Determinants of dividend payout and dividend propensity in an emerging market, Iran: an application of the LASSO

ABSTRACT

Accurate prediction of dividends is important for market participants such as investors, firm managers, and monitoring authorities, as they can, respectively, invest, manage dividend decisions, and monitor dividend policies more effectively. We identify the most relevant variables for predicting the dividend payout of the firms in an emerging market, Iran, using the least absolute shrinkage and selection operator (LASSO). The advantages of the LASSO include: enhancing the prediction accuracy of the dividend model, improving interpretation of the results, and applicability to high-dimensional data. We obtain several results. First, some fundamental determinants of dividends in the industrialized economies such as market-to-book ratio and current ratio, do not play a role in deciding dividends in Iran. Second, LASSO-selected variables outperform the variables commonly used in the literature in terms of model fit and prediction accuracy. Third, business risk, leverage, return on assets and effective tax rate are the most important predictors of dividend propensity of the Iranian firms. Fourth, if the support vector machine algorithm, an often-used classification method, is combined with LASSO-selected variables, it can better discriminate between dividend-paying and dividend non-paying firms than other methods such as logistic regression and linear discriminant analysis.

Abbreviations: LASSO: Least Absolute Shrinkage and Selection Operator; TSE: Tehran Stock Exchange; RMSE: Root Mean Squared Errors; MAE: Mean Absolute Errors; ROC: Receiver Operating Characteristics; GMM: Generalized Method of Moments; MENA: Middle East and North Africa region; AIC: Akaike Information Criterion; BIC: Bayesian Information Criterion; LARS: Least Angel Regression; OLS: Ordinary Least Squares; AUC: Area Under Curve; BS: Brier Score ; OA: Overall Accuracy; LDA: Linear Discriminant Analysis; SVM: Support Vector Machine algorithm; LR: Logistic Regression.

KEYWORDS:

• Dividend payout prediction
• emerging markets
• Iran
• least absolute shrinkage & selection operator (LASSO)
• variable selection

JEL CLASSIFICATION:

• G35
• C51
• C52
• C53

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ http://www.donya-e-eqtesad.com/fa/tiny/news-699564 and Board Activity Report of TSE for the financial year 2015.

² Breiman (1996) defines stability as follows: ‘If the regression equations generated by a procedure do not change drastically with a small change in the data, the procedure is called stable’.

³ See, for example, Rencher and Schaalje (2008, 169–174).

⁴ For a detailed description of LASSO and its generalizations see Hastie, Tibshirani, and Wainwright (2015).

⁵ Dickens, Casey, and Newman (2002), Aivazian, Booth, and Cleary (2003) and Mancinelli and Ozkan (2006) use the natural logarithm of sales as a measure of size.

⁶ Patra, Poshakwale, and Ow-Yong (2012) and Nizar Al-Malkawi (2007) use the ratio of total liability to book value of common equity as an alternative measure of leverage.

⁷ Patra, Poshakwale, and Ow-Yong (2012) and Aivazian, Booth, and Cleary (2003) use the Return on Equity (ROE) as the measure of profitability.

⁸ Aivazian, Booth, and Cleary (2003) and Akhtar (2015) use the standard deviation of the return on investment and standard deviation of stock return an index of risk, respectively. Also, Dickens, Casey, and Newman (2002) use coefficient of variation of net income for the same purpose.

⁹ Anderson et al. (2011) define the firm’s age as the number of years a company is listed on the stock market.

¹⁰ http://en.intamedia.ir/pages/default.aspx?mode=show&lan=en&id=corporate_income_tax.

¹¹ In this paper to calculate the LASSO, we use the GLMNET package in the R software (Friedman, Hastie, and Tibshirani 2010). The results are based on the 10-fold Cross-Validation (CV) for tuning parameter estimation and cyclical coordinate descent method (Friedman et al. 2007) for regression coefficient estimation.

¹² According to La Porta et al. (2000, 2008), common law countries (England and its former colonies, such as the US, Canada, Australia, New Zealand, etc.) have the strongest protection of outside investors (both shareholders and debtholders) than civil law countries (France and many countries Napoleon conquered, former French, Dutch, Belgian and Spanish colonies, and Germany and Germanic countries in Europe, etc.). Common law originated in British law and civil law originated in Roman law and particularly French civil law. Common law is associated with better contract enforcement, and greater security of property right and civil law is associated with a higher level of government ownership and regulation than common law.

¹³ The result remains robust if we use 2012–2015 as an alternative testing period.

¹⁴ The result remains consistent if we use 2012–2015 as an alternative testing period.

¹⁵ The ROC curve is a suitable graphical method for displaying the two types of errors rate: true positive rate and false-positive rate. The former is the sensitivity: the fraction of dividend-paying firms that are correctly identified. The latter is 1-specificity: the fraction of non-dividend paying firms that are classified incorrectly as dividend-paying firms (James et al. 2013, 148).

¹⁶ See, for example, James et al. (2013, chapter 4 and 9) for description of classification methods.

¹⁷ To do support vector machine, we use e1071 package in the R software.

¹⁸ In the testing period of 2012–2016, the measure of Area Under Curve (AUC) of logistic regression (LR) (0.941) is slightly higher than that of support vector machine (SVM) (0.940).