A few machine learnings projects will be posted here soon.
A test result for an algo trading based on what I have below will be posted here over the Summer.
----------------------------------------
I have worked on several studies using data on blockholder activists (Schedule 13D). The blockholding information is filed with the SEC and disseminated via the SEC EDGAR. Considering the significance, these filings often accompany with quite positive reaction from the market.
While the holding information (%; blockholder identity) gets the most attention, some other aspect in Schedule 13D can make quite difference when it's evaluated by the market. For example, early communication between the blockholder activist and the target management can be perceived more positively (see the figure below, and the details are available in the following paper: Let’s Talk Sooner Rather Than Later: The Strategic Communication Decisions of Activist Blockholders).
Then, I use a Machine Learning algorithm to establish a link between several aspects that are ignored in the literature (input) and the market reaction based on abnormal returns for the [+1, +5] window (output). This link allows me to identify target firms with better market assessment so I can maximize my investment performance by following the selected target firms.
Why not follow all the target firms by Schedule 13D?
1. Often target firms are very small and suffer from poor market liquidity. You need to consider your investment performance after deducting round-trip transaction costs (bid-ask spreads).
2. Some Schedule 13D filings do not convey much new information (thus no positive market reaction) because the blockholding information may be disclosed via other associated filings. Certain criteria need to be met to find suitable target firms, and those criteria should be used for the data preparation stage in your Machine Learning algorithm.
Investment Performance (based on my actual transactions):
#1. Personalis (PSNL): 13D filed on 8/24/2020; Bought at 22.0969 on 8/25/2020; Sold at 23.0000 (limit order) on 8/31/2020; Return 4.087%; Note - trying to strictly follow the model specs for the 5-business day holding period
Currently I am developing an automated trading system based on this algorithm. More details will be posted.