new idea generation screens (SCREENER) and their holdings monitor (PULSE). StarMine M&A Target Model scores will be available in these two LSEG Workspace Apps in the near future.
The model should also be interesting to investment bankers, consultants and M&A law practices seeking potential new clients, to help originate new deals or trying to anticipate upcoming buyout offers. Investment banks, in particular, are investing significantly in technology to drive efficiency, growth opportunities and the automation of workflows.
