Liu, Huanliang

Use Deep Learning to Predict Cash Dividend Initiation

Cash dividend initiation marks an important milestone in a public corporation’s life. Before cash dividend initiation, companies are usually in the fast-growing stage of their corporate lives. Cash flows generated from business activities are usually used for reinvestments. As a company matures and grows, the returns on their investments decline. At the same time, cash flows generated from prior investments start to accumulate. Firms usually start to pay regular cash dividends to shareholders – the so-called dividend initiation.

The existing literature uses traditional econometric methods such as logit or probit regressions to predict when a public company will initiate cash dividends. The variables used to make the prediction include firm size, firm age, book to market ratio of equity, recent stock returns, stock return volatility, institutional ownership, liquidity, and other macroeconomic and industry features.

One weakness of these traditional econometric methods is that they assume a linear relation between the dependent variable (in this case, whether a firm will initiate dividends or not) and independent variables (the features used to make predictions). In reality, the true relation can be nonlinear. Adding interaction terms and higher moments will partly alleviate the concern but will not eliminate it because the true relation is unknown and can take any functional form.

Deep neural networks (DNNs) are function approximators. If we feed the model with enough data, DNNs will learn the approximate relation between the target and the features, after many rounds of training. Therefore, DNNs are ideal to overcome the weakness of the traditional econometric methods and have the potential to uncover the true relation between various features and corporate dividend initiation decisions.

I plan to use quarterly data from Center for Research in Security Prices (CRSP) and Compustat to examine corporate dividend initiation decisions. I will look at U.S. publicly traded firms from 1980 to 2021 and use five years of data to train a DNN to make predictions on dividend initiation in the next quarter. Because there are thousands of public firms in the U.S. in any quarter and I am using a rolling window method to make predictions each quarter, this will take hundreds rounds of training. Each round of training has thousands of iterations. The project is extremely computationally costly. Therefore, I would request that UKY MCC assist me in this project.

Students:

Computational Method:

Conduct deep learning.

Software:

Python

Grants:

Publications: