Handbook of Economic Forecasting part 104 pptx

In this section the focus will be on the attraction model and on the Bass model, where the expressions for out-of-sample forecasts will be given.. Addition-ally, there will be a discussi

closed-form solutions to these expressions, and hence one has to resort to simulation-based techniques In this section the focus will be on the attraction model and on the Bass model, where the expressions for out-of-sample forecasts will be given Addition-ally, there will be a discussion of how one should derive forecasts for market shares when forecasts for sales are available

4.1 Attraction model forecasts

As discussed earlier, the attraction model ensures logical consistency, that is, market shares lie between 0 and 1 and they sum to 1 These restrictions imply that (functions of) model parameters can be estimated from a multivariate reduced-form model with

I − 1 equations The dependent variable in each of the I − 1 equations is the natural

logarithm of a relative market share, that is, log m i,t ≡ logMi,t

M I,t , for i = 1, 2, , I − 1,

where the base brand I can be chosen arbitrarily, as discussed before.

In practice, one is usually interested in predicting M i,t and not in forecasting the logs of the relative market shares Again, it is important to recognize that, first of all,

exp(E [log mi,t ]) is not equal to E[mi,t] and that, secondly, E[M i,t

M I,t] is not equal to E[M i,t]

E[M I,t]. Therefore, unbiased market share forecasts cannot be directly obtained by these data transformations

To forecast the market share of brand i at time t , one needs to consider the relative

market shares

(48)

m j,t = M j,t

M I,t

for j = 1, 2, , I,

as m 1,t , , m I −1,t form the dependent variables (after log transformation) in the

reduced-form model As M I,t = 1 −Ij−1=1M j,t, it holds that

(49)

M i,t = m i,t

I

j=1m j,t

for i = 1, 2, , I.

market shares as follows First draw η (l) t from N(0, ˜ Σ ), then compute

(50)

m (l) i,t = exp˜μi + η (l)

i,t

I

j=1



k=1

x ˜β k,j,i k,j,t

5

,

with m (l) I,t = 1 and finally compute

(51)

M i,t (l)= m

(l) i,t

I

j=1m (l) j,t

for i = 1, , I,

where l = 1, , L denotes the simulation iteration Each vector (M (l)

1,t , , M I,t (l) )

gen-erated this way is a draw from the joint distribution of the market shares at time t Using

the average over a sufficiently large number of draws one can calculate the expected value of the market shares This can be modified to allow for parameter uncertainty,

similar lines

4.2 Forecasting market shares from models for sales

The previous results assume that one is interested in forecasting market shares based on models for market shares In practice, it might sometimes be more easy to make models for sales One might then me tempted to divide the own sales forecast by a forecast for category sales, but this procedure leads to biased forecasts for similar reasons as before

A solution is given inFok and Franses (2001)and will be discussed next

An often used model (SCAN*PRO) for sales is

(52)

log S i,t = μi+

I

j=1

K

k=1

β k,j,i x k,j,t+

I

j=1

P

p=1

α p,j,i log S j,t −p + εi,t ,

with i = 1, , I, where εt ≡ (ε 1,t , , ε I,t )∼ N(0, Σ) and where xk,j,tdenotes the

kth explanatory variable (for example, price or advertising) for brand j at time t and where β k,j,i is the corresponding coefficient for brand i, seeWittink et al (1988) The

market share of brand i at time t can of course be defined as

(53)

M i,t = S i,t

I

j=1S j,t

.

Forecasts of market shares at time t+1 based on information on all explanatory variables

up to time t + 1, denoted by Πt+1, and information on realizations of the sales up to

period t , denoted by S t, should be equal to the expectation of the market shares given the total amount of information available, denoted by E[Mi,t+1| Πt+1, S t], that is,

(54)

E[Mi,t+1| Πt+1, S t] = E



S i,t+1

I

j=1S j,t+1



Π t+1, S t



.

Due to non-linearity it is therefore not possible to obtain market shares forecasts di-rectly from sales forecasts A further complication is that it is also not trivial to obtain a

forecast of S i,t+1, as the sales model concerns log-transformed variables, and it is well

known that exp(E [log X]) = E[X] See also Arino and Franses (2000)andWieringa

series models In particular,Wieringa and Horvath (2005)show how to derive impulse response functions from VAR models for marketing variables, and they demonstrate the empirical relevance of a correct treatment of log-transformed data

outlined inGranger and Teräsvirta (1993) Naturally, unbiased forecasts of the I market

shares should be based on the expected value of the market shares, that is,

E[Mi,t+1| Πt+1, S t]

=

0

.

0

s i,t+1

I

j=1s j,t+1

(55)

× f (s 1,t+1, , s I,t+1| Πt+1, S t ) ds1,t+1, , ds I,t+1,

where f (s1,t+1, , s I,t+1 | Πt+1, S t ) is a probability density function of the sales conditional on the available information, and s i,t+1denotes a realization of the

stochas-tic process S i,t+1 The model defined in the distribution of S t+1, given Π t+1andS t, is log-normal, but other functional forms can be considered too Hence,

(56)



exp(S 1,t+1), , exp(S I,t+1)

∼ N(Zt+1, Σ ), where Z t = (Z 1,t , , Z I,t )is the deterministic part of the model, that is,

(57)

Z i,t = μi+

I

j=1

K

k=1

β k,j,i x k,j,t+

I

j=1

P

p=1

α p,j,i log S j,t −p

The I -dimensional integral in(55)is difficult to evaluate analytically.Fok and Franses

In short, using the estimated probability distribution of the sales, realizations of the sales are simulated Based on each set of these realizations of all brands, the market shares can be calculated The average over a large number of replications gives the expected value in(55)

Forecasting h > 1 steps ahead is slightly more difficult as the values of the lagged

sales are no longer known However, for these lagged sales appropriate simulated values can be used For example, 2-step ahead forecasts can be calculated by averaging over

simulated values M i,t (l)+2, based on draws ε t (l)+2from N(0, ˆ Σ ) and on draws S i,t (l)+1, which are already used for the 1-step ahead forecasts Notice that the 2-step ahead forecasts do not need more simulation iterations than the one-step ahead forecasts

An important by-product of the simulation method is that it is now also easy to calcu-late confidence bounds for the forecasted market shares Actually, the entire distribution

of the market shares can be estimated based on the simulated values For example, the lower bound of a 95% confidence interval is that value for which it holds that 2.5% of the simulated market shares are smaller Finally, the lower bound and the upper bound always lie within the[0, 1] interval, and this should be the case for market shares indeed.

4.3 Bass model forecasts

The Bass model is regularly used for out-of-sample forecasting One way is to have several years of data on own sales, estimate the model parameters for that particular series, and extrapolate the series into the future Asvan den Bulte and Lilien (1997)

demonstrate, this approach is most useful in case the inflection point is within the sam-ple If not, then one might want to consider imposing the parameters obtained for other markets or situations, and then extrapolate

The way the forecasts are generated depends on the functional form chosen, that is, how one includes the error term in the model TheSrinivasan and Mason (1986)model

seems to imply the most easy to construct forecasts Suppose one aims to predict X n +h,

where n is the forecast origin and h is the horizon Then, given the assumption on the

error term, the forecast is

(58)

ˆXn +h = ˆmF

n + h; ˆθ− Fn − 1 + h; ˆθ.

When the error term is AR(1), straightforward modifications of this formula should be made If the error term has an expected value equal to zero, then these forecasts are

unbiased, for any h.

This is in contrast with the Bass regression model, and also its Boswijk and Franses modification, as these models are intrinsically non-linear For one-step ahead, the true

observation at n+ 1 in the Bass scheme is

(59)

X n+1= α1+ α2N n + α3N n2+ εn+1.

The forecast from origin n equals

(60)

ˆX n+1= ˆα1+ ˆα2N n + ˆα3N n2

and the squared forecast error is σ2 This forecast is unbiased

For two steps ahead matters become different The true observation is equal to

(61)

X n+2= α1+ α2N n+1+ α3N n2+1+ εn+2,

which, as N n+1= Nn + Xn+1, equals

(62)

X n+2= α1+ α2(X n+1+ Nn ) + α3(X n+1+ Nn )2+ εn+2.

Upon substituting X n+1, this becomes

X n+2= α1+ α2



α1+ α2N n + α3N n2+ εn+1+ Nn

(63)

+ α3



α1+ α2N n + α3N n2+ εn+1+ Nn2

+ εn+2.

Hence, the two-step ahead forecast error is based on

X n+2− ˆXn+2

= εn+2+ α2ε n+1

(64)

+ α3



2α1 ε n+1+ 2(α2+ 1)Nn ε n+1+ 2α3N n2ε n+1+ ε2

n+1

.

This shows that the expected forecast error is

(65)

E

X n+2− ˆXn+2= α3σ2.

It is straightforward to derive that if h is 3 or more, this bias grows exponentially with h Naturally, the size of the bias depends on α3 and σ2, which both can be small As the

sign of α3is always negative, the forecast is upward biased

Franses (2003b)points out that to obtain unbiased forecasts for the Bass-type

re-gression models for h = 2, 3, , one needs to resort to simulation techniques, the

same ones as used in Teräsvirta’sChapter 8in this Handbook Consider again the Bass regression, now written as

(66)

X t = g(Zt−1; π) + εt ,

where Z t−1contains 1, N t−1and N t2−1, and π includes p, q and m A simulation-based one-step ahead forecast is now given by

(67)

X n +1,i = g(Zn; ˆπ) + ei ,

where e i is a random draw from the N(0, ˆσ2) distribution Based on I such draws, an

unbiased forecast can be constructed as

(68)

ˆX n+1= 1

I

I

i=1

X n +1,i

Again, a convenient by-product of this approach is the full distribution of the forecasts

A two-step simulation-based forecast can be based on the average value of

(69)

X n +2,i = g(Zn , X n +1,i ; ˆπ) + ei ,

again for I draws, and so on.

4.4 Forecasting duration data

Finally, there are various studies in marketing that rely on duration models to describe interpurchase times These data are relevant to managers as one can try to speed up the purchase process by implementing marketing efforts, but also one may forecast the amount of sales to be expected in the next period, due to promotion planning Inter-estingly, it is known that many marketing efforts have a dynamic effect that stretches beyond the one-step ahead horizon For example, it has been widely established that there is a so-called post-promotional dip, meaning that sales tend to collapse the week after a promotion was held, but might regain their original level or preferably a higher level after that week Hence, managers might want to look beyond the one-step ahead horizon

In sum, one seems to be more interested in the number of purchases in the next week or next month, than that there is an interest in the time till the next purchase The modelling approach for the analysis of recurrent events in marketing, like the purchase timing of frequently purchased consumer goods, has, however, mainly aimed at ex-plaining the interpurchase times The main trend is to apply a Cox (mixed) Proportional Hazard model for the interpurchase times, seeSeetharaman and Chintagunta (2003)for

a recent overview In this approach after each purchase the duration is reset to zero This transformation removes much of the typical behavior of the repeat purchase process in

a similar way as first-differencing in time series Therefore, it induces important lim-itations to the use of time-varying covariates (and also seasonal effects) and duration dependence in the models

An alternative is to consider the whole path of the repeat purchase history on the time scale starting at the beginning of the observation window.Bijwaard, Franses and Paap

all the current and past information available for all purchases as time continues to run along the calendar timescale It is based on theAndersen and Gill (1982)approach It

delivers forecasts for the number of purchases in the next period and for the timing of the

first and consecutive purchases Purchase occasions are modelled in terms of a counting process, which counts the recurrent purchases for each household as they evolve over time These authors show that formulating the problem as a counting process has many advantages, both theoretically and empirically Counting processes allow to understand survival and recurrent event models better

(i) as the baseline intensity may vary arbitrary over time,

(ii) as it facilitates the interpretation of the effects of co-variates in the Cox propor-tional hazard model,

(iii) as Cox’s solution via the partial likelihood takes the baseline hazard as a nui-sance parameter,

(iv) as the conditions for time-varying covariates can be precisely formulated and finally, and finally

(v) as by expressing the duration distribution as a regression model it simplifies the analysis of the estimators

5 Conclusion

This chapter has reviewed various aspects of econometric modeling and forecasting in marketing The focus was on models that have been developed with particular applica-tions in marketing in mind Indeed, in many cases, marketing studies just use the same types of models that are also common to applied econometrics In many marketing re-search studies there are quite a number of observations and typically the data are well measured Usually there is an interest in modeling and forecasting performance mea-sures such as sales, shares, retention, loyalty, brand choice and the time between events, preferably when these depend partially on marketing-mix instruments like promotions, advertising, and price

Various marketing models are non-linear models This is due to specific structures imposed on the models to make them more suitable for their particular purpose, like the Bass model for diffusion and the attraction model for market shares Other models that are frequently encountered in marketing, and less so in other areas (at least as of yet) concern panels of time series Interestingly, it seems that new econometric methodology (like the Hierarchical Bayes methods) has been developed and applied in marketing first, and will perhaps be more often used in the future in other areas too

There are two areas in which more research seems needed The first is that it is not yet clear how out-of-sample forecasts should be evaluated Of course, mean squared forecast error type methods are regularly used, but it is doubtful whether these criteria meet the purposes of an econometric model In fact, if the model concerns the retention

of customers, it might be worse to underestimate the probability of leaving than to overestimate that probability Hence the monetary value, possibly discounted for future events, might be more important The recent literature on forecasting under asymmetric loss is relevant here; see, for example,Elliott, Komunjer and Timmermann (2005)and

Second, the way forecasts are implemented into actual marketing strategies is not trivial, see Franses (2005a, 2005b) In marketing one deals with customers and with competitors, and each can form expectations about what you will do The successfulness

of a marketing strategy depends on the accuracy of stake-holders’ expectations and their subsequent behavior For example, to predict whether a newly launched product will be successful might need more complicated econometric models than we have available today

References

Allenby, G., Rossi, P (1999) “Marketing models of consumer heterogeneity” Journal of Econometrics 89, 57–78.

Andersen, P.K., Gill, R.D (1982) “Cox’s regression model for counting processes: A large sample study” Annals of Statistics 10, 1100–1120.

Andrews, D.W.K., Ploberger, W (1994) “Optimal tests when a nuisance parameter is present only under the alternative” Econometrica 62, 1383–1414.

Arino, M., Franses, P.H (2000) “Forecasting the levels of vector autoregressive log-transformed time series” International Journal of Forecasting 16, 111–116.

Assmus, G., Farley, J.U., Lehmann, D (1984) “How advertising affects sales: Meta-analysis of econometric results” Journal of Marketing Research 21, 65–74.

Bass, F.M (1969) “A new product growth model for consumer durables” Management Science 15, 215–227 Bass, F.M., Leone, R (1983) “Temporal aggregation, the data interval bias, and empirical estimation of bimonthly relations from annual data” Management Science 29, 1–11.

Bewley, R., Griffiths, W.E (2003) “The penetration of CDs in the sound recording market: Issues in specifi-cation, model selection and forecasting” International Journal of Forecasting 19, 111–121.

Bijwaard, G.E., Franses, P.H., Paap, R (2003) “Modeling purchases as repeated events” Econometric Insti-tute Report, 2003-45, Erasmus University Rotterdam Revision requested by the Journal of Business and Economic Statistics.

Blattberg, R.C., George, E.I (1991) “Shrinkage estimation of price and promotional elasticities – Seemingly unrelated equations” Journal of the American Statistical Association 86, 304–315.

Boswijk, H.P., Franses, P.H (2005) “On the econometrics of the Bass diffusion model” Journal of Business and Economic Statistics 23, 255–268.

Bronnenberg, B.J., Mahajan, V., Vanhonacker, W.R (2000) “The emergence of market structure in new repeat-purchase categories: The interplay of market share and retailer distribution” Journal of Marketing Research 37, 16–31.

Chandy, R.K., Tellis, G.J., MacInnis, D.J., Thaivanich, P (2001) “What to say when: Advertising appeals in evolving markets” Journal of Marketing Research 38, 399–414.

Clarke, D.G (1976) “Econometric measurement of the duration of advertising effect on sales” Journal of Marketing Research 8, 345–357.

Cooper, L.G., Nakanishi, M (1988) Market Share Analysis: Evaluating Competitive Marketing Effective-ness Kluwer Academic, Boston.

Davies, R.B (1987) “Hypothesis testing when a nuisance parameter is present only under the alternative” Biometrika 64, 247–254.

Dekimpe, M.G., Hanssens, D.M (2000) “Time series models in marketing: Past, present and future” Inter-national Journal of Research in Marketing 17, 183–193.

Elliott, G., Komunjer, I., Timmermann, A (2005) “Estimation and testing of forecast rationality under flexi-ble loss” Review of Economic Studies In press.

Elliott, G., Timmermann, A (2004) “Optimal forecast combinations under general loss functions and forecast error distributions” Journal of Econometrics 122, 47–79.

Fok, D., Franses, P.H (2001) “Forecasting market shares from models for sales” International Journal of Forecasting 17, 121–128.

Fok, D., Franses, P.H (2004) “Analyzing the effects of a brand introduction on competitive structure using a market share attraction model” International Journal of Research in Marketing 21, 159–177.

Fok, D., Franses, P.H (2005) “Modeling the diffusion of scientific publications” Journal of Econometrics.

In press.

Fok, D., Franses, P.H., Paap, R (2002) “Econometric analysis of the market share attraction model” In: Franses, P.H., Montgomery, A.L (Eds.), Econometric Models in Marketing Elsevier, Amsterdam,

pp 223–256 Chapter 10.

Franses, P.H (2003a) “On the diffusion of scientific publications The case of Econometrica 1987” Sciento-metrics 56, 29–42.

Franses, P.H (2003b) “On the Bass diffusion theory, empirical models and out-of-sample forecasting” ERIM Report Series Research in Management ERS-2003-34-MKT, Erasmus University Rotterdam.

Franses, P.H (2004) “Fifty years since Koyck (1954)” Statistica Neerlandica 58, 381–387.

Franses, P.H (2005a) “On the use of econometric models for policy simulation in marketing” Journal of Marketing Research 42, 4–14.

Franses, P.H (2005b) “Diagnostics, expectation, and endogeneity” Journal of Marketing Research 42, 27– 29.

Franses, P.H., Paap, R (2001) Quantitative Models in Marketing Research Cambridge University Press, Cambridge.

Franses, P.H., van Oest, R.D (2004) “On the econometrics of the Koyck model” Econometric Institute Report 2004-07, Erasmus University Rotterdam.

Franses, P.H., Vroomen, B (2003) “Estimating duration intervals” ERIM Report Series Research in Man-agement, ERS-2003-031-MKT, Erasmus University Rotterdam Revision requested by the Journal of Advertising.

Granger, C.W.J., Teräsvirta, T (1993) Modelling Nonlinear Economic Relationships Oxford University Press, Oxford.

Hansen, B.E (1996) “Inference when a nuisance parameter is not identified under the null hypothesis” Econometrica 64, 413–430.

Klapper, D., Herwartz, H (2000) “Forecasting market share using predicted values of competitor behavior: Further empirical results” International Journal of Forecasting 16, 399–421.

Kotler, Ph., Jain, D.C., Maesincee, S (2002) Marketing Moves A New Approach to Profits, Growth, and Renewal Harvard Business School Press, Boston.

Koyck, L.M (1954) Distributed Lags and Investment Analysis North-Holland, Amsterdam.

Kumar, V (1994) “Forecasting performance of market share models: An assessment, additional insights, and guidelines” International Journal of Forecasting 10, 295–312.

Leeflang, P.S.H., Reuyl, J.C (1984) “On the predictive power of market share attraction model” Journal of Marketing Research 21, 211–215.

Leone, R.P (1995) “Generalizing what is known about temporal aggregation and advertising carryover” Marketing Science 14, G141–G150.

Naert, P.A., Weverbergh, M (1981) “On the prediction power of market share attraction models” Journal of Marketing Research 18, 146–153.

Nijs, V.R., Dekimpe, M.G., Steenkamp, J.-B.E.M., Hanssens, D.M (2001) “The category-demand effects of price promotions” Marketing Science 20, 1–22.

Pauwels, K., Srinivasan, S (2004) “Who benefits from store brand entry?” Marketing Science 23, 364–390 Russell, G.J (1988) “Recovering measures of advertising carryover from aggregate data: The role of the firm’s decision behavior” Marketing Science 7, 252–270.

Seetharaman, P.B., Chintagunta, P.K (2003) “The proportional hazard model for purchase timing: A com-parison of alternative specifications” Journal of Business and Economic Statistics 21, 368–382 Srinivasan, V., Mason, C.H (1986) “Nonlinear least squares estimation of new product diffusion models” Marketing Science 5, 169–178.

Talukdar, D., Sudhir, K., Ainslie, A (2002) “Investigating new product diffusion across products and coun-tries” Marketing Science 21, 97–114.

Tellis, G.J (1988) “Advertising exposure, loyalty and brand purchase: A two stage model of choice” Journal

of Marketing Research 25, 134–144.

Tellis, G.J., Chandy, R., Thaivanich, P (2000) “Which ad works, when, where, and how often? Modeling the effects of direct television advertising” Journal of Marketing Research 37, 32–46.

Tellis, G.J., Franses, P.H (2006) “The optimal data interval for econometric models of advertising” Market-ing Science In press.

van den Bulte, C., Lilien, G.L (1997) “Bias and systematic change in the parameter estimates of macro-level diffusion models” Marketing Science 16, 338–353.

van Nierop, E., Fok, D., Franses, P.H (2002) “Sales models for many items using attribute data” ERIM Report Series Research in Management ERS-2002-65-MKT, Erasmus University Rotterdam.

van Oest, R.D., Paap, R., Franses, P.H (2002) “A joint framework for category purchase and consumption behavior” Tinbergen Institute report series TI 2002-124/4, Erasmus University Rotterdam.

Wedel, M., Kamakura, W.A (1999) Market Segmentation: Conceptual and Methodological Foundations Kluwer Academic, Boston.

Wieringa, J.E., Horvath, C (2005) “Computing level-impulse responses of log-specified VAR systems” In-ternational Journal of Forecasting 21, 279–289.

Wittink, D.R., Addona, M.J., Hawkes, W.J., Porter, J.C (1988) “SCAN*PRO: The estimation, validation, and use of promotional effects based on scanner data” Working Paper, AC Nielsen, Schaumburg, Illinois.

n indicates citation in a footnote.

Abeysinghe, T 670

Abidir, K 585

Abou, A 758

Abraham, B 304

Adams, F 753, 769

Addona, M.J 1005

Aggarwal, R 599

Aguilar, O 14, 852

Ahn, S.K 313, 318, 677

Ainslie, A 1002

Aiolfi, M 138, 155, 161, 164, 183, 184, 186,

187, 439, 524

Aït-Sahalia, Y 797, 798, 838, 864

Akaike, H 291, 315, 318, 480

Al-Qassam, M.S 637n

Alavi, A.S 291

Albert, J 635, 900, 915, 930

Albert, J.H 8, 14

Alexander, C 812, 853

Alizadeh, S 838, 864

Allen, D 480

Allen, H 760

Allenby, G 1002

Alt, R 623

Altissimo, F 535, 896

Amato, J.D 978

Amemiya, T 304

Amirizadeh, H 70

Amisano, G 64

Andersen, A 182

Andersen, P.K 1009

Andersen, T.G 416, 696–700, 784n; 805, 812,

817, 818, 821, 828–830, 832–838, 840, 850,

851, 853, 856, 858, 864

Anderson, B.D.O 362, 365

Anderson, H.M 446, 927

Anderson, O 739, 743, 769

Anderson, T.W 524, 529, 529n; 530

Andreou, E 813

Andrews, D 570

Andrews, D.W.K 105, 116, 201, 202, 214,

215, 215n; 216, 218, 218n; 219, 228, 426,

607, 623, 637, 995 Andrews, M.J 633 Andrews, R.C 333 Ang, A 852 Aoki, M 294 Arino, M 1005 Armstrong, J.S 161, 162, 607 Artis, M.J 534, 889, 891, 897, 898, 910, 917,

918, 923, 924, 926, 927, 937 Ashley, R 103, 107, 119, 220 Assimakopoulos, V 333 Assmus, G 995 Auerbach, A.J 904 Avramov, D 116, 541, 542

Bac, C 693 Bachelier, L 818 Bacon, D.W 418 Bai, J 203, 204, 206–210, 213, 220, 223, 276,

424, 530, 531, 623, 627, 627n; 896, 910 Baillie, R.T 317, 614, 805, 812, 814 Bakirtzis, A 24

Bakshi, G 798 Balke, N.S 625–627 Baltagi, B.H 202 Bandi, F 838, 864 Banerjee, A 424, 534, 564, 566n; 641, 813, 910

Banerjee, A.N 910, 920 Banerji, A 922 Barnard, G.A 23 Barndorff-Nielsen, O.E 58, 833, 835, 838,

849, 853 Barnett, G 59 Barnett, W.A 635, 655 Barone-Adesi, G 797 Barquin, J 797 Barrett, C 796 I-1

Arino, M., Franses, P.H (2000) ? ?Forecasting the levels of vector autoregressive log-transformed time series” International Journal of Forecasting 16, 111–116.

Assmus,