Sie sind auf Seite 1von 7

International Journal of Forecasting 19 (2003) 743749 www.elsevier.

com / locate / ijforecast

Just-in-time inventory systems innovation and the predictability of earnings


a, b c d Thomas A. Carnes *, Jefferson P. Jones , Timothy B. Biggart , Katherine J. Barker
a

University of Arkansas, Department of Accounting, Sam M. Walton College of Business, Fayetteville, AR 72701 1201, USA b Auburn University, Auburn, AL, USA c University of North Carolina, Greensboro, NC, USA d Lander University, Greenwood, SC, USA

Abstract Firms that adopt just-in-time (JIT) inventory practices do so in order to realize cost savings and improve product quality, but an unexpected benet to such rms could be a more predictable earnings stream. We examine the relationship between implementation of just-in-time inventory practices and the predictability of future quarterly earnings for a matched-pair sample of 82 rms, half of which have publicly announced that they have adopted JIT inventory practices. We nd that one- and four-step-ahead forecasts of quarterly earnings, using either a BrownRozeff [Journal of Accounting Research (1979) 179189] ARIMA or a seasonal random walk expectation model, are more accurate for the rms that have adopted JIT. 2002 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved. Keywords: Just-in-time; Inventory utilization; Quarterly earnings forecasts

1. Introduction Firms that implement just-in-time inventory practices typically expect benets such as reduced inventory costs, heightened quality awareness, a leaner manufacturing system, and greater productivity. Prior research, such as Huson and Nanda (1995), has found that JIT adoption leads to an increase in earnings per share and inventory turnover and a decrease in labor per sales dollar. However, an additionalthough possibly unexpectedpotential benet from successful JIT adoption is a smoother,

more predictable earnings stream.1 We examine this empirically by determining whether a sample of rms that have adopted JIT inventory practices has more predictable one- and four-step-ahead quarterly earnings than a control sample of similar non-JIT adopters. Prior research in accounting and nance has found multiple benets for rms with smoother earnings streams. As discussed by Hunt, Moyer, and Shevlin (1996), rms that do not smooth earnings may incur larger tax, debt and implicit claims costs. There
1 We describe this as unexpected since our examination of the public announcements of JIT adoption by over 300 rms nds no mention of this benet. In contrast, many rms cite expected cost reductions and quality improvements.

*Corresponding author. Tel.: 11-479-575-4117. E-mail address: carnes@walton.uark.edu (T.A. Carnes).

0169-2070 / 02 / $ see front matter 2002 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved. doi:10.1016/S0169-2070(02)00079-1

744

T. A. Carnes et al. / International Journal of Forecasting 19 (2003) 743749

appears to be a direct market reaction to a smoother earnings stream as well; Lipe (1990) nds that the earnings response coefcientthe coefcient that measures the stock-price response to a one-dollar earnings shockis positively related to both the predictability of earnings series and the time-series persistence of earnings. Inventory costs are one of the more volatile, unpredictable costs for many manufacturing rms (Oglove, 1987). Typically, rms that adopt JIT expect total inventory costs to be reduced. Reducing such costs should result in a more predictable, smoother earnings stream, and rms that successfully adopt JIT practices may enjoy these benets. We test our hypothesis that JIT adopters have more predictable earnings than non-JIT adopters by comparing the accuracy of quarterly EPS forecasts, employing a seasonal random walk and a Brown and Rozeff (1979) ARIMA model as our expectation models, for a sample of 41 publicly acknowledged JIT adopters and 41 non-JIT adopters, matched by size and industry. We nd that regardless of the expectation model, the JIT adopters have more predictable earnings than non-JIT adopters, as determined through comparison of the mean absolute percentage errors for one- and four-step-ahead predictions of earnings per share. We therefore conclude that JIT adopters in our sample enjoy a smoother earnings stream and may reap the benets thereof, as described in prior research. Empirical evidence is mixed regarding the nancial benets of JIT adoption. Huson and Nanda (1995) nd that rms implementing JIT had signicant increases in EPS and inventory turnover and signicant reductions in interest expenses and other nancing costs needed to service inventory and in labor per sales dollar. Balakrishnan, Linsmeier, and Venkatachalam (1996) nd ROA declines in their sample, but there is no signicant difference in ROA changes between 46 adopters and a matched set of non-JIT adopters. However, the ROA performance of JIT adopters free from signicant customer inuence does exceed that of matched non-JIT adopters. Adoption of JIT inventory practices, and the consequent lower inventory levels, may have a positive effect upon rms quality of earnings. Inventory levels can be adjusted in order to manipulate net income. As Palepu, Healy, and Bernard

(2000, p. 712), point out, (T)he valuation of the cost of sales provides several opportunities for managers to exercise judgment in nancial reporting. Managers can select the inventory ow method, increase or reduce production to allocate capacity costs over more or less units, or, if the rm uses LIFO, deplete inventory to match old costs against revenues. In the most extreme case, inventory levels can be fraudulently overstated in order to increase net income, as documented in Beasley, Carcello, and Hermanson (1998) in their examination of SEC fraud investigations between 1987 and 1997. Such adjustments are more difcult for JIT rms, due to the lower inventory levels on hand. Consequently, our hypothesis, stated in the alternative form, is as follows: H0 : Firms that have adopted JIT inventory practices will have more predictable one-step (four-step) ahead quarterly EPS than a matched-pair sample of non-JIT adopters. The remainder of this paper is organized as follows. We describe our data and research design in Section 2, and we report our results in Section 3. In Section 4, we discuss our conclusions and implications for future research.

2. Data and research design

2.1. Sample selection


We obtained an initial sample of rms that publicly disclosed a JIT adoption between 1975 and 1995 from Mead Datas LexisNexis database. We searched the PR Newswire and Major Newspaper les for the keyword JIT as well as several synonyms for JIT: Zero Inventory Production System (ZIPS); Minimum Inventory Production System (MIPS); kanban production; kaizan production; and Quick Response Inventory Systems (QR). We repeated this search pattern using the National Automated Retrieval System (NAARS) database and using the Wall Street Journal Index subject listing for the categories inventory and inventory management. In addition, rms from previous JIT research (Moras & Dieck, 1992; Billesbach & Hayen, 1994) were added to the sample, which resulted in the identication of approximately 530 rms.

T. A. Carnes et al. / International Journal of Forecasting 19 (2003) 743749

745

Our study design required that sample rms have publicly available nancial statements. Of the rms identied, 238 met this requirement. We deleted 126 of those rms since they adopted JIT after 1990, which meant there would not be a sufcient time series of data available on Standard and Poors 1999 Research Insight to perform the ARIMA estimation and the subsequent forecasts. We also deleted any rms with an SIC Code not between 2000 and 3999, thereby restricting our study to manufacturing rms. This left us with 94 rms. We were forced to eliminate 28 of those rms because there was not an uninterrupted time series of at least 36 quarterly EPS observations available for them on Compustats Research Insight (a minimum of 30 to estimate the ARIMA model and a six-quarter holdout sample for forecasting). We matched the remaining 66 rms that were JIT adopters based on 2-digit SIC code and 10-year average total assets (19901999)making sure that none of the control rms were in our original sample of JIT adopters.2 This left a matchedpair sample of 41 JIT adopters and 41 non-JIT adopters. This sample size is comparable to that of similar studies using public announcement of JIT adoption such as Balakrishnan et al. (1996) (46 matched pairs), and Huson and Nanda (1995) (55 rms). The majority of our sample comes from SIC codes 3500 (computer equipment, 26.8 percent of sample), 3600 (electrical, other electric equipment, 24.3 percent of sample) and 3700 (transportation equipment, 19.5 percent of sample). As noted earlier, the cost of liquidating obsolete inventory can be signicant in the computer and electrical industries. It is therefore unsurprising that such rms are strongly represented in our sample of JIT adopters. We searched the previously mentioned databases and veried that our sample of non-JIT adopters did not publicly announce JIT adoption during the 1975 1995 period, but we may have JIT rms in the control sample, since JIT adoption disclosure is voluntary. In order to minimize such contamination of our control sample, we contacted approximately half the rms in the control sample, selected at
2

random, and found no rms that had adopted JIT at the relevant time. Should our sample be contaminated in this manner, it should reduce the power of our tests and make it less likely that we could reject our null hypothesis.

2.2. Research design


We considered ve different expectation models for quarterly EPS. The rst two, a seasonal random walk (SRW) and seasonal autoregressive model (SAR), are nave statistical models. The other three models employed are parsimonious ARIMA models that prior accounting research has shown outperform the SRW and SAR models and rm-specic ARIMA models in predicting one-step-ahead quarterly earnings. All ARIMA models can be described in ( p,d,q) 3 (P,D,Q) notation, where p and P are the order of the autoregressive polynomial; q and Q are the order of the moving average polynomial; and d and D represent the degrees of ordinary and seasonal differencing, respectively, necessary to achieve stationarity. The three ARIMA models we consider are the Foster (1977) model, a (1,0,0)3(0,1,0) model with a constant; the Brown and Rozeff (1979) model, a (1,0,0)3(0,1,1) model; and the Grifn (1977) and Watts (1975) model, a (0,1,1)3(0,1,1) model. We rst estimated rm-specic parameters, employing at least 30 quarters of data for each rm, for each of the ve expectation models. We then performed goodness-of-t diagnostic checking of the ve models using the mean Q-statistic in order to determine which models had the minimum residual autocorrelation.3 This allowed us to determine which of the models best described the EPS series. The preferred model is one whose residuals behave as a white-noise series, so the lower the Q-statistic, the better the t. After determining which models best described the underlying data, we generated one- and four-step-ahead quarterly EPS forecasts employing those models. The one-step-ahead forecasts were generated for a six-quarter holdout period that ended with the second quarter of 1999. The four-step-ahead forecasts were generated for a three-quarter holdout

Matching rms on size is critical, since Bathke, Lorek, & Willinger (1989) demonstrate a pervasive rm-size effect upon predictive ability of the three ARIMA models we employ and the seasonal random walk.

We used 12 values of the residual autocorrelation to compute Q.

746

T. A. Carnes et al. / International Journal of Forecasting 19 (2003) 743749 Table 1 Summary of LjungBox Q-statistics Model BrownRozeff (1,0,0)3(0,1,1) Seasonal random walk Seasonal autoregressive Foster (1,0,0)3(0,1,0) GrifnWatts (0,1,1)3(0,1,1) JIT 9.948 42.199 78.850 14.767 10.210 P-Value 0.650 ,0.0001 ,0.0001 0.310 0.560 Non-JIT 8.552 53.807 84.269 16.506 11.124 P-Value 0.735 ,0.001 ,0.001 0.190 0.470

period ending at the same time. As is typical in accounting time-series research, the parameters of the model were re-estimated before each prediction in order to incorporate the information from the previous holdout period. We used the mean absolute percentage error (MAPE) to assess the accuracy of the predictions, calculated as follows: 4 1 MAPE 5 ] N Q 2 E(Q OU]]]]) U Q
1 t t N t

(1)

Where N represents the total number of rmquarters in the sample; Q t is the actual value of EPS for rm n in quarter t; and E(Q t ) is the forecast EPS in quarter t as generated by each forecast model. As is common in time-series research, values of the APE above 100 percent were truncated at 100 percent in order to reduce the effects of outliers.

P-Values are computed using a Chi-square test.

3. Results As discussed above, we estimate rm-specic parameters for each of the ve expectation models and perform goodness-of-t diagnostic checking of the ve models using the mean Q-statistic in order to determine which models have the minimum residual autocorrelation. The mean Q-statistics for all models are reported in Table 1. The P-values are computed using a Chi-square test. The Brown and Rozeff (1979) ARIMA model has the minimum residual autocorrelation of all models tested, with Q-values signicantly lower than those for the Foster (1977), Grifn (1977) and Watts (1975) models. We therefore employ it to make one- and four-step ahead predictions of quarterly earnings per share for our tests of our hypothesis. As a form of sensitivity analysis, we also test our hypothesis using the seasonal random walk model, the nave statistical model with the minimum residual autocorrelation. We report our one-quarter-ahead predictive results

in Table 2, panel A. We compare MAPEs on a quarter-by-quarter and a pooled basis, employing a one-tailed t-test, in order to determine if they are signicantly lower for the JIT adopters than for non-JIT adopters. For the BrownRozeff model, the most accurate model in describing our data for the 30-quarter estimation period, the one-step-ahead EPS predictions are signicantly more accurate for the JIT adopters for the pooled sample (P,0.01) and for three of the six quarters of the holdout period (quarters 1, 2 and 4), enabling us to accept our hypothesis as stated in these case.5 One-step-ahead predictions are signicantly more accurate (P50.07 for the pooled sample) for the JIT adopters on a pooled basis when using the SRW model, as well. However, when using the SRW, one-step ahead predictions are not signicantly more accurate for any of the six quarters of the holdout period. Our evidence regarding predictions of one-quarter-ahead EPS for JIT rms is therefore much stronger when using the BrownRozeff model. For the pooled sample, we also computed the mean forecast error, as discussed in Bathke et al. (1989), as a measure of bias. The mean forecast error is calculated as follows:
5

We calculated all results for the Mean Squared Error, as well, but in the interest of brevity, we report results only for the MAPE. Results for the MSE do not affect our conclusions and can be obtained from the corresponding author.

As noted, the holdout period runs from the rst quarter of 1998 to the second quarter of 1999; therefore, the BrownRozeff predictions are signicantly more accurate for the JIT sample in the rst quarter of 1998, the second quarter of 1998, and the fourth quarter of 1998.

T. A. Carnes et al. / International Journal of Forecasting 19 (2003) 743749 Table 2 Panel A Mean Absolute Percentage Error (MAPE) of one-step ahead EPS predictions Holdout 1 2 3 4 period BrownRozeff ARIMA model JIT 0.301 Non-JIT 0.469 P-Value 0.045 Seasonal Random Walk model JIT 0.380 Non-JIT 0.317 P-Value 0.254

747

Pooled

MFE

0.226 0.386 0.038

0.355 0.452 0.120

0.404 0.553 0.43

0.373 0.426 0.272

0.292 0.357 0.203

0.326 0.441 0.001

0.080 0.192 0.010

0.315 0.402 0.171

0.384 0.448 0.226

0.433 0.482 0.299

0.466 0.523 0.239

0.409 0.400 0.454

0.398 0.429 0.066

0.072 0.211 0.002

Panel B Mean Absolute Percentage Error (MAPE) of four-step ahead EPS predictions Holdout 1 2 3 Pooled period BrownRozeff ARIMA model JIT Non-JIT P-Value 0.444 0.629 0.015 0.486 0.453 0.357 0.389 0.479 0.139 0.440 0.521 0.050

MFE

0.075 0.172 0.090

Seasonal Random Walk model JIT Non-JIT P-Value 0.444 0.482 0.340 0.477 0.523 0.283 0.419 0.400 0.410 0.446 0.468 0.311 0.035 0.206 0.011

All P-values are comparisons of JIT and non-JIT results, employing one-tail t-tests. In panel A, we generate forecasts one quarter ahead, beginning with the rst quarter of 1998. In panel B, we generate forecast four quarters ahead, beginning with the fourth quarter of 1998.

1 MFE 5 ] N

Q 2 E(Q OS]]]]) D Q
N t t 1 t

(2)

The MFE allows forecast errors of different signs to cancel, thereby providing information on the directionality of the forecast errors. The MFE for the JIT adopters is signicantly lower than for the nonJIT adopters, at P50.01 for the BrownRozeff model and P,0.01 for the SRW. Again, this conrms our hypothesis that one-quarter-ahead predictions of EPS are more accurate for JIT adopters than for non-JIT adopters. We report our four-quarter-ahead predictive results in Table 2, panel B, comparing MAPEs in the same manner as for one-quarter-ahead predictions. For the

JIT adopters in the pooled sample, the MAPE is signicantly smaller for the BrownRozeff model (P50.05), enabling us again to reject our null hypothesis of no difference in prediction accuracy. When examining the BrownRozeff prediction errors on a quarterly basis, those for the JIT adopters were signicantly smaller only in the fourth quarter of 1998. Prediction errors are smaller for the JIT adopters when predictions are made using the SRW, but the difference is not statistically signicant (P5 0.31). It is not surprising that our results are weaker for the four-quarter-ahead predictions than for the one-quarter-ahead predictions, as prior research (summarized in Brown, 1993) has shown statisticalbased forecasting models have increased accuracy over shorter horizons, since confounding events

748

T. A. Carnes et al. / International Journal of Forecasting 19 (2003) 743749

make the four-quarter-ahead predictions signicantly less accurate. However, it is important to point out that for the most accurate forecasting model, our hypothesis is strongly conrmed on a pooled basis. When comparing MFE on a pooled basis, the four-quarter-ahead predictions are marginally more accurate for the BrownRozeff model (P50.09) and signicantly more accurate for the SRW model (P5 0.01).

4. Conclusions We examine whether rms that have adopted justin-time inventory practices have a more predictable quarterly earnings stream than non-JIT adopters. Firms typically adopt JIT practices in order to reduce inventory costs (e.g. holding and obsolescence costs), eliminate many types of waste, and heighten quality awareness. We hypothesize that an additional, perhaps unexpected, benet to JIT adopters is a smoother, more predictable earnings stream. We compare the accuracy of one- and four-quarter-ahead predictions of EPS, using the Brown and Rozeff (1979) ARIMA model and the seasonal random walk as our expectation models. We make predictions for a set of 41 publicly acknowledged JIT adopters and compare them to 41 non-JIT adopters matched by industry and size, comparing the mean absolute percentage errors and mean squared errors as measures of prediction accuracy. We nd that for the BrownRozeff model, both one- and four-quarter-ahead EPS predictions are signicantly more accurate for the JIT adopters. For the seasonal random walk model, one-quarter-ahead EPS predictions are more accurate for the JIT adopters; there is no statistically signicant difference in four-quarterahead predictions using the seasonal random walk. Our ndings demonstrate that JIT adopters receive the possibly unexpected benet of a smoother earnings steama benet that prior research has shown may lower the expected costs of dealing with shareholders and strengthen the relationship between accounting earnings and market returns. There are future implications of these results for both business managers and accounting researchers. For managers, the possibility of an operational innovation such as JIT increasing earnings predictability needs to be

included in the cost / benet decision model that managers use to make innovation adoption decisions such as JIT, Total Quality and ISO 9000 programs, and lean production methodologies. For researchers, the possibility that specic operational innovations can affect the behavior of earnings creates the opportunity to link nancial accounting research with operations management. Possible extensions of this research include investigation of whether similar results hold for non-manufacturing rms, which may be affected in different ways by the adoption of JIT policies, and whether other management-accounting initiatives have a similar effect upon the predictability of EPS.

Acknowledgements The authors thank Karen Pincus, Erv Black, Lee Willinger, Bill Collins, and participants in the International Institute of Forecasters 2000 Annual meeting and the American Accounting Association Managerial Section 2002 Midyear meeting for helpful discussions regarding this paper.

References
Balakrishnan, R., Linsmeier, T. J., & Venkatachalam, M. (1996). Financial benets from JIT adoption: effects of customer concentration and cost structure. The Accounting Review, 71, 183205. Bathke, A. W., Lorek, K. S., & Willinger, G. L. (1989). Firm-size and the predictive ability of quarterly earnings data. The Accounting Review, 64, 4968. Beasley, M. S., Carcello, J. V., & Hermanson, D. R. (1998). Fraudulent nancial reporting: 1987 1997. An analysis of U.S. public companies. New York: The Committee of Sponsoring Organizations of the Treadway Commission. Billesbach, T. J., & Hayen, R. (1994). Long-term impact of just-in-time on inventory performance measures. Production and Inventory Management Journal, 35, 6266. Brown, L. D., & Rozeff, M. S. (1979). Univariate time-series models of quarterly accounting earnings per share: a proposed model. Journal of Accounting Research, 17, 179189. Brown, L. D. (1993). Earnings forecasting research: its implications for capital markets research. International Journal of Forecasting, 9, 295320. Foster, G. (1977). Quarterly accounting data: time-series properties and predictive-ability results. The Accounting Review, 52, 121.

T. A. Carnes et al. / International Journal of Forecasting 19 (2003) 743749 Grifn, P. A. (1977). The time-series behavior of quarterly earnings: preliminary evidence. Journal of Accounting Research, 15, 7183. Hunt, A., Moyer, S. E., & Shevlin, T. (1996). Managing interacting accounting measures to meet multiple objectives: A study of LIFO rms. Journal of Accounting and Economics, 21, 339374. Huson, M., & Nanda, D. (1995). The impact of just-in-time manufacturing on rm performance in the US. Journal of Operations Management, 12, 297310. Lipe, R. (1990). The relation between stock returns and accounting earnings given alternative information. The Accounting Review, 65, 4971. Moras, R. G., & Dieck, A. J. (1992). Industrial applications of just-in-time: lessons to be learned. Production and Inventory Management Journal, 33, 2529. Oglove, T. (1987). Quality of earnings: the investor s guide to how much money a company is really making. New York: Free Press. Palepu, K. G., Healy, P. M., & Bernard, V. L. (2000). Business analysis & valuation: using nancial statements, 2nd edition. Cincinnati, OH: South-Western College Publishing. Watts, R. L. (1975). The time-series behavior of quarterly earnings. Unpublished paper, University of New Castle. Biographies: Thomas A. CARNES is an Assistant Professor of Accounting at the University of Arkansas. His primary areas of research are international capital markets and the market valuation of rm reputation. He has published articles in Review of Quantitative Finance and Accounting, Asia-Pacic Journal of Accounting and Economics, Accounting Enquiries, and the Corporate Reputation Review.

749

Jefferson P. JONES is an Associate Professor of Accounting at Auburn University. His primary area of research concerns the quality of reported accounting numbers as well as the quality of the related disclosures. He has published articles in Advances in Accounting, Review of Quantitative Finance and Accounting, Issues in Accounting Education, The Journal of Accounting and Finance Research, The CPA Journal, The Journal of Corporate Accounting and Finance, and the Ohio CPA Journal. Tim BIGGART is an Assistant Professor of Accounting at the University of North Carolina, Greensboro. His research interests are the effects of management innovation on rm performance and the interaction of business decision-making and taxes. He has published articles in The Journal of Business Finance and Accounting, The Engineering Economist, Accounting Enquiries, Tax Notes, and Industrial Management and Data Systems. Katherine J. BARKER is an Associate Professor of Accounting at Lander University. Her research interests include empirical cost and managerial accounting, information systems, logistics, and accounting education. She has published in Accounting Enquiries.

Das könnte Ihnen auch gefallen