The extant literature argues that a company’s decision to change the existing dividend policy is influenced to a significant extent by the past and present financial profiles of the firm (see, for example, Lintner, 1956; Wansley & Lane, 1987; Healey & Palepu, 1988; Jensen & Johnson, 1995; Lonie et al., 1996; Benartzi et al., 1997). In addition, the dividend signalling hypothesis suggests that any changes in current dividend levels provide information about the future performance of the firm; firms that increase dividends demonstrate an improvement in their financial performance over the long run while those that decrease dividends experience deterioration in their future performance.
In this context, if future dividend changes could be predicted using past and present financial profiles, then investors should be able to implement potentially lucrative trading strategies. In particular, they could use such predictions to categorise companies into two groups: ‘good future performers’ and ‘bad future performers’; they could then take a long position in the first group and a short position in the second group. The objective of this study is twofold. First, we test the predictability of subsequent dividend changes and second, we examine the profitability of an investment strategy based on these forecasted variations in dividends. In terms of methodology, we evaluate logit models that use past and present financial profiles for a sample of firms to predict the probability of one year-ahead changes in dividend policy. Specifically, we use financial information for New Zealand companies for the six-year period 1995-2000 to determine financial ratios which, ex-post, are found to be good indicators of subsequent dividend changes. Employing a stepwise approach, we determine the multi-logit model that provides the best explanatory power for upcoming dividend changes during this model development period. Thereafter, we use this model to forecast the probability of a company altering its annual dividends in the subsequent 2001 to 2006 testing period and compare the results from these forecasts to the dividend changes actually observed. Next, probability predictions are pooled and then ranked with the top 40 per cent of observations (where the model suggests that a dividend increase is most likely) assigned to a long portfolio, while the bottom 40 per cent (where the model indicates that a dividend increase is unlikely) are included in a short portfolio. The profitability of this investment strategy is finally examined by calculating market adjusted buy-and-hold returns for these two portfolios over holding periods of up to 24 months. In addition, we attempt to explain whether the characteristics of these two portfolios contribute to the performances which they achieve. The remainder of the paper proceeds as follows: section 2 provides a review of the relevant literature. Section 3 describes the data and methodology. The logit models generated and their prediction accuracies are explained in section 4. Section 5 presents the results relating to the returns generated by the investment strategy. The last section offers some conclusions.
The argument that past and present financial profiles influence the current dividend decision of a firm, and provide signals about its future profitability to the market, has a well established pedigree. Lintner (1956) was the first researcher to adopt a behavioural approach where US executives were asked about their perceptions of corporate dividend decisions. He found that the most influential determinant of company dividend policy was corporate earnings – both past and present. He developed a behavioural model which explained how companies partially adjusted their dividend payout ratios in the direction of a previously set target payout. Fama and Babiak (1968), who evaluated a number of alternative models, concluded that Lintner’s behavioural model performed well relative to its competitors and any change in a firm’s current dividend payout was a function of its target payout, current earnings, past earnings and past dividend payout. Survey-based studies conducted subsequent to Lintner’s pioneering work provide overwhelming support for the argument that corporate managers place a significant emphasis on the level of past and current earnings as well as on the variability of expected future earnings when they alter the existing dividend policy of the firm (see for example, Baker et al., 1985; Partington, 1989; Baker et al., 2001). This phenomenon of basing dividend decision on past earnings performance has been empirically observed by researchers. For example, Healey and Palepu (1988) found that the decision to initiate dividend payments was preceded by an improvement in earnings growth that started at least one year before the announcement while the decision to omit a dividend was preceded by a significant decline in earnings which started two years before the announcement date; these earning trends prevailed in the year of the dividend initiation or omission as well. Benartzi et al. (1997) corroborated Healey and Palepu’s (1988) findings using a sample of regular dividend changes. They found that the firms that increased (decreased) dividends experienced significant increases (decreases) in their earnings in the year before and the year of the announcement. A number of studies have uncovered evidence that the other elements of a company’s financial performance also influence the dividend policy of the firm. For example, Wansley and Lane (1987) found that the dividend initiating firms in their sample experienced a significant reduction in their debt levels in the years prior to the payment of their first dividend. In a number of survey- based studies, the maintenance of a target capital structure has been identified as an important determinant of the dividend policy by corporate managers (see Baker et al., 1985; Baker &Powell, 2000; Baker et al., 2001). In particular, Baker et al. (2001) found a NASDAQ firm’s debt-equity mix to be the sixth most important determinant of its dividend policy; companies tended not to pay large dividends relative to reported earnings if such dividends had to be financed by the issuance of new debt which might alter the existing target capital structure of the firm. They also found that highly levered firms that paid a high proportion of their earnings as interest were more likely to cut dividends than their low- geared counterparts. Lonie et al. (1992), who analysed the pressure exerted by interest rate rises on firms, found that the incidence of a dividend cut was more common among companies with ‘high’ interest-to-operating profit ratios compared to their peers with ‘low’ interest-to-operating profit ratios. The results of De Angelo and De Angelo (1990) corroborate this evidence; a minority of firms in their sample indicated that a rising level of interest expenses forced them to cut dividends, indicating that such firms used the cash saved from dividend reductions to service debt obligations. Large and mature companies tended to pay high and stable dividends while small, fast-growth firms maintained low payout ratios. Fox and Green’s (1992) study revealed that the members of FT-100 Index (FTSE) with large market capitalisations maintained a high dividend payout ratio of 50 per cent between 1984 and 1990 while the members of the Unlisted Securities Market (USM) with much smaller market capitalisations paid only 30 per cent of their earnings as dividends. An earlier study of Chowdhury and Miles (1987) found that small firms (i.e. total assets less than £181 million) were more likely to cut their dividend levels when faced with severe financial pressure than their larger-sized counterparts. Lintner (1956) observed that the management employed flexible standards on their firms’ current liquidity position in order to provide a buffer between the current investment programme of the firm and a more definite dividend policy. Darling (1957) found that Lintner’s behavioural model worked well during periods of improved firm liquidity and business sentiments; he documented that the availability of liquid assets was an important determinant of a firm’s capacity to pay dividends. Baker et al. (1985) agreed with this notion; in their study, managers ranked the availability of cash as the third most important determinant of their firms’ dividend policy. Wansley and Lane (1987) also observed that the competing demand for cash in their sample firms declined in the years prior to the initiation of dividends. Two empirical studies – one for the UK, the other for the US – analysed various financial profiles of dividend changing companies in the years prior to regular dividend changes. Lonie et al. (1996) analysed profitability, operating activity, gearing, liquidity and size measures of 617 UK firms during the six-year period prior to the announcement of changes in dividends and earnings. Their findings revealed that companies which increased both dividends and earnings reported a statistically significant higher level of profitability (represented by return on equity, return on capital employed and net profit margin) compared to those that announced a decrease in both dividends and earnings. Also, the companies that reduced dividends while reporting a reduction in earnings were found to have significantly higher leverage levels and interest/ operating profit ratios than their counterparts who reported increases in both dividends and earnings. The latter group demonstrated an extreme level of operational efficiency by holding stocks for fewer days and extracting more trade credit from suppliers than the former group. The dividend–increasing firms were much larger and more liquid than their dividend decreasing counterparts during the periods leading up to the change in their regular dividend payments. Jensen and Johnson (1995) analysed a sample of US firms which reduced their dividends. They found that the decision to cut dividends was accompanied by a deterioration in the overall financial performance of the firm; in the years prior to the dividend drop, a typical firm experienced a significant decrease in profitability, stock price, current assets, cash position, number of employees and new external financing and an increase in leverage and rising operating expenses.
On the basis of the evidence reported in relation to the influence of the past financial profiles on current dividend changes, we propose the following testable hypothesis:
Hypothesis 1: Past and present financial profiles can be used to predict one year-ahead changes in dividends.
The dividend signalling hypothesis asserts that the changes in a firm’s current dividend level convey information about the future performance of the firm. The theoretical models in the dividend signalling literature suggest that in a world of information asymmetry where managers have superior information about the current operations and future prospects of the firm compared to outside investors, the announcements of changes in current dividend levels convey information about the future payoffs from current investments. Accordingly, dividend increases are regarded as positive signals that convey favourable news to the market while dividend decreases are regarded as negative signals that convey unfavourable information (Bhattacharya, 1979 & 1980; John & Williams, 1985; Miller & Rock, 1985). The bulk of the supporting evidence for this hypothesis comes from the studies that have employed an event study methodology to observe the market reaction to dividend news during the period when a dividend change is announced to investors. The existing evidence suggests that dividend increases are associated with statistically significant positive abnormal returns while dividend decreases are associated with statistically significant negative abnormal returns (see, for US evidence, Pettit, 1972; Charest, 1978; Aharony & Swary, 1980; Woolridge, 1982; Asquith & Mullins, 1983; Brickley, 1983; Divecha & Morse, 1983; Benesh et al. 1984; Dielman & Oppenheimer, 1984; Eades et al., 1985; Wansley & Lane, 1987; Ghosh & Woolridge, 1988; Aharony et al., 1988; Healey & Palepu, 1988; Ghosh & Woolridge, 1991; & for UK evidence, Lonie et al., 1996; Gunasekarage & Power, 2001). The studies that have analysed the post-announcement financial performance of dividend changing companies, however, do not provide overwhelming support for the dividend signalling hypothesis. While there is some evidence to suggest that dividend–increasing firms perform well in the subsequent years, the evidence does not necessarily suggest that the performance of dividend–decreasing companies deteriorates during the same period. For example, Healey and Palepu (1988) analysed the earnings performance of firms after an initiation, and omission of dividends had taken place. They found that dividend–initiating firms demonstrated a sustained improvement in their earnings over subsequent years. However, dividend–omitting firms reported a rebound in their earnings and went on to report positive earnings in the two years after the omission was announced. Nissim and Ziv (2001) used a sample of ordinary dividend changes in their examination of the relationship between current dividend changes and the subsequent earnings performances of dividend changing firms. The authors found strong evidence that dividend changes were positively related to the future earnings of the firm - irrespective of whether the earnings performance was measured by the change in earnings, the absolute level of earnings or abnormal earnings – and thereby provided strong evidence in support of the prediction of the dividend–signalling hypothesis. On the basis of the dividend–signalling hypothesis and the related empirical evidence we argue that the dividend increasing companies will outperform their dividend–decreasing counterparts in the long run. Therefore, we propose the following hypothesis:
Hypothesis 2: Taking a long position in dividend increasing shares and a short position in dividend decreasing shares will generate abnormal returns for investors.
We employed data on New Zealand companies to examine the two hypotheses outlined in the previous section. Our sample included all the companies listed on the New Zealand Stock Exchange (NZX) for which the relevant accounting and share price data were available for analysis. The period under investigation spanned the years from 1995 to 2006 with the annual financial statement information sourced from the NZX Deep Archive Service. This online archive contains extensive details of company financial statement information. The monthly adjusted stock prices together with the market index were obtained from Datastream Advance Database. We used the six-year period 1995-2000 to develop a multiple logit model to predict the probability of the one-year-ahead dividend changes according to a stepwise procedure. Specifically, we identified significant elements in the companies’ financial profiles as measured by a total of 59 financial ratios for 261 firm-year observations during this calibration period. The multiple logit model with maximum explanatory power was then used to forecast one year-ahead dividend changes during the 2001 to 2006 testing period. To study hypothesis 1, we examined the accuracy of the dividend change prediction during the testing period. Hypothesis 2 was explored by analysing market excess returns for two investment portfolios formed on the basis of dividend change signals generated during the testing period. We assumed that we invested in a portfolio of shares with high probabilities of forecast dividend increases but take short positions in shares where the probabilities of dividend increases are low. In the remainder of this section, this analytical procedure is explained in detail. The method of calibrating the prediction model is outlined and then its forecasts described. This is followed by a description of the technique for determining market excess returns when testing hypothesis 2. We hypothesized that a multiple logit model could help us predict probabilities of one-year-ahead dividend changes. Explanatory variables were accounting ratios); 16 ratios were dropped purely due to the unavailability of complete data. It accounting ratios); 16 ratios were dropped purely due to the unavailability37of complete data was our objective "to let the data speak”; i.e. we did not make any conscious attempt to pick was our objective "to let the data speak”; i.e. we did not make any conscious attempt to p historic financial ratios for a firm. We started with 75 accounting ratios that suitable accounting descriptors ourselves but rather seek the broadest possible selection of suitable have been accounting discussed descriptors in text books andourselves but in employed rather priorseek the broadest research articles possible to selection capture the potential financial drivers profiles of dividend of a as changes firm. For example, a starting Ou model point for the and Penman (1989) development. As discussed potential used similar drivers ofdescriptors accounting dividend changes as a starting to predict point the future for thein changes model development. earnings per As discus in the literature share. However,review only 59section, previous of these studies suggest were included thatanalysis in the final firms use both present (Appendix 1 and past in the literature review section, previous studies suggest that firms use both present and p provides the definitions of these accounting ratios); 16 ratios were dropped purely financial theprofiles due to financial as the ofbasis unavailability for changing complete existing dividend “to levels. Accordingly, our profiles as the basis data. for It was ourexisting changing objective let the dividend data Accordingly, levels. speak”; i.e. we did not make any conscious attempt to pick suitable accounting explanatory variables became the three-year averages of each financial ratio. Therefore, we descriptors ourselves explanatory but rather variables seekthe became the three-year broadest possible averagesselection of each of potential financial ratio. Therefore, drivers ofthedividend changes as of a starting the ratio,point for the to bemodel development. As the end of averaged current observation assumed disclosed 3 months after averaged discussed in thetheliterature current observation of the previous review section, ratio, assumed to be studies disclosed suggest that3 firms months after the end useaccounting the both present and topast period, financial which profiles it relates, as the with the basis for ofchanging observations existing the previous two accounting the levels. dividend accounting period, to our Accordingly, which it relates, with explanatory the observations variables became theofthree-year the previous two account periods. averages of each financial ratio. Therefore, we averaged the current observation periods. of the ratio, assumed to be disclosed 3 months after the end of the accounting period, todependent The which it relates, variablewith in thethe observations model of the was a binary previous variable two which accounting took the value 0 for firms periods. The dependent variable in the model was a binary variable which took the value 0 for fir which decreased dividends The dependent and 1inforthefirms variable modelthatwas increased a binary dividends. variableWe did took which not use a naive which decreased dividends and 1 for firms that increased dividends. We did not use a na the value 0 for firms which decreased dividends and 1 for firms that increased definition dividends. ofWedividend did change not however, usechange a naive since observations definition with absolute dividend decreases are definition of dividend however, sinceofobservations dividend change however, with absolute dividend decreases since observations with absolute dividend decreases are relatively rare. Instead relatively rare. Instead we compared dividend changes observed relative to the drift in past relativelydividend we compared rare. Instead changes we observed compared relative dividendtochanges the driftobserved in past relative dividendto the drift in p dividend changes. If it is above (below) the drift, this binary variable Π becomes one changes. If it is above (below) the drift, this binary variable Π becomes one (zero). The (zero).dividend changes. The definition ofIfΠit is is shown above (below) the drift, [1] in the equation thisbelow. binary variable Π becomes one (zero). T definition of Π is shown in the equation [1] below. definition of Π is shown in the equation [1] below. 1 if DPS i,t DPS i,t 1 DPS i, t 2 DPS i,t 3 3 t 1 if DPS i,t DPS i,t 1 DPS i, t 2 DPS i,t 3 3 [1] if DPS i, t DPS i, t 1 DPS i, t 2 DPS i, t 3 3 0 t [1] [1] 0 if DPS i, t DPS i, t 1 DPS i, t 2 DPS i, t 3 3 where where DPSi ,t DPSi ,t DPSi ,t 1 is is the theabsolute absolutechange changeinin dividend perper dividend share for for share firm i in year t where DPSi ,t DPSi ,t DPSi1,t 1 is the absolute change in dividend per share for firm i in ye firm i in year t compared to year t-1 . compared Totoidentify year t-1financial . ratios with explanatory power during the 1995-2000 compared to year t-11. model development period, we first ran univariate logit models individually on all 59To ratios included identify in this analysis. financial ratios withUnless the ratio was explanatory power found to bethe during significant 1995-2000 model at least at theTo10identify per cent financial level, it wasratios with explanatory disregarded power during in the subsequent the 1995-2000 mo analyses. development period, The next we was step first ran univariatealllogit to include models ratios individually found to generateon all 59 ratios included in significant development period, we first ran univariate logit models individually on all 59 ratios included coefficients into a multivariate logit model. Using a stepwise procedure, non- this analysis.regressors significant Unless thewere ratio then was found to be significant eliminated at least at theof10regressors per cent level, it was this analysis. Unless the ratio was foundone bysignificant to be one. Removal at least at the 10 per cent level, it w stoped once all of the remaining variables were significant at the 20 per cent disregarded in the subsequent analyses. level ordisregarded if the goodness of fit as measured in the subsequent analyses.by the Schwarz information criterion deteriorated. The coefficients of this best-fit multivariate logit model were subsequently used to forecast dividend changes during the 2001 to 2006 testing period. The predicted probability of dividend changes (PR) for time t+1 based on observations at time t is shown in equation [2]. subsequently subsequentlyused usedtotoforecast forecastdividend dividendchanges changesduring duringthe the2001 2001toto2006 2006testing testingperiod. period.The The predicted probability of dividend changes (PR) for time t+1 based on ob predicted predictedprobability probabilityofofdividend dividendchanges changes(PR) (PR)for fortime timet+1 t+1based basedon onobservations observationsatattime timet tisis shown shown in The equation [2]. shownininequation equation[2]. [2].
1 1 PR PRi ,ti,t11 PR1 ( i,t 1 ... j j j ,tj),t)( 11,t 2 2 ,t ... j j ,t ) [2] [2] [2] 11ee( 111,t1,t 222 ,t2 ,t ...1 e where where XX1,t1,to where X1,tXto t to Xj ,tj ,Xisj,tisais avector X 1,tofto vector a vector ofjof jaccounting accounting j accounting variables variables atattime variables time t,t,and at time t, 1 1and and toto j j isto isaavector vectorofof j ,t is a vector of j accounting variables at time t, and 1 twhere X is a vector of coefficients from the multiple logit model for the 1995 to 2000 coefficients period.from coefficients To the fromtest multiple multiplelogit thehypothesis 1, model logit wemodel for forthe compared the1995 1995toforecasted these to2000 2000period. PR To period. Totest values hypothesis testto the hypothesis 1,1,we we ones actually observed coefficients in thefrom the multiple subsequent period. logit model For this purposefor we theconducted 1995 to 2000 period. To compared compared these theseforecasted a Chi-Squared Test toPR forecasted seevalues PR values whether totothetheones our ones model actually actually observed observed had superior ininthethesubsequent forecasting abilitiesperiod. subsequent period.For For beyond a model compared of randomthese dividendforecasted changes.PR values to the ones actually observed in the s this thispurpose purposeThe weweconducted PR valuesaaChi-Squared conducted Chi-Squared were then used Test Testtototorank see seewhether whether our ourmodel observations withhad model thesuperior had highest forecasting superior forecasting (lowest) PR values, this purpose being those we conducted with highesta (lowest) Chi-Squared Testfor likelihood to see whether our model ha upcoming abilities abilitiesbeyond beyond dividend aamodelmodelofFor increases. ofrandom random dividend the purposedividend ofchanges. changes. this analysis, we allocated the top 40 per cent of theabilities observations beyond to the a model expected ofdividend random increase dividendportfolio changes. and the The ThePR bottom PR40values per cent values were then to the were used usedtoto dividend then rank rankobservations decrease with portfolio; for observations both with theofhighest the these highest (lowest) (lowest)PR portfolios, PRvalues, values, we explored return performance for holding periods of up to 24 months. For being those beingthis those with withhighest purpose, highest we used The (lowest) (lowest)PR monthly values returns were likelihood likelihood for for then forupcoming upcoming the used companies toinrank dividend dividend the observations increases. increases. sample For Forthe and with thepurpose the purpose ofofhighes return on the market index to generate the market-adjusted buy-and-hold return this thisanalysis, followswe analysis, as we being those 2 allocated : allocated the thetopwith top40 40highest per percent cent (lowest) ofofthe likelihood observationsfor theobservations totoupcoming the theexpected expected dividend dividend dividendincreas increase increaseportfolio portfolioand the this bottom bottom40 theanalysis, 40per percent centtotoMthe thedividend top decrease portfolio; for forboth bothofofthese 1 N M we allocated the 40 per cent of the observations and dividend decrease portfolio; these to t MABHR P ,M 1 Ri ,t 1 Rm,t [3] [3] portfolios, portfolios,we weexplored explored N i 1 performance return return performance for t 1holding for holding periods periodsof of up up to to24 24 months. months. For For this this increase portfolio and the bottom 40 per cent to the dividend decrease portf t 1 purpose, purpose, we weused monthly used[3], monthly returns returnsfor for the the companies companiesininthethesample sampleand andthe return thereturn on onthe returnfor themarket market In equation In equation MABHR portfolios, [3], MABHRwe P , M explored isP,Mthe the return performance is market-adjusted market-adjustedbuy-and-hold for holding buy-and-hold periods return of up to for portfolio indexportfolio P from month 1 (which is the fourth month after the 2accounting 2 year indextotogenerate generatethe themarket-adjusted market-adjustedbuy-and-hold buy-and-holdreturn follows: : returnasasfollows end of a firm3)purpose, to month M; used Ri ,t ismonthly the returnreturns of firmfor i inthe month t; Rm ,t is the sample month 1 (which is we the fourth month after the accounting companies year end in of the and the a firm3) to month M return of the market index in month t and N is the number of observations in the portfolio. index to generate the market-adjusted buy-and-hold return as follows2: the return of firm i in month t; Rm ,t is the return of the market index in month t and N
4. The Logit in number of observations Model and its Prediction Accuracy the portfolio.
As mentioned in the previous section, we used the financial statement information
for the six-year period from 1995 to 2000 to develop the multiple logit models; 55 companies were used for this purpose with annual observations ranging As mentioned between in the previous section, we used the financial statement information for 34 and 52 firms. When the univariate logit models were estimated, 10 accounting year period descriptors from as emerged 1995 to 2000 to important develop because variables the multiple theylogit hadmodels; 55 companies were u significant coefficients at the 10 per cent level; six of these had coefficients that were this purpose significant at the 5 with annual per cent levelobservations ranging four while the remaining between were34 and 52 firms. significant at the 10 per cent level. Columns 3, 4 and 5 of Appendix 1 provide the results for the coefficient and When their significance the univariatelevel logitformodels each ratio. wereInestimated, addition, the 10final column descriptors eme accounting indicates whether or not a variable has been accepted or rejected for inclusion in important variables because they had significant coefficients at the 10 per cent level; six had coefficients that were significant at the 5 per cent level while the remaining fou the final model. The accounting ratios that emerged as the important measures in predicting the year-ahead dividend change encompassed a number of aspects of a firm’s financial profile (which have been found to influence the dividend policy of the firm). They included profitability, leverage, liquidity, operating activity, capital expenditure (assets base) and sales volume. These accounting descriptors were then included in the model using a stepwise procedure, in order to derive the final multiple logit model. This was then used to predict the direction of the one-year-ahead dividend change. Table 1 presents the coefficient estimates for the final logit model. Our dividend change prediction model contains the six financial ratios which capture four important financial attributes of the firm – i.e. profitability (represented by percentage change in net profit to EBITDA), liquidity (represented by percentage change in working capital), operating activity (represented by percentage change in inventory level and percentage change in sales to average total assets) and capital expenditure (represented by percentage change in depreciation and percentage change in average fixed assets and depreciation). Interestingly, three coefficients for these accounting variables are significant at the five per cent level; out of the remaining coefficients, two are significant at the 10 per cent level while the remaining one is significant at the 20 per cent level.
| Accounting Descriptor | Coefficient | z-statistic | p-value |
|---|---|---|---|
| Intercept | 1.1234 | 2.99 | 0.00 |
| %Δ in Inventory Level (absolute) | -2.0497 | -1.77 | 0.08 |
| %Δ in Depreciation | 0.4660 | 2.25 | 0.03 |
| %Δ in Average Fixed Assets & | |||
| Investments (excl. land) | -0.0395 | -1.90 | 0.06 |
| %Δ in Sales to Average Total | |||
| Average Assets | -1.0689 | -1.30 | 0.19 |
| %Δ in Net Profit to EBITDA | -39.3090 | -2.67 | 0.01 |
| %Δ in Working Capital | 0.0001 Note: This table reports the output for the final Logit regression model which was developed through a step-by-step process by dropping explanatory variables which failed to generate significant slope coefficients at each step. Selection criteria: p-value <= 20 per cent but stops if Schwarz information criteria no longer decreases. It reports the accounting ratios that entered into the final model, the slope coefficients of the explanatory variables, and the associated z-statistics and p-values. | 2.00 | 0.05 |
| Panel A: Observed Dividend Changes (2001-2006) | |||
|---|---|---|---|
| No. of observed dividend changes | 338 | ||
| No. of dividend increases | 184 | ||
| No. of dividend decreases | 154 | ||
| % actual DPS increasing firms | 54.4% | ||
| % actual DPS decreasing firms | 45.6% | ||
| Panel B: Predicted Dividend Changes | |||
| Calibration period | 1995-2000 | ||
| Prediction period | 2001-2006 | ||
| No. of accounting variables in logit model | Selection Rule: Top 40% = Dividend Increase; Bottom 40% = Dividend decrease | 6 | |
| Total no. of signals generated | 337 | ||
| of which | |||
| No. of dividend increase signals generated (top 40%) | 135 | ||
| No. of dividend decrease signals generated (bottom 40%) | 135 | ||
| No. of inconclusive signals generated (middle 20%) | 67 | ||
| Panel C: Accuracy of Dividend Change Predictions | |||
| Total no. of signals for which we can check accuracy (i.e. | 198 | ||
| they are matched with subsequent dividend observations) | (continued) | ||
| Details: | Increases | No. of Dividends Decreases | Total |
| Cases | 111 | 87 | 198 |
| % correct predictions | 60.0% Chi-squared (d.f. 1)/p-value | 46.9% | 52.5% 3.83*/0.05 |
| Notes: | The table reports information relating to the prediction accuracy of the multiple logit model. The * indicates statistical significance at the 10 per cent level. The following calibration parameters were | ||
| applied in the prediction model: | Unilogit models: Variable selection is based on the rule p-value <= 10 per cent. An explanatory accounting variable must have at least 75 per cent of the observations. Multilogit model: Variable selection is based on the step-by-step process and the rule of p-value | ||
| <= 20 per cent. | The next step was to test the accuracy of the dividend forecasts. We matched the dividend predictions with their corresponding actual dividend change observations. However, subsequent dividend information was not available for 72 observations and these observations were disregarded for the purpose of this comparison. The evaluation of prediction accuracy was thus based on a final sample of 198 matched observations with the results shown in Panel C of Table 2. The overall prediction accuracy rate was 52.5 per cent. As the chi-squared statistic reveals, this prediction accuracy is significant at the 10 per cent level. However, our logit model seemed to be able to classify dividend increases more accurately than dividend decreases; this model predicted 60.0 per cent (46.9 per cent) DPS increases (DPS decreases) correctly. These prediction accuracy rates are in line with the actual dividend changes observed for this | ||
| market during the 2001-2006 period. | |||
Before showing the returns for our own strategy, we first present the excess return generated for the perfect foresight strategy, i.e. the return earned by an investor if he/she knew the direction of the actual dividend change before that news was announced and took a long position in dividend increasing firms and a short position in dividend decreasing firms. As the previous section highlighted, during this six-year period there were 338 dividend change observations; of these, 184 were related to dividend increases and 154 involved dividend decreases. We calculated the market-adjusted buy-and-hold return for a number of holding periods for these two groups. The results are presented in Table 3, Panel A. According to the statistics reported in this table, dividend increasing companies reported gradually increasing positive returns across the different holding periods examined. Even though the dividend decreasing firms reported negative returns across all the holding periods analysed, a decreasing trend in negative returns could be observed for this category.
| and Dividend Prediction Model Based Strategy (2001-2006 Investment Period) | |||||
|---|---|---|---|---|---|
| Portfolio | No. of | Investment Horizon | |||
| Firms | 6 Months | 12 Months | 18 Months | 24 Months | |
| Panel A: Perfect Foresight Strategy | |||||
| Observed Increase | 184 | 2.74% | 4.68% | 4.76% | 7.11% |
| Observed Decrease | 154 | -3.47% | -3.10% | -2.62% | -0.59% |
| Strategy | Panel B: Dividend Prediction Model-based Strategy | 6.22% | 7.78% | 7.38% | 7.70% |
| Long Position | 136 | -2.61% | -1.30% | -4.56% | -4.98% |
| Short Position | 135 | -1.75% | -2.31% | -4.02% | -2.65% |
| Strategy | – dividend increasing companies outperformed in the market subsequent to the announcement of a dividend change while dividend decreasing firms when the dividends for the subsequent years became known to the market. | -0.86% Notes: Panel A assumes perfect foresight and reports returns of shares which turned out to be dividend increasers and dividend decreasers in an accounting period. Investment takes place 3 months after the end of the accounting period. The strategy return is the outcome of taking a long position in the ‘dividend increase’ portfolio and a short position of the ‘dividend decrease’ portfolio. Panel B shows buy– and– hold returns for portfolios of firms predicted to increase, respectively decrease dividends. Predictions are based on a multiple logit model using past company financial ratios. The model was calibrated during 1995 to 2000 and then used to forecast dividend increasing firms during the 2001 to 2006 investment period. Investment takes place 3 months after the end of the accounting period preceding the year of the predicted dividend; at this time company financial statements would be available to investors. Strategy returns are calculated as in panel A. On average, the evidence supports the dividend signalling hypothesis underperformed. Therefore, the strategy of investing in dividend increasing companies and short–selling stocks in dividend decreasing firms generated positive market adjusted returns for investors; for a 24-month holding period, for instance, this strategy earned a market adjusted return of 7.70 per cent. Therefore, if the information about the direction of future dividend changes had been available to investors, they could have earned positive market-adjusted returns by implementing this strategy. The annualised returns generated by this strategy for different holding periods were as follows: for a six-month holding period 12.43 per cent, for a 12-month holding period 7.78 per cent, for an 18-month holding period 4.92 per cent, and for a 24-month holding period 3.85 per cent. In this study we found an optimal holding period of 12 months when the holding-period returns reached a peak. One reason for the decline in returns after 12 months for the strategy might be the loss of any information advantage | 1.01% | -0.54% | -2.34% |
We now analyse the market adjusted returns for a strategy based on our dividend change prediction model. In contrast to the evidence reported in Panel A, the statistics in Panel B reveal that the strategy did not produce any significant returns for investors. The return to the strategy demonstrated an inconsistent pattern as the holding period extended for up to 24 months. It generated a negative return of 0.86 per cent for a six–month holding period; this increased to a positive return of 1.01 per cent for a 12-month holding period and subsequently declined to a -0.54 per cent for an 18-month holding. An investor who implemented this dividend prediction– based strategy for a 24-month holding period would have earned a negative return of 2.34 per cent (or an annualized return of -1.17 per cent). The return behaviour of the dividend decreasing firms in Panel B is in agreement with the dividend signalling hypothesis. These firms reported negative market- adjusted returns for all the holding periods analysed where these negative returns increased for a period of 18 months. However, dividend increasing firms did not report positive returns as postulated by the dividend signalling hypothesis. Their returns were negative for all the holding periods analysed and, on average, these negative returns demonstrated an increasing trend as the length of the holding period expanded. The general conclusion that emerges from this analysis is that, even though the multiple logit model is able to classify companies with some accuracy as dividend increasing firms and dividend decreasing firms based on their past and present financial profiles, a strategy of investing in firms which were predicted to be dividend increasers and short selling stocks of firms which were predicted to be dividend decreasers proved to be unprofitable. In our sample, such a strategy generated negative market adjusted returns. This may indicate that the market is semi-strong form efficient; the investors had interpreted the dividend prediction based on financial profiles correctly and impounded that information into share prices in a quick and unbiased manner so that a trading strategy based on such predictions did generate abnormal returns for investors.
In order to identify potential cycles in excess returns, we analysed the above portfolio returns by year of investment. This means the cumulative excess returns, say for 2001, is the return of an investment portfolio formed 3 months after the end of the 2000 financial year and then held for 24 months. The results of this analysis are shown in Figure 1. It shows extreme volatilities in the outcomes to investors in each year of investment. In 2003, for example, the strategy outperformed with an excess return of 11.4 per cent for the long portfolio, -3.8 per cent for the short portfolio, resulting in a strategy return of 15.2 per cent. In the other years, however, the results are mixed; in only three of theresults six years (i.e. in2003, are mixed; 2005; only three 2006) of the did the six years investor (i.e. 2003, 2005;achieve 2006) didpositive returns the investor achievefor the long/short strategy. positive returns for the long/short strategy.
20% 24–month Cumulative Excess Returns 24 month Cumulative Excess Returns
15% LONG
0% SHORT
-10% Strategy (=LONG - SHORT) -15%
| istics introduced | returns; |
|---|---|
| earlier this relationship | |
| in first | |
| this section. | is positive, however. Such evidence contradicts |
| with the We well known estimated | |
| size effect whichunivariate | predictsregressions and the between a negative relationship results are reporte |
coefficients. High book-to-market and large firms tended to earn
| Constant | Pr | BETA | B/M | E/P | LOGMV | CRR12 |
|---|---|---|---|---|---|---|
| (t-stat) | (t -stat) | (t -stat) | (t -stat) | (t -stat) | (t -stat) | (t -stat) |
| 0.0300 | -0.1380 | |||||
| (0.46) | (-1.15) | |||||
| 0.0151 | -0.0925 | |||||
| (0.32) | (-1.36) | |||||
| -0.0754* | 0.4532 | |||||
| (-1.91) | (1.52) | |||||
| -0.0372 | -0.0115 | |||||
| (-1.25) | (-1.06) | |||||
| -1.1999*** | 0.0619*** | |||||
| (-4.19) | (4.09) | |||||
| -0.0384 | 0.3697 | |||||
| (-1.27) | (0.48) | |||||
| -1.3069*** | -0.1328 | -0.1465** | 0.6628** | -0.0216** | 0.0719*** | 0.72314 |
| (-4.31) | (-1.10) | (-2.18) significance at the 10 per cent, 5 per cent and 1 per cent levels respectively. | (2.23) Note: The table reports coefficient estimates and corresponding t-statistics in parentheses when the 24-month market-adjusted buy-and-hold return is regressed on a set of independent variables as defined in equation [4]. The independent variables considered in the estimation of the model include the dividend increase probability (PR) of the stock as defined in equation [2], firm beta (BETA), book-to-market value (B/M), earnings-to-price ratio (E/P), natural logarithm of market value (LOGMV) and the 12-month cumulative raw return (CRR12). The *, (**), (***) indicate statistical | (-1.96) | (4.63) | (0.94) |
As a robustness test, we used alternative cut-off points to allocate companies into the predicted dividend increase group of firms and the predicted dividend decrease group of firms. As one example, we present here the results for a 30/30 portfolio, i.e. the top 30 per cent of observations were allocated to the former portfolio (long position) and the bottom 30 per cent combined to form the latter portfolio (short position). This process allocated 101 firms into each portfolio during this six-year period with a prediction accuracy of 55 per cent. This prediction accuracy was significant at just below 10 per cent level, i.e. slightly lower than the significance of the 40/40 portfolio presented earlier. Appendix 2 reports the buy–and–hold returns generated by these portfolios. According to this table, both portfolios generate negative market-adjusted buy–and–hold returns for all the holding periods analysed. On average, this negative return increases as the holding period expands. For a 24 month holding period, our strategy with these new cut-off points generates -7.59 per cent for investors. Again, even though the logit model has some moderate power explaining future dividend changing companies, investors are unable to translate such knowledge into a profitable investment strategy.
Overall, the findings of this paper suggest that although past financial statement information may help investors to forecast future dividend changes (especially dividend increases), any attempt to exploit these predictions using a trading strategy is unsuccessful; buying shares in companies where the probability of a dividend increase is high and selling shares–where the probability of a dividend increase is low yield negative abnormal returns. One implication of these results is that the New Zealand stock market seems to be semi-strong form efficient in that attempts to exploit dividend forecasts based on publically available financial statement information does not offer risk-adjusted profits. This result is robust to variations in the cut-off used to identify those firms where a dividend increase (decrease) is most likely. However, an analysis over a longer time span and in different markets is needed before firm conclusions can be reached about profitability of a trading strategy based on a dividend prediction model.
Author Information:
Abeyratna Gunasekarage is a senior lecturer in finance at the Department of Banking and Finance of Monash University, Australia. Prior to joining Monash university he has worked at the University of Canterbury and the University of Waikato in New Zealand. Abey received his BSc Honours from the University of Sri Jayewardenepura, Sri Lanka and MAcc and PhD from the University of Dundee, UK. He has published in international journals in the areas such as financial communication, performance of investment strategies, efficiency of emerging capital markets and corporate governance. Kurt Hess is a Senior Credit Advisor at the Independent Credit View, Zurich, Switzerland. Besides his applied financial consulting work, Kurt had been lecturing and researching as a senior fellow in Finance at the University of Waikato, New Zealand. Prior to this academic career Kurt was a director in Fixed Income and Credit Research at Credit Suisse First Boston. Kurt Hess holds a Masters in Industrial Engineering from ETH (Zurich, Switzerland), a PhD in Economics from the University of Waikato as well as an MBA from the University of British Columbia (Vancouver, Canada).
David Power is Professor of Business Finance at the Department of Accounting and Finance of the University of Dundee, UK. He is a graduate on the National University of Ireland (BCom), the London School of Economics (MSc Econ) and the University of Dundee (PhD). He has published widely in the areas of emerging stock markets, market efficiency and the communication between companies and the financial markets
Appendix 1: Univariate Model Parameters
(continued)
Short Note: Position The table shows buy–and–hold101 -1.45% returns for portfolios of firms predicted to-1.76% increase, respectively -4.43%
Strategy -3.19% -1.26% decrease dividends. Predictions are based on a multiple logit model using past company financial -4.07% ratios. The model was calibrated during 1995 to 2000 and then used to forecast dividend increasing Note: firms The during thetable shows 2001 to buy–and–hold 2006 investment period. returns Investmentfortakes portfolios of firms place 3 months afterpredicted the end of to increase, re the accounting period preceding the year of the predicted dividend; at this time companycompany dividends. Predictions are based on a multiple logit model using past financial financial rat statements would be available to investors. The strategy return is the outcome of taking a long firms during calibrated during 1995 to 2000 and then used to forecast dividend increasing investment period. Investment takes place 3 months after the end of the accounting period prece position in the ‘dividend increase’ portfolio (shares in top 30% of dividend increase predictions) predicted dividend; at this time company financial statements would be available to investors. T and a short position of the ‘dividend decrease’ portfolio (shares in bottom 30% of dividend increase the outcome of taking a long position in the ‘dividend increase’ portfolio (shares in top 30% o predictions). predictions) and a short position of the ‘dividend decrease’ portfolio (shares in bottom 30% o predictions).
End Notes
1Note that this definition implies that no Π value is defined for Note that this definition implies that no Π value is DPS i,t DPS i,t 1 DPS i,t 2 DPS i,t 3 3 -- i.e. where i.e. where current dividen current dividend to the past three changes areofequal years’ drift to the past three years’ drift of changes. changes. 2 2 IfIfa acompany companyis is delisted during delisted duringa holding period, a holding suchsuch period, a company a company is assigned is assigned zero mon remainder zero monthly ofreturns the holding period. for the remainder of the holding period. 3 3 We keep We keep a gap of of three threemonths monthsininorder to to order allow for for allow the possible time time the possible gap between gap the end of a firm and between the the end publication of itsyear of the financial annual of areports. firm and Conover et al. (2008), the publication of itswho analysed financia annual sampleConover reports. of 22 industrialised et al. (2008),nations, reported financial who analysed that during the 1992-96 reporting lag forperiod, the New Zealand f a sample of 22 industrialised nations, reported that during the 1992-96 period, the New statements. took a median number of 87.5 days (nearly three months) to publish their financial Zealand firms in their sample took a median number of 87.5 days (nearly three months) to publish their financial statements.