Logistic Regression as a Baseline for Takeover Probability Models
Summary
This discussion considers estimating the probability that a company receives a takeover bid from signals such as pre-offer price moves, trading activity, rumors, and financial ratios. One answer recommends a logit-link generalized linear model, commonly called logistic regression. It can relate candidate explanatory variables to takeover log-odds and support tests of whether their associations are statistically informative. The answer presents it as an interpretable baseline against which more complex machine-learning approaches can be compared.
A second response notes that sell-side research on takeover likelihood had already appeared and points to examples spanning several years, including some described as using machine learning. The document provides methodological suggestions and leads to prior research, but no dataset, model evaluation, predictive results, or guidance on event labeling and class imbalance. Its claims therefore motivate a research starting point rather than demonstrate that any approach predicts takeovers successfully.
Key ideas
- Takeover likelihood can be modeled using market, news, and company-level explanatory variables.
- Logistic regression maps predictors to takeover log-odds and offers an interpretable statistical baseline.
- More complex models should be compared with a simpler baseline before their added value is assumed.
- Prior sell-side studies are suggested as sources to review, but the discussion reports no model performance.
Tags
Full text
# Machine Learning approach for the probability estimation of certain events # Machine Learning approach for the probability estimation of certain events I am planning a research project on estimating the probability of corporate takeovers. I think that different variables could be indicators to predict takeover bids. For example, price increases in the days prior to the offer, increased trading volume, takeover rumors in stock market magazines, or business ratios. From a methodological point of view, I think that machine learning (ML) techniques are very promising with regard to estimating the probability of a takeover. However, I am not very familiar with this field. Can anyone recommend techniques that can be used to model the probability? Perhaps someone has estimated a probability model with an ML technique in a similar area and can share their experience with me. I would be very happy to receive feedback. Many thanks in advance! ## Answer by kurtosis (score 2) https://quant.stackexchange.com/a/63827 Your best bet would be to look into a logit-link generalized linear model (GLM) -- what some call logistic regression. That would allow you to put various variables into the model and see how each affects the log-odds of a takeover; and, you can test hypotheses about how significant a given variable is. Some people might suggest something more complicated; others might claim that logistic regression GLMs are machine learning. Nonsense. Something more complicated will obscure which variables are more vs less informative. As for ML claiming GLMs: GLMs are solid statistical methods that has worked well for decades. Even if you were to do one of the fancier ML techniques, this would be the baseline to compare against. If you need a reference for this type of model, either Dobson's book or McCullagh and Nelder are the best resources. ## Answer by user42108 (score 0) https://quant.stackexchange.com/a/63831 "I am planning a research project on estimating the probability of corporate takeovers" This is not a new idea. There have been at least four sellside papers on this topic between ~2006 and 2017 (those are the ones I'm aware of). At least two of those use ML. You might want to try to get hold of them to avoid reinventing the wheel. EDIT: papers are - DB, Systematic M&A Arbitrage, 2015 - SocGen, Training your computer to find potential M&A candidates, 2010 - Wolfe Research, MACHINE LEARNING TAKEOVERS, 2017 - MS, Introducing ALERT-E: Our European M&A Likelihood Model, 2017 There was also a JPM paper on this topic from, IIRC, 2006 or 2007.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.