“Changing dividend payout behavior and predicting dividend policy in emerging markets: Evidence from India”

Dividends have become increasingly important for capital market participants to achieve financial goals in the rapidly changing Indian economy. This study aims to simplify the evolving Indian dividend puzzle by analyzing the dividend trends, examining the evolving nature of firm and macroeconomic determinants of dividends, and developing a dividend policy prediction model. Dividend trends of 3,162 non-financial listed Indian firms from 2006–2022 are studied to gain insights about the Indian dividend puzzle. Regularization and logit models are used to explore the nature of impact of important dividend determinants. Data-mining methods are employed to build a robust model for dividend policy prediction. Trend analysis reveals a decline in the quantum of dividends and proportion of dividend-paying firms with approximately 90% of the dividend-payers belonging to the manufacturing and service sector. Further findings suggest that size, age, maturity, profitability, past dividends, earnings, and bank monitoring of firms had a favorable impact on the likelihood of dividend payments. Macroeconomic indicators such as GDP growth rate, repo rate, percentage change in equity issues, listings, gross fixed assets formation also had a positive impact. The annual percentage change in debt issues and new project announcements at the macro level with investment prospects at firm level negatively impacted dividends. Dividend prediction model based on the random forest technique achieved the highest prediction accuracy of 90.77% and 77.31% under binomial and multi-class situations. These findings are expected to help corporate executives, portfolio managers and in-vestors proactively design optimal dividend policies and formulate their investment strategies.


INTRODUCTION
Dividends entail the distribution of surplus generated by business operations to shareholders.Although various studies have already examined the dividend phenomenon in India, in light of the various changes observed in the Indian economy in the recent past, it has become essential to revisit the dividend puzzle to provide further insights about its evolving nature.Indian firms frequently opt for regular dividends as a signaling tool for reducing the prevailing information asymmetry (Kim et al., 2021;Kanojia & Bhatia, 2023).An increase in participation of retail investors in the equity markets over the last two decades has further bolstered the importance of dividends in India.Majority of retail investors belong to middle income groups from small towns and prefer dividends over capital gains (Graham & Kumar, 2006).Additionally, disincentivizing of fixed income schemes like fixed deposits, Public Provident Fund (PPF), etc. have made dividends a preferred source of recurring income for retail investors.The low level of investor protection in India coupled with the growing retail investor base in the rapidly expanding Indian equity markets further adds to the rising concerns for protecting the shareholders' interests, particularly in the light of recent corporate governance scams.This further increases the relevance of dividends as an instrument for reducing agency problems by disbursing the excess cash to the shareholders (Pahi & Yadav, 2019).Enactment of the new Companies Act in 2013, adoption of the Indian Accounting Standards, increase in obligation of auditors and stricter corporate governance norms after the Satyam scam have further impacted the role of dividends as a substitute of governance mechanism in Indian firms.India has also experienced rapid changes in macro-level indicators alongside the changes in firm level characteristics in the past two decades.Interestingly, high GDP growth rate in India presents the twin contrary arguments of higher investments by firms resulting in lower dividends versus higher profits (propelled by higher growth) resulting in larger dividends.Additionally, the private sector has taken a greater role in the economy in the last twenty years.Private sector firms are highly growth-oriented and thus focus more on capital appreciation through reinvestments than on paying dividends.Further, there was a change in the dividend taxation policy in 2019-2020 that provides for the taxability of dividends in the shareholders' hands against the earlier law of dividends being exempt in the shareholders' hands.The drastic change in the economic and regulatory policy stance of the new government since 2014 and the impact of global shocks (financial crisis of 2008 and COVID-19 pandemic) have further made dividends a highly intriguing research issue.The earlier discussion indicates that the influence of the various determinants of dividends may have undergone a large change.The above context makes it essential to revisit the dividend puzzle amidst the new and increasingly changing economic setup in India.Also, despite the numerous market reforms to improve the information environment, dividends' ability to influence market value of share, especially in emerging markets with high information asymmetry highlights the importance of the need for dividend policy prediction in the Indian context.

LITERATURE REVIEW
The role and importance of dividends has changed with the evolution of the Indian economy.Its importance for protecting the interest of investors highlights its dual nature as an outcome and substitute of good governance.Agency hypothesis posits the use of a firm's resources on extravagant payments to managers and investment in bad projects (Jensen & Meckling, 1976;Manos, 2003).This results in the decline in firm value (Jensen, 1986).Effective governance prevents such misuse of funds and thereby results in higher dividends thus providing evidence for the outcome hypothesis (Sharma, 2011;Pahi & Yadav, 2019; Kanojia & Bhatia, 2022;Fayyaz et al., 2023).On the contrary, Benjamin and Jain (2015) found evidence of substitution argument between dividends and corporate governance in Indonesia.This is especially due to weak shareholder protection mechanisms in the developing markets.Further, Michiels et al. (2015) found that dividends are an effective tool to cool off the intra-family conflicts in a privately held firm.This relationship is strengthened by the governance practices of the family, which subsequently results in the increase in efficacy of dividend policies in minimizing agency issues.This points to the role of dividends as an agency problem trouble-shooter across all types of firms.Additionally, the evolving nature of financial markets has significantly impacted the signaling role of dividends.
Therefore, the trend in dividend payments and propensity to pay was investigated by various studies to understand the dynamic nature of dividends.Fama and French (2001) analyzed the dividend trend of US firms between 1926-1999 revealing a sharp fall in dividend paying firms post 1972 and especially after 1978.The study attributed it to the substantial increase in the newly listed firms that were smaller, had high growth rates and low earnings.A widespread decline in the propensity to pay amongst US firms was also noted.A decline in dividend-payers, combined with an increase in total dividends paid, further reflected the increase in the degree of concentration (Allen & Michaely, 1995).This decline was also explained by the increased substitution of dividends with share repurchases.Firms preferred offloading cash reserves through the share buyback route as it was more tax efficient than dividends.It also increased the EPS with the option of using repurchased shares for issuing employee stock options in future (Dittmar, 2008).Baker and Wurgler (2004) found a new aspect in the form of adjustments in the market dividend premiums by investors to be the reason for the disappearing and reappearing pattern in dividends.Another study by Bates et al. (2009) further explained that the decreasing trends in dividends was also due to the high cash holdings by firms because of the increased riskiness and requirement of timely payment of debt.Ferris et al. (2006) examined the declining dividend pattern among UK firms and found a drop in dividend-paying firms from 75.9% to 54.5%.Unlike US, tax policy amendments and substitution effect of repurchases could not explain this trend.Rather, the catering hypothesis by Baker and Wurgler (2004) seemed to strongly drive the dividend payments.Denis and Osobov (2008) analyzed a dataset of multiple countries to test the global presence of disappearing dividends.They also found evidence of declining propensity to pay dividends along with the concentration phenomenon across US, Canada, France, UK, Germany, and Japan during 1989 to 2002.Similar cross-country analysis by Vieira and Raposo (2007) suggested that the trend was different across UK, France, and Portugal during 1994-2002.
They found no decline in case of France but highly erratic dividend stream in case of Portugal, thereby indicating varying trends across economies with different characteristics.Another study by Ferris et al. (2009) examined the disappearing dividends across 25 countries and found that propensity to pay had declined more in common law countries than civil law countries during 1994-2007.However, an extensive study covering 33 nations by Fatemi and Bildik (2012) found the decline to be more prominent in weak investor protection nations with civil law.Also, the long-established linkage between dividends and future earnings was negated and along with share repurchase substitution raised a question on their signaling use (Grullon & Michaely, 2002;Grullon et al., 2005).Dividend-increasing firms experienced a decline in market risk and an increase in future prices (Grullon et al., 2002).However, another study suggested that the use of dividends (firms experiencing regular stable cash flows) and stock repurchases (firms with non-recurring non-operating cash flows) were done by firms with different characteristics and thus are not substitutes (Jagannathan et al., 2000).
Although, it was well established that propensity to pay dividends had decreased, but a study by Grullon et al. (2011) suggested an increase in net cash disbursements to shareholders through other routes during the same time.This highlighted the shift in preference of firms from dividends to other ways of distributing cash to shareholders.However, dividend payments bounced back in the early years after 2000 due to an increase in the maturity of newly listed firms and restoring of investor confidence by reduction in the investment in wealth destroying projects.It was further supported by the tax rate cut of 2003 (Julio & Ikenberry, 2004).
The change in propensity to pay dividends was observed across 18 economies in another study by Kuo et al. (2013).It was impacted by risk and liquidity with catering playing an important role in the riskreward criterion especially in the high investor protection regimes of common law economies.
Michaely and Moin (2022) enlarged the time frame of the analysis by studying dividend pattern from 1970 to 2018 and found that dividend decline was partly due to changes in firm characteristics and partly due to changing firm tendency.The reappearance of dividends, post 2000 was due to delisting of non-paying firms at a fast pace.Banks on the contrary did not exhibit any decline in dividends since 1978 till the onset of crisis in 2008.Later, banks resorted to aggressive reductions to preserve and support capital requirements after the crisis deepened (Floyd et al., 2015).
Developing economies on the other side had a different experience.Higher propensity to pay was manifested in the form of dividends by state-owned firms in China as a tunnelling tool rather than to alleviate the agency costs of minority shareholders (Lee & Xiao, 2004).Additionally, leverage, growth opportunities, concentrated ownership and smoothened dividends played a crucial role in the payment of dividends in excess of the regulatory requirement in Brazil (Procianoy & Vancin, 2014).
India experienced the dividend-disappearance phenomenon during 1995-2002, reappearance between 2003-2008 and repeat of disappearance from 2009-2013.Dividend payout ratio and dividend yield showed a volatile path during 1995-2013.The number of dividend-payers had decreased but not as strongly seen in the developed world.Similar to the earlier findings in the developed countries, concentration of earnings and dividends with increase in aggregate dividends and earnings was experienced in India during that period.However, dividend payers experienced higher growth than non-payers in India as opposed to the findings in the developed countries.Catering considerations also seemed to have a significant influence on the propensity to pay dividends (Labhane, 2017).Another study on India by Pahi and Yadav (2021) revealed that dividend tax was the major culprit of decline in the propensity to pay dividends.Effective governance mechanisms gave a boost to dividends during 1990-2016.They found dividends to appear, disappear and followed by reappearance in later part of the time period 1990-2016.
Thus, the earlier discussion points to the varied dividend trends observed across time and region, which triggers the need for dividend policy prediction.Only a few studies in the past have attempted to build empirical models to achieve this objective.Three separate studies used data-mining techniques on a small sample of 137 listed Korean companies from 1980-2000 to predict the dividend policy under a binary classification scenario of dividend-payers and non-payers.Bae (2010) found support vector machine (SVM) to achieve the highest accuracy.Kim et al. (2010) found knowledge integration (KI) method to be more accurate than decision tree (DT) and neural networks (NN).Subsequently, Won et al. (2012) found genetic algorithm-based knowledge refinement (GAKR) to give the highest prediction accuracy.
Longinidis and Symeonidis (2013) used DT and NN on a sample of 244 Greek companies from 2007-2009 to predict dividend policy in a binary class scenario and compared the results with logistic model.DT and NN were found to be far more accurate.Further, Kosala (2017) also found DT to more accurately predict the likelihood of paying dividends in comparison to SVM, logistic regression and multi-layer processing (MLP) in the context of 366 Indonesian listed firms between 2007-2009.
The above discussion suggests that the nature of disappearing dividends phenomenon has changed with the passage of time.There is also a dearth of literature on empirical analysis deciphering the div-idend puzzle particularly in India after the numerous changes observed in the Indian economy along with the global shocks in the form of COVID-19 crisis and geopolitical tussles.These events along with world economic slowdown and rising Indian stock markets further adds to the complexity of the issue.Increased understanding of the dividend puzzle will have wider implications for market participants.It will also help in accurate prediction of dividend policy of firms.Lack of accurate prediction models impair efficient investment and financial planning by various financial market participants.Therefore, this study tries to find answers to the existing gaps in the literature.Specifically, this study aims to examine the changes in the dividend pattern in India, investigate the evolving role of micro and macro-level determinants of dividend on the likelihood of dividend payments by Indian firms and develop a dividend policy prediction model in the Indian context.

METHODS
Initial sample of the study included all non-financial Indian companies listed on the National Stock Exchange and Bombay Stock Exchange.The sample period from 2006-2022 was covered to analyze the impact of major shift in the business environment in India over the last two decades.Companies with missing data for large number of years were excluded with 3,162 companies remaining in the final sample.The factors that are studied for analyzing the likelihood to pay dividends and further for dividend policy prediction are determined from an exhaustive set of potential dividend-influencing variables.These variables capture firm level characteristics, macroeconomic indicators, and crisis shocks of 2007-2008 (financial crisis) and 2019-2020 (COVID pandemic).Equity dividend to the net worth ratio was used as the dependent variable representing dividend payments.All the variables used in the study have been explained in Appendix, Table A1.All the company specific continuous variables were winsorized at the 5 th and 95 th percentile values to remove the effect of outliers.Data for firm-level variables was extracted from the Prowess IQ database maintained by Centre for Monitoring Indian Economy (CMIE).Data of macroeconomic variables was obtained from CMIE Economic Outlook and RBI database.
Investment Management and Financial Innovations, Volume 21, Issue 1, 2024 http://dx.doi.org/10.21511/imfi.21(1).2024.20 The paper examined the dividend trends from 2006-2022 to understand the evolving behavior and characteristics of dividends in India.Number of firms, proportion of firms, proportion of firms with positive and negative earnings and industry composition was calculated on annual basis.These calculations were performed for both dividend-paying and non-paying firms to capture the different aspects of dividend trends.Yearly mean, standard deviation and quartile values were calculated for equity dividend to net worth ratio and dividend payout ratio to study the trend in size of payout by firms during the sample period.
Regularization techniques consisting of ridge regression (Hoerl & Kennard, 1970), five approaches (cross-validation (CV), adaptive, plugin, Bayesian Information Criterion-BIC and square-root) of lasso regression (Tibshirani, 1996;Hastie et al., 2015) and elastic net regression (Zhou & Hastie, 2005;Zhou, 2013) were used to select the determinants with the maximum out of sample prediction.A final dataset was divided into training and testing sets having equal data points with 10-fold cross-validation.Logistic regression was used for investigating the impact of selected determinants on the likelihood to pay dividends (Labhane, 2017;Pahi & Yadav, 2019).Logit regression model is represented by equation 1: ( ) where Y it takes a value equal to one if the firm pays dividend in a year, and zero otherwise.X 1it to X nit represent the independent variables determining dividends.β 0 is the constant term and β ) and non-paying firms (represented by 0) in each year.This was done to develop a model for predicting whether a firm will pay dividend in a particular year or not.Further, the multi-class dividend prediction scenario transformed the dividend policy changes into five categories of zero dividend in the previous year to positive dividend in the current year, current year dividend was greater than last year dividend, current year dividend was same as last year dividend, current year dividend was less than last year dividend and dividend in the previous year to zero dividend in the current year.These five categories were represented by five categorical variables of 2, 1, 0, -1 and -2.These categories captured if the dividend was initiated, increased, constant, decreased or stopped.SVM, KNN, Logit, RF and DT techniques were used for predicting these five changes in dividend payments.Data was divided into two sets of training and testing data with each set containing 50% observations.Five-fold cross-validation was used for obtaining the results.Grid search was employed to find out the optimal value of the hyper-parameter for all the techniques used in dividend policy prediction under binary and multi-class scenarios.Linear kernel function was found to be appropriate for SVM.The cost function (C) was selected from a range of 0.1, 1, 10, 100 and 1000 for SVM and Logit.Range of 1 to 25 and 2 to 100 was taken for KNN and DT to find the optimal value of K=nearest neighbors and maximum depth.Range for the number of trees in RF was 100 to 1,000 with an interval of 100.Singular value decomposition solver was used in LDA as it does not compute the covariance matrix and thus is appropriate for a model with a large number of features.Note: Table 3 presents the proportion of firms with positive and negative EPS amongst dividend-payers and non-payers on yearly basis.

Dividend trend analysis in
Comparison of firm characteristics (annual mean values) of dividend-payers and non-payers reveals that investment prospects of payers were 1.5 times higher than non-payers with greater increase seen amongst non-payers.Sales growth for payers has been higher than non-payers except in 2022 with visible trend of decline in the difference amongst them.Return on assets and EPS of payers has been five to eight times higher than non-payers dur-ing the sample period.Beta and change in profits have been observed to be higher for payers than non-payers except in the initial years of the study period.This is different from the findings of the developed world (Fama & French, 2001;Denis & Osobov, 2008) and has been found to be more evident after the financial crisis of 2007-2008 to 2009-2010.Liquidity of payers became greater than non-payers after 2013 and free cash flows of payers was mostly greater than non-payers during the sample period.Payers were much larger in size and have half the leverage than non-payers.Liquidity of shares of payers ranges on an average between one-third to one-fifth of the non-payers thereby reducing the need for non-payers to pay dividends.The tangibility of payers was slightly less than non-payers till 2014 and thereafter it was slightly higher for them.Payers were older and more mature than non-payers.Retained profits share in the equity increased from more than 50% to 80% for payers whereas it increased from close to 15% to 31% for non-payers during the sample period.Effective tax rate was higher for the payers whereas greater proportion of shareholding was owned by institutions and promoters in payers than non-payers.Institutional stake was more than double of that of non-payers whereas the difference in promoter stake was small.Both have raised majority of debt from banks and the payers had a greater number of banking relations indicating better bank monitoring of operations of payers.
Descriptive statistics consisting of number of observations, mean, standard deviation, minimum and maximum values of the variables are provided in table 5.The absence of the multi-collinearity was indicated by low correlation (less than +0.8) between the independent variables.Table 6 presents the two most parsimonious set of dividenddetermining variables identified by regularization techniques.Plugin and BIC lasso selected the most parsimonious model with 19 (Sparse Model 1) and 24 (Sparse Model 2) determinants.These are further employed in the logit model to gauge their impact on the likelihood to pay dividends.They are also used in the dividend policy prediction models under binary and multiclass scenarios.Note: Table 4 presents the proportion of firms from manufacturing, service, real estate, electricity and mining sectors amongst dividend-payers and non-payers on yearly basis.

CONCLUSION
This study aims to investigate the dynamic nature of dividend phenomenon in India using a holistic approach.Further, the study aims to develop a contemporary model to predict the dividend policy based on modern data-mining techniques.The results demonstrate a substantial decrease in the percentage of dividend-payers and quantum of dividend payments during 2006-2022.COVID-19 affected years of 2021 and 2022 experienced a large decrease followed by a substantial jump in the percentage of dividend-payers and size of dividends.The percentage of firms with positive earnings decreased and then increased.This indicates that these firms stopped paying dividends due to negative earnings and continued to do so even after moving into the positive earnings category.Manufacturing and service sector dominated the dividend-payers category.Analysis of the annual trend in firm characteristics reveal that dividend-payers were bigger, older, more mature, more profitable; had higher sales growth, investment avenues, banking relations, institutional shareholding, and lower liquidity of shares in comparison to non-payer firms.Further logistic regression results suggest that larger, older, mature and profitable firms with higher past year dividends, current year earnings, effective tax rate, institutional stake holding, bank monitoring and audit quality are more likely to pay dividends.GDP growth rate, repo rate, income velocity of money, annual percentage change in equity issues, listings and gross fixed assets formation at the macro level increase the likelihood to pay dividends.Firms with higher investment opportunities are less likely to pay dividends.Annual percentage change in debt issues and new project announcements at the macro level decrease the chances of dividend payment.Dividend policy prediction models based on regularization and data-mining found random forest attain the highest prediction accuracy of 90.77% and 77.31% under binary and multiclass scenario.
This study will be helpful for firms' management to understand the evolving trend in dividends and the contemporary forces influencing payment of dividends.It will be useful for portfolio managers and investors to choose the target firms according to the characteristics that are more appropriate for their investment objectives (preference for dividends versus retention of profits).This study also provides a prediction model that can be helpful for firms' management, portfolio managers, retail investors, government, market analysts and other market participants to devise their future actions proactively by leading the market occurrence of dividends.
1 to β n are the coefficients that indicate the change in the probability to pay dividends with a unit change in the respective variable, ε it represents the prediction errors.
analysis (LDA), logistic regression (Logit), decision tree (DT), K nearest neighbors (KNN), support vector machine (SVM) and random Forest (RF) were used for this purpose.DT has been ear-

Table 1 .
Dividend payer and non-payer firm composition It rose for the next 3 years to reach 25.1% in 2016.It fell thereafter before rising sharply to reach 27.98% in 2020 (low denominator ef-fect) and then fell sharply again to reach 19.95% in 2022.The top 25% of the firms paid very large dividends as can be seen in quartile 3 values for both the measures of dividend payments (equity dividend to net worth ratio and dividend payout ratio).This highlights the concentration of dividends amongst the top quartile firms.

Table 2 .
Trends in dividend payouts

Table 3
shows that the percentage of payers with positive EPS (earnings per share) have been in range of 95% to 98%.Non-payers with positive EPS decreased in the initial 8 years from 99.01% in 2006 to reach 56.75% in 2013 and then it increased in the future years to reach 76.36% in 2022.Growth in the proportion of firms with negative EPS both amongst dividend-payers and non-payers also led to decrease in the proportion of dividend-payers and size of dividends.

Table 3 .
Decomposition of payers and nonpayers into firms with positive and negative EPS

Table 6
also provides the results of logistic regression analysis that shows the nature of impact of variables selected by plugin and BIC lasso on the Investment Management and Financial Innovations, Volume 21, Issue 1, 2024 http://dx.doi.org/10.21511/imfi.21(1).2024.20

Table 4 .
Industry composition

Table 5 .
Descriptive statistics See Appendix, TableA1for variable description and variable identification codes.The table shows the number of observations, mean, standard deviation, minimum and maximum value for all the firm-specific continuous variables. (1)p://dx.doi.org/10.21511/imfi.21(1).2024.20

Table 5 (cont.). Descriptive statisticsTable 6 .
Variables selected by regularization methods and logit results of the two sparse models

Sparse Model 1)- Direction of impact with significance of selected variables Determinants selected by BIC lasso Logit Results (Sparse Model 2)- Direction of impact with significance of selected variables
Note: Legend x represents the variables selected as determinants by Plugin and BIC lasso.The variable codes and definition can be referred from Appendix, TableA1.Symbols ***, ** and * indicate significance at 1%, 5% and 10%, respectively.Investment Management and Financial Innovations, Volume 21, Issue 1, 2024 http://dx.doi.org/10.21511/imfi.21(1).2024.20

Table 7 .
Binomial scenario dividend policy prediction results

Sparse Model 2 (24 determinants selected by BIC Lasso) No. of Observations = 26,636 Technique Best Parameter AUC Cross Validation
Note: AUC -Area under the curve.Cross validation mean score, accuracy, precision, recall and F1 score values are out of 1. 0 in precision, recall and F1 score represent the values for dividend not paid category, and 1 represents values for the dividend paid category.
Age of the firm Natural log of number of years since incorporation AGE Four different dividend tax regimes during 2006-07, 2007-08 to 2013-14, 2014-15 to 2019-20 and 2020-21 to 2021-22 are represented by 3 dummy variables representing years from the latter 3 tax regimes.These dummy variables take a value equal to 1 if a year belongs to the particulars tax regime, else 0. Audit Quality Dummy variable is equal to 1 if the auditor is a Big5 Firm for a year and 0 if it is not a Big5 firm.AUDIT 20 Major Offices Location Dummy variable is equal to 1 if the location of either the head, corporate or registered office is in Mumbai, Delhi, Kolkata, Chennai, Bengaluru and Hyderabad and 0 for any other city (representing mining, service, manufacturing and electricity) showed the classification of firms into five Industries (mining, service, manufacturing, electricity and real estate).Dummy variable is equal to 1 if a firm belongs to the particular industry, else 0