Entrepreneur: Start & Grow Your Business

Aggregation of price risk over commodities: an economic index number approach.


by Coyle, Barry T.

Index number theory is commonly used to motivate aggregation of prices over commodities. In particular, three price indexes (Fisher, Tornqvist, Walsh) are favored by the most popular criteria (economic test and weighted stochastic approaches), and these three indexes closely approximate each other in practice (Diewert 2004a,b). Index number methods for aggregating prices and quantities over commodities are essential to empirical economics.

Other conceptual approaches are also used to aggregate prices over commodities or agents. Prices can be aggregated over commodities under distributional assumptions such as perfect contemporaneous covariance (Hicks' [1936] aggregation) or more generally mean scaling (Lewbel 1996; Davis, Lin, and Shumway 2000; Shumway and Davis 2001; Davis 2003). Different prices for the same commodity can be aggregated over firms assuming Gorman (Gorman 1953) polar form type relations in prices or more generally knowledge of second moments of cross-section distributions of price (Pope and Chambers 1989; Chambers and Pope 1991, 1994).

However, apparently an index number or related approach has not been presented for aggregation of price risk over commodities. (1,2) Such aggregation involves price covariances as well as variances, so this is a higher dimensional problem than aggregation over prices.

It is a stylized fact that agricultural producers are risk averse and decisions are influenced by risk and uncertainty (e.g., see references in Moschini and Hennessy 2001), and of course there are multiple agricultural commodities. So, it is important to develop summary measures of price risk in a multicommodity setting. The absence of an index number approach to this aggregation is a serious problem for policy and empirical research.

Common approaches to aggregating price risk are (a) to reduce the dimension of the problem by omitting many covariances and variances or (b) to calculate a variance of an aggregate price index. However, the first approach obviously ignores information about risk and the second approach has no basis in index number theory.

This article extends index number theory to aggregation of price risk over commodities in production, and includes an application to major Manitoba crops. The analysis is an extension of the economic approach to aggregation of prices under certainty due largely to Diewert.

We focus on price risk rather than quantity risk primarily because it is well known that output risk is underestimated by regional data on output (Eisgruber and Schuman 1963), and farm-level output data are less common than regional data. Since there is higher contemporaneous covariance of prices than output levels over farms, this problem is less serious for prices.

The study is organized as follows. The first section summarizes relevant standard index number theory. The next section presents preliminaries and initial results. The subsequent section presents the general economic theory approach to aggregation of price risk over commodities, and this is related to Tornqvist-type output quantity indexes. Then the analysis is extended to Fisher-type output quantity indexes, and the following section considers higher moments, subindexes, and output quantity risk. Finally the methodology is applied to price data for major Manitoba crops. Results illustrate the empirical importance of the index number approach to aggregation of price risk developed here.

Standard Index Number Theory

We first provide a brief summary of standard index number theory before extensions to price risk. Readers unfamiliar with standard index number theory should read Diewert (1976 or 2004a). Index number theory has focused on aggregation under certainty or risk neutrality. Consider aggregation of output prices and quantities, given output prices p = ([p.sub.1],., [p.sub.m]), output levels y = ([y.sub.1],., [y.sub.m] for m outputs and input prices w = ([p.sub.1],., [p.sub.n]), input levels x = ([x.sub.1],., [x.sub.n]) for n inputs, and two time periods 0, 1. A fundamental goal of index number theory (in both economic and test approaches) is to decompose the value (revenue) change between the two periods, [p.sub.1][y.sub.1]/[p.sub.0][y.sub.0] [equivalent to] [[summation].sub.i=1,.,m][p.sub.i1][y.sub.i1]/[[summation].sub.i=1,.,m] [p.sub.i0][y.sub.i0] into a price change part [P.sub.1][P.sub.0] and a quantity change part YflY0, i.e., [P.sub.1][P.sub.0] and [Y.sub.1]/[Y.sub.0] should satisfy

(1) ([P.sub.1][P.sub.0])([Y.sub.1]/[Y.sub.0]) = [p.sub.1][y.sub.1]/ [p.sub.0]/[y.sub.0].

[P.sub.1]/[P.sub.0] is an aggregate price index over commodities for period t = 1 relative to t = 0, and similarly [Y.sub.1]/[Y.sub.0] is an aggregate quantity index over commodities for t = 1 relative to t = 0. This criterion for index numbers was first proposed by Fisher (1911, p. 418) and was called the product test equation by Frisch (1930, p. 399).

Diewert (1976) notes that indexes aggregating input quantities or input prices should preserve the contributions to output quantity or marginal/average cost, respectively (the index should be exact), and he shows that Tornqvist indexes accomplish this under Translog flexible functional forms and common behavioral assumptions (the Tornqvist index is superlative). Analogous arguments apply to aggregation of output quantities and prices, as summarized next.

A common output quantity index is the following Tornqvist discrete time approximation [Y.sub.1]/[Y.sub.0] to a continuous Divisia aggregate index:

(2) log([Y.sub.1]/[Y.sub.0])

= [[summation].sub.i=1,.,m]{([S.sub.i1] + [S.sub.io]/2} log([y.sub.i1]/[y.sub.i0])

where [s.sub.it] [equivalent to] [p.sub.it][y.sub.it]/ [[summation].sub.J=1,.,n][W.sub.jt][X.sub.jt]. (3) [Y.sub.1]/[Y.sub.0] is a superlative output quantity index corresponding to a Translog joint cost function C(w, y) and static competitive profit maximization (Diewert 1980a, 1976). Given [Y.sub.1]/[Y.sub.0], a corresponding implicit index of output prices [P.sub.1]/[P.sub.0] is defined by the product test equation (1). This implicit index [P.sub.1]/[P.sub.0] can be interpreted as a superlative index for ratio of aggregate price (Diewert 1976).

A Fisher output quantity index is a prominent alternative:

(3) [Y.sub.1]/[Y.sub.0] = [[[p.sub.0][y.sub.1]/ [p.sub.0][y.sub.0]].sup.1/2] [[p.sub.1][y.sub.1]/[p.sub.1][y.sub.0]].sup.1/2].

This is also a superlative output quantity index, and it is superior to a Tornqvist index by the test approach (Diewert 1992).

Aggregation of Price Variances and Covariances: Preliminaries and Initial Laspeyres/Fisher Indexes

A common approach to approximating an index of price risk over commodities is simply to construct a variance from time series data for an aggregate price index such as the above Tornqvist (e.g., Coyle 1992, 1999). However, this ad hoc approach cannot be rationalized in terms of index number theory.

This weakness of the ad hoc approach can be explained as follows. A k parameter distribution for price implies that the first k moments generally characterize the distribution (higher moments are dependent on these moments); so these moments can evolve differently over time. (4) For example, assuming a normal distribution whose two parameters change over time, the mean and variance are independent constructs (determining higher moments) and hence can evolve separately over time. So, the relation between evolution of expected prices and of price variances/ covariances over time can be complex, i.e., price variances/covariances are unlikely to be determined by expected prices over time. In turn, the evolution of sample price, price mean, and price variance over time can be complex. Consequently, a simple transformation such as a variance of a correct price index is unlikely to provide a correct price risk index, which is an aggregation of price variances/covariances rather than of sample prices or expected prices. In sum, observed prices and risk can evolve differently over time (e.g., prices known with certainty can change over time, or consider a GARCH model of price). So, a variance of an index of observed prices does not provide a correct index number aggregate of risk.

A related approach is to specify a GARCH model for an aggregate price index and calculate the conditional variance of the disturbance. This approach is common in the literature on stock market price indexes (Bollerslev, Chou, and Kroner 1992). However, again the estimated aggregate of price risk is a (complex) transformation of data on aggregate prices rather than an aggregate of commodity level price variances and covariances, which evolve differently from observed prices or expected prices. A more appropriate procedure is to estimate a multivariate GARCH model over all prices and then aggregate the conditional price variances and covariances by a correct index number procedure.

In sum, common approaches for measuring aggregate price risk from a price index do not provide a correct index number approach to aggregating price risk over commodities. Moreover, as we shall demonstrate in the empirical section of this article, this distinction in theory between a variance of a price index and a correct price risk index is also important in practice.

Our approach to constructing index numbers for aggregate price risk is in the spirit of the economic analysis of index numbers under certainty. In our case, price risk for individual outputs contributes to risk regarding total revenues: variances and covariances for output prices contribute to variance of total revenue. Ignoring output (quantity) risk or uncertainty, price risk and output levels jointly contribute to revenue risk as [VR.sub.t] = [y.sup.T.sub.t][Vp.sub.t][y.sub.t], where Vp is the price covariance matrix and y is a vector of output levels. Our index number theory is invariant to the time dimension of price risk, i.e., the price covariance matrix Vp can be viewed as reflecting either (e.g.) annual, monthly, or daily price risk. Of course, the particular time dimension is important in empirical applications of the theory.

An appropriate aggregation procedure for price risk Vp will preserve the contribution of Vp to revenue risk while controlling for effects of output levels y. This is the fundamental criterion in designing index number approaches to aggregation of price risk over commodities, and it is similar in spirit to standard index number theory. Standard index number problems are best addressed in terms of value ratios rather than levels (Diewert 2004a), and we proceed in a similar manner.

The following "Laspeyres" index is the most obvious approach to aggregating price risk over commodities:

(4) [(VP.sub.1]/[VP.sub.0])L = [y.sup.T.sub.0][Vp.sub.1][y.sub.0]/ [y.sup.T.sub.0][Vp.sub.0][y.sub.0]

using base period weightings [y.sub.0] throughout the index. An analogous "Paasche" index [(VP.sub.1][VP.sub.0]).sup.P] can be defined using base period weightings [y.sub.1]. Such weightings would obviously be appropriate if outputs were in fixed proportions. However, in principle this approach misrepresents the contribution of price risk Vp to revenue risk VR under general changes in output levels y, somewhat as aggregation with standard Laspeyres indexes generally loses the economic meaning of the subaggregates (Diewert 1981). (5) An analogous "Fisher" index is

(5) [([VP.sub.1]/[VP.sub.0]).sup.F]

= [{[([VP.sub.1]/[VP.sub.0]).sup.L]([VP.sub.1]/ [VP.sub.0]).sup.P]}.sup.1/2].

Fisher indexes typically have better properties than do Laspeyres or Paasche indexes. Nevertheless, I am unaware of any applications or references even to these Laspeyres, Paasche, or Fisher indexes for aggregation of price risk.

Aggregation of Price Variances and Covariances: A Tornqvist Index Approach

Similar to the product test equation (1), a fundamental goal of index number theory for aggregating output price risk should be to decompose the change in revenue risk between two periods, [y.sup.T.sub.1][Vp.sub.1][y.sub.1]/[y.sup.T.sub.0][Vp.sub.0][y.sub.0], into a price risk change part [VP.sub.1]/[VP.sub.0] and a quantity change part [Y.sub.1]/[Y.sub.0]. So in the spirit of standard index number theory and elementary statistics, [VP.sub.1]/[VP.sub.0] and [Y.sub.1]/[Y.sub.0] should satisfy the following equation analogous to (1):

(6) [([VP.sub.1]/[VP.sub.0])([Y.sub.1]/[Y.sub.0]).sup.2] = [y.sup.T.sub.1][Vp.sub.1][y.sub.1]/ [y.sup.T.sub.0][Vp.sub.0][y.sub.0].

Then a price risk index can be correctly calculated from (6) given an appropriate aggregate output index. This connects aggregation of price risk to the formulation of aggregate index numbers for multiple outputs.

Suppose an output quantity index [Y.sub.1]/[Y.sub.0] is superlative under appropriate behavioral assumptions. Then (6) implies an aggregate price risk index [VP.sub.1]/[VP.sub.0]

(7) [VP.sub.1]/[VP.sub.0] = ([y.sup.T.sub.1][Vp.sub.1][y.sub.1]/ [y.sup.T.sub.0][Vp.sub.0][y.sub.0]/ [([Y.sub.1]/[Y.sub.0]).sup.2]

that is exact and superlative in terms of preserving the contribution of price risk for commodities to revenue risk.

Assuming risk aversion, index numbers for multiple outputs should not be calculated from profit maximization, and in general index numbers depend on knowledge of risk preferences or the corresponding dual utility function (Chambers 1983). Nevertheless, we can show that a Tornqvist-like aggregate output quantity index is appropriate assuming a (static) Translog cost function and constant returns to scale (CRTS) in nonjoint technologies, which are common assumptions in index number theory. Later, we will relax these assumptions. This result is stated as the following Proposition (see Appendix for proof).

PROPOSITION 1. Define the output quantity index [Y.sub.1]/[Y.sub.0]:

(8) log([Y.sub.1]/[Y.sub.0])

= [[summation].sub.i=1,.,m]{([[theta].sub.i1] + [[theta].sub.i0]/2} log([Y.sub.i1]/[Y.sub.i0])

where [[theta].sub.i0] [equivalent to] [C.sub.it]/[C.sub.t] is the share of output i in total cost at time t. Assume CRTS non joint technologies, (static) cost minimization, and a Translog cost function C(w, y). Then

(9) [Y.sub.1]/[Y.sub.0] = ([C.sub.1]/[C.sub.0])/([AC.sub.1]/ [AC.sub.0])

where [C.sub.1]/[C.sub.0] is ratio of total cost and log ([AC.sub.1]/[AC.sub.0]) = [[summation].sub.i=1,.,n][S.sub.wj] log([w.sub.j1]/[w.sub.j0]) [S.sub.wj] [equivalent to] {[w.sub.j1][x.sub.j1]/[C.sub.1] + ([w.sub.j0][x.sub.j0]/[C.sub.0])}/2. [AC.sub.1]/[AC.sub.0] is an index of aggregate average cost. Thus [Y.sub.1]/[Y.sub.0] is an exact output quantity index, and it is superlative.

In contrast, if technology is joint or not CRTS, then the above result does not apply. Then an aggregate output index generally depends on the properties of risk preferences as well as technology. For example, assume constant absolute risk aversion (CARA) and the firm solves the utility maximization problem

(10) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where [alpha] is the constant coefficient of absolute risk aversion. This leads to the following Proposition.

PROPOSITION 2. Define the output quantity index [Y.sub.1]/[Y.sub.0]:

(11) log([Y.sub.1]/[Y.sub.0])

= [[summation].sub.i=1,.,m]{([[gamma].sub.i1] + [[gamma].sub.i0]/2log([y.sub.i1]/[y.sub.i0])

where [[gamma].sub.it] [equivalent to] ([P.sub.it][y.sub.it] - [alpha][VR.sub.t])/[C.sub.t] Assume CARA utility maximization (10) and a Translog cost function C(w, y). Then [Y.sub.1]/[Y.sub.0] satisfies (9).

However, it is important to note that this index requires an estimate of the coefficient of absolute risk aversion [alpha]. (6)

Extensions to a Fisher Index

A Fisher ideal index [([Y.sub.1]/[Y.sub.0]).sup.F] (3) is often advocated for aggregation. The economic interpretation of this index is typically based on revenue maximization: then an aggregator function of the form f(y)= [[[summation].sub.i=1,.,m] [[summation].sub.i=1,.,m][a.sub.ij][y.sub.i][y.sub.j]].sup.1/2] implies that [([Y.sub.1]/[Y.sub.0]).sup.F] = f(y.sub.1])/f([y.sub.0]). Thus, this index is superlative assuming revenue or profit maximization (Diewert 1976, 2004a). However for our purposes, a Fisher output quantity index should be defined from cost minimization and a cost function, independently of profit maximization or risk preferences. (7)

A Fisher-like output quantity index can be developed from a cost minimization model as follows. Assume nonjoint CRTS technologies, so that average cost for output i is [AC.sub.i](w). Define a new Fisher-like index by substituting [AC.sub.i](wt) for [p.sub.t] in (3):

(12) [([Y.sub.1]/[Y.sub.0]).sup.F]*

= [[[AC.sub.0]/[y.sub.1]/[AC.sub.0]/[y.sub.0].sup.1/2] [[[AC.sub.1]/[y.sub.1]/[AC.sub.1]/[y.sub.0].sup.1/2]

where [AC.sub.t] [equivalent to] ([AC.sub.1]([W.sub.t]),., [AC.sub.m]([W.sub.t])). Here, e.g., [AC.sub.0][y.sub.1] [equivalent to] [[summation].sub.i=1,.,m][AC.sub.i]([w.sub.0])[y.sub.i1] = C([w.sub.0], [y.sub.1]). The following Proposition shows that (under standard assumptions in the literature) this is a superlative output quantity index.

PROPOSITION 3. Assume nonjoint CRTS technologies and a cost function C(w, y) = c(w)h(y). Then

(13)

[([Y.sub.1]/[Y.sub.0]).sup.F*] = [[Y([w.sub.0], [y.sub.1])/Y([w.sub.0], [y.sub.0])].sup.1/2]

x [[Y([w.sub.1], [y.sub.1])/Y([w.sub.1], [y.sub.0])].sup.1/2]

where Y(w, y) is an output aggregator function.

This implies that [([Y.sub.1]/[Y.sub.0]).sup.F*] is an output aggregator, independently of profit maximization or risk preferences (assuming cost minimization and nonjoint CRTS technologies). Since c(w) can be a flexible functional form, [([Y.sub.1]/[Y.sub.0]).sup.F*] is a superlative output quantity index.

Thus, we have developed two economic index numbers for aggregating outputs based on cost minimization, the Tornqvist index [([Y.sub.1]/[Y.sub.0]).sup.T] (8) and the Fisher index [([Y.sub.1]/[Y.sub.0]).sup.F*] (12). In either case, the output quantity index can be substituted into (7) to calculate a corresponding index of price risk over commodities.

Extensions to Higher Moments, Subindexes, and Output Risk

The above analysis can easily be extended to aggregation over commodities of higher moments of price risk. Given that y is non-stochastic, the [k.sup.th] moment of the probability distribution for revenue R = [[summation].sub.i=1,.,m][p.sub.i][y.sub.i] is defined as [M.sup.k]R = E{[[[summation].sub.i][p.sub.i][y.sub.i] - E([[summation].sub.i][p.sub.i][y.sub.i])].sup.k]} = E{[[[[summation].sub.i]([p.sub.i] - E[p.sub.i])[y.sub.i]].sup.k]} (k [greater than or equal to] 2) For example, the third moment of the distribution for revenue with two commodities (m = 2) is [M.sup.3]R = [m.sub.111][y.sup.3.sub.1] + 3 [m.sub.112][y.sup.2.sub.1][y.sub.2] + 3 [m.sub.122][y.sub.1][y.sup.2.sub.2] + [m.sub.222][y.sup.3.sub.2] where (e.g.) [m.sub.112] = E[[([p.sub.1] - E[p.sub.1]).sup.2] ([p.sub.2] - E[p.sub.2]) is a third moment of price. The Laspeyres and Fisher indexes (4) and (5) extend in an obvious manner to higher moments.

Let [M.sup.k]p denote an aggregator of the kth moments of the joint probability distribution for prices. In the spirit of index number theory and elementary statistics, an aggregate output price risk index [M.sup.k]P and an aggregate output quantity index Y should satisfy the following identity in value ratios:

(14) [([M.sup.k][P.sub.1]/[M.sup.k][P.sub.0])([Y.sub.1]/ [Y.sub.0]).sup.k]

= [M.sup.k][R.sub.1]/[M.sup.k][R.sub.0] k [greater than or equal to] 2.

This is a generalization of (6). Thus, an aggregate index of higher moments of price risk can be calculated from (14) given an appropriate aggregate output index, such as the Tornqvist and Fisher indexes (8,12) of the previous section. Given such an output quantity index Y that is superlative under our behavioral assumptions, then (14) implies an aggregate price risk index

(15) ([M.sup.k][P.sub.1]/[M.sup.k][P.sub.0]

= [([M.sup.k][R.sub.1]/[M.sup.k][R.sub.0]/ ([Y.sub.1]/[Y.sub.0]).sup.k]

that is exact and superlative in terms of preserving the contribution of price risk for commodities to revenue risk.

Our analysis can also easily be extended to partition price risk into aggregate subgroups. For example, variance of total revenue can be partitioned in terms of two commodity groups A and B as

(16)

VR = [y.sup.T.sub.A][Vp.sub.A][y.sub.A] + [y.sup.T.sub.B][Vp.sub.B[y.sub.B] + [2y.sup.T.sub.A][Vp.sub.AB[y.sub.B].

Matrices [Vp.sub.A] and [Vp.sub.B] can be aggregated similarly to (7) using (e.g.) Tornqvist subindexes [Y.sub.A] and [Y.sub.B]. A Laspeyres index analogous to (4) could be defined for [Vp.sub.AB]. However, a more appropriate approach is to define a "Tornqvist" or "Fisher" index for [Vp.sub.AB] using the following simple generalization of (7):

(17) [VP.sub.AB1]/[VP.sub.AB0] = ([y.sup.T.sub.A1][Vp.sub.AB1] [y.sub.B1]/[y.sup.T.sub.A0][Vp.sub.AB0][y.sub.B0]) /{(Y.sub.A1]/[Y.sub.A0])([Y.sub.B1]/[Y.sub.B0])}. (8)

Extensions of the analysis to include risk for both output price and output quantity are more problematic. Assuming both p and y are stochastic but coy(p, y) = 0, then variance of revenue is

(18) VR = [Ey.sup.T]VpEy + [Ep.sup.T]VyEp + [[summation].sub.i][[summation.sub.j][Vp.sub.ij][Vy.sub.ij]

where Vy is the covariance matrix for output risk, and Ey is the vector of expected outputs. The more realistic assumption that p and y covary typically leads to a more complex expression for variance of revenue (see Bohrnstedt and Goldberger [1969] for an exception). An aggregate index [VP.sub.1]/[VP.sub.0] can be defined as in (7), and an aggregate index for Vy can be defined similarly as

(19) [VY.sub.1]/[VY.sub.0] = ([Ep.sup.T.sub.1][Vy.sub.1]/ [Ep.sup.T.sub.0][Vy.sub.0][Ep.sub.0])/[([EP.sub.1]/[EP.sub.0]).sup.2]

where [EP.sub.1]/[EP.sub.0] is an appropriate Tornqvist or Fisher (expected) output price index. Since the last term in (18) depends solely on Vp and Vy (independently of Ey and Ep), it directly defines an aggregate Z or [Z.sub.1]/[Z.sub.0]. These indexes [VP.sub.1]/[VP.sub.0], [VY.sub.1]/[VY.sub.0] (and [Z.sub.1]/[Z.sub.0]) can be viewed as aggregate measures in ratio form of the components of revenue risk VR. For certain purposes, this may be satisfactory.

On the other hand, note that (18) implies that VR is additive rather than multiplicative in these components. Therefore (in contrast to the product test equation (1) and its generalization (6)) the value ratio [VR.sub.1]/[VR.sub.0] for revenue risk is not a simple transformation of these indexes, e.g., [VR.sub.1]/[VR.sub.0] [not equal to] ([VP.sub.1]/[VP.sub.0])[([EY.sub.1]/[EY.sub.0]).sup.2] + ([VY.sub.1]/[VY.sub.0]) [([EP.sub.1]/[EP.sub.0]).sup.2] + [Z.sub.1]/[Z.sub.0]. Thus these indexes may provide a poor approximation to [VR.sub.1]/[VR.sub.0].

Alternatively, a Laspeyres-type index for revenue risk under both price risk and quantity risk can be defined as

(20) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

However, this index does not facilitate decomposition into price and quantity aggregates. Paasche and Fisher indexes can be defined similarly.

Empirical Applications

This article has shown that price risk should be aggregated over commodities by applying index number theory to aggregate price variances and covariances rather than by calculating a variance of an aggregate price index, and several alternative measures have been suggested for such aggregation. This section illustrates the practical importance of this methodology in the case of major Manitoba crops.

The index number theory developed above is formally independent of the time dimension of risk, e.g., the price covariance matrix Vp above can be defined as an annual, monthly, or daily measure of price risk. The appropriate time dimension may vary with the application. Consequently, this section considers variation in crop prices at both the annual and monthly level (daily data were not as readily available).

Different measures of price risk have been used in empirical studies. The most common alternatives are ad hoc naive models and GARCH models. This section considers both models of price risk.

In all cases, the aggregation measures proposed here are highly correlated with each other but are less correlated with variances of aggregate price indexes. So the common approach to aggregation, a variance of an aggregate price index, is not a close proxy in practice to approaches that are superior in theory. Thus, our theoretical arguments have considerable practical importance.

A simple measure of price risk based on naive expectations approximates expected prices by a one-period lag and price variances and covariances using a simple three period weighted average of prediction errors:

(21) [E.sub.t-1][p.sub.t] = [P.sub.t-1]

(22)

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

Such approximations are common with annual data (we follow the weightings of Chavas and Holt [1990]). Price covariances and higher moments are easily calculated in this manner, in contrast to GARCH models.

Univariate ARCH and GARCH models are commonly used to estimate the second moment of price risk, but extensions to price covariances and higher moments are problematic. A multivariate GARCH model with time-varying conditional covariances was developed by Bollerslev, Engle, and Wooldridge (1988), but this model has been intractable for empirical research. Alternatively, Bollerslev (1990) simplifies the model by assuming constant conditional correlations and then jointly estimates variances and covariances. More recently, Engle (2002) advocates a two-step multivariate GARCH process: (a) first univariate GARCH models are estimated for each separate price or return, and then (b) standardized residuals from the GARCH models are used to estimate parameters of correlation. Engle and Sheppard (2001) extend this analysis. The two-step procedure is particularly simple under constant conditional correlations. ARCH/GARCH models have recently been extended to higher moments (Hansen 1994; Harvey and Siddique 1999; Wang et al. 2001; Roberts 2001; Haas, Mittnik, and Paolella 2004; Alexander and Lazar 2006). (9)

This article adopts the simple two-step process for estimating price risk (price variances and covariances): univariate GARCH models are first estimated, then standardized residuals are used to estimate constant conditional correlations (as assumed by Bollerslev), and in turn time varying conditional covariances are calculated. First, univariate GARCH models are specified for each of the six prices assuming a GARCH(1,1) conditional disturbance [u.sub.it] = [[epsilon].sub.it][h.sup.1/2.sub.it]. Second, assuming constant correlations for u's, pairwise regressions of standardized residuals are estimated: [e.sub.it] = [[beta].sub.ij0] + [[beta].sub.ij][e.sub.jt] + [v.sub.ijt] where e are estimates of [[epsilon].sub.it] = [u.sub.it]/[h.sup.l/2.sub.it], and coefficient [[beta].sub.ij] estimates a constant cov([[epsilon].sub.it][[epsilon].sub.jt]) over t assuming var([[epsilon].sub.jt]) - 1. Then estimates of regression coefficients [[beta].sub.ij] and conditional variances hit are used to calculate time--varying conditional covariances, cov([u.sub.it][.sub.jt]) = [[beta].sub.ij][([h.sub.it][h.sub.jt]).sup.1/2]. More complex approaches can be used to estimate covariances (Engle 2002), but research on this is still preliminary (Engle 2004).

Annual data on market prices ($/tonne) for the six major crops in Manitoba (wheat, barley, canola, oats, rye, and flax) over 1950-2002 were obtained from the Manitoba Agriculture Yearbook (2002). Correlations between market prices and coefficients of variation for each crop are presented in table 1. The high correlations between prices indicate the importance of including price covariances in aggregation of price risk over crops. For simple autoregressive models of price, homoscedasticity is generally rejected at the 0.05 level for multiplicative heteroscedasticity (Harvey 1976, 1990), but homoscedasticity is not rejected for ARCH (by Engle Lagrange multiplier test). Nevertheless, we consider GARCH models of price risk as well as naive and multiplicative heteroscedastic models.

First, consider the common naive model of price risk (21)-(22) with annual data. In principle, this naive model is a poor forecast for the distribution of prices. Nevertheless, estimates of (e.g.) the univariate GARCH models for prices (discussed below) offer some support for this model for all crops excluding oats: coefficients for a one-year lag in price are very close to +1.0 and coefficients for other lags are smaller and approximately cancel out; and correlations between predicted prices from GARCH and a one-period lag in prices are high (ranging from 0.9729 for wheat to 0.9179 for oats).

The resulting estimates of time-varying conditional variances and covariances for all six crops are aggregated using the methodology developed above. In constructing indexes of price risk from (21)-(22), estimates of the covariance matrix Vp from (22) are used to calculate revenue uncertainty VR = [y.sup.T]Vpy. A chained Laspeyres index (4) [VP.sub.t]/[VP.sub.t-1] is calculated. This does not require the calculation of an aggregate output quantity index, in contrast to our other approaches based on (7). Since we do not have reliable time series data on average costs by crop, our Tornqvist and Fisher-type output quantity indexes (8) and (12) cannot be calculated. So instead we calculate standard chained Tornqvist, Fisher, and Laspeyres output quantity indexes [Y.sub.t]/[Y.sub.t-1] (in effect substituting [p.sub.it] for [AC.sub.it] in (8) and (12)). (10) Table 2 indicates that the Tornqvist and Fisher output quantity indexes are highly correlated with the Laspeyres index (r = 0.9912, 0.9992, respectively). Given these output quantity indexes, aggregate indexes for price risk are calculated from

(23) [VP.sub.t]/[VP.sub.t-1] = ([y.sup.T.sub.t][Vp.sub.t][y.sub.t]/ [y.sup.T.sub.t-l][Vp.sub.t-1][y.sub.t-1]) /[([Y.sub.t]/[Y.sub.t-1]).sup.2]

which is analogous to (7).

For comparison, we also construct a Tornqvist output price index [q.sub.t] [equivalent to] [([P.sub.t]/[P.sub.t-1]).sup.Torn] and then calculate its variance as [var.sub.t-1]([q.sub.t]) = 0.50 [([q.sub.t-1]- [q.sub.t-2]).sup.2] + 0.33 [([q.sub.t-2] - [q.sub.t-3]).sup.2] + 0.17 ([q.sub.t-3] - [q.sub.t-4]) similarly to (22). This provides an aggregate measure of price risk [VP.sup.B], which is in theory highly inferior to the other indexes of aggregate price risk.

Table 3 presents correlations between these price risk indexes based on naive expectations and annual data. [VP.sup.L] denotes the Laspeyres index (4). [VP.sup.Torn], [VP.sup.Fish,] [VP.sup.Lasp] denote the aggregate indexes of price risk constructed from (23) using Tornqvist, Fisher, and Laspeyres output quantity indexes, respectively. [VP.sup.B] denotes the price variance for the aggregate Tornqvist output price index. The first four indexes can be viewed as alternative index number approaches for aggregating price risk over commodities, but the last index VPB does not have a valid interpretation as an index number that preserves the contribution of commodity price variances and covariances to revenue risk.

The three indexes constructed from (23) are very highly correlated, with correlations r ranging between 0.9997 and 0.9846. This reflects very high correlations between the corresponding output quantity indexes (table 2). These price risk indexes are also highly correlated with the Laspeyres index (4), with r ranging between 0.9897 and 0.9678. However, all four of these indexes show much smaller correlations with the last index [VP.sup.B], with r ranging between 0.5129 and 0.4656. This last result illustrates that the distinctions in theory between the first four indexes and [VP.sup.B] are also important in practice. (11)

Although GARCH models are inappropriate for annual data, for comparison the two-step multivariate GARCH model under constant conditional correlations was also estimated with this data set. Current price Pi is specified as a function of a four-period lag in Pi (longer lags are insignificant) and a time trend. Univariate GARCH(1,1) models are estimated for each crop by maximum likelihood using Broyden--Fletcher--Goldfarb--Shanno (BFGS) algorithms as encoded in Shazam (White 1997). Second, assuming constant correlations for u's, pairwise regressions of standardized residuals are estimated and conditional covariances are calculated, as discussed above. Estimates of the covariance matrix are used to calculate measures of revenue risk [VR.sub.t] = [y.sub.t][Vp.sub.t][y.sub.t]. The correlation between this measure and the measure using Vp from naive models is 0.860. Then index numbers for price risk are calculated similarly to the naive case.

The three indexes constructed from (23) are very highly correlated, with correlations r ranging between 0.991 and 0.971. These indexes are also highly correlated with the Laspeyres index (4), with r ranging between 0.974 and 0.925. However, all four of these indexes are negatively correlated with the index [VP.sup.B] (based on a univariate GARCH model of a Divisia price index), with r ranging between -0.274 and -0.420.

A model of multiplicative heteroskedasticity (Harvey 1976, 1990) is more appropriate than GARCH with annual data. The variance of the disturbance for the price equation is specified as multiplicative in explanatory variables [z.sub.t] = (1, [p.sub.t-1], x , [p.sub.t-4], t), and price equations are estimated as in Harvey (1976) using a BFGS maximum likelihood algorithm in Shazam. As before, the three indexes constructed from (23) are very highly correlated, with correlations r ranging between 0.9998 and 0.9873, and are highly correlated with the Laspeyres index (4), with r between 0.9628 and 0.9147. These indexes are somewhat less highly correlated with the index [VP.sup.B] (based on a multiplicative heteroscedastic model of a Divisia price index), with r ranging between 0.8742 and 0.8191.

Since GARCH models are inappropriate with annual data, monthly data were also collected for Manitoba crop prices. Monthly price data for wheat, barley, canola, oats, and flax were obtained from January 1990 to July 2005 (Agriculture and Agri-Food Canada, Winnipeg). Correlations between market prices and coefficients of variation are presented in table 4. The high correlations between prices again indicate the importance of including price covariances in aggregation of price risk over crops.

For simple autoregressive models of price, homoscedasticity is rejected at the 0.01 level for ARCH (Engle Lagrange multiplier test) for all crops except canola. However, homoscedasticity is seldom rejected for multiplicative heteroscedasticity (Harvey 1976, 1990). So we focus on GARCH models with monthly data.

A multivariate GARCH model is estimated by the two-step process assuming constant conditional correlations as presented above. First, univariate GARCH models are specified for each of the five prices. Current price [p.sub.i] is initially specified as a function of a 24-month lag on [p.sub.i], monthly dummies and a time trend, and a GARCH(1,1) model is estimated by BFGS maximum likelihood in Shazam. Price lags beyond 13 months are statistically insignificant, so longer lags are omitted. Various monthly dummies are statistically significant, so first-step univariate GARCH models are estimated without and with dummies. Then conditional covariances are estimated as before, and aggregate indexes of price risk are constructed as before.

Table 5 presents correlations between these price risk indexes based on multivariate GARCH and monthly data, and omitting dummies. Notation is similar to table 2. [VP.sup.L] denotes the Laspeyres index (4). [VP.sup.Torn], [VP.sup.Fish], [VP.sup.Lasp] denote the aggregate indexes of price risk constructed from (23) using Tornqvist, Fisher, and Laspeyres output quantity indexes [Y.sub.t]/[Y.sub.0], respectively. [VP.sup.B] denotes the price variance for the aggregate Tornqvist output price index. The first four indexes can be viewed as alternative valid index number approaches for aggregating price risk over commodities, in contrast to the last index [VP.sup.B].

The three indexes constructed from (23) are very highly correlated, with correlations r ranging between 0.9994 and 0.9978. These indexes are also highly correlated with the Laspeyres index (4), with r ranging between 0.9568 and 0.9536. However, all four of these indexes show smaller correlations with the last index [VP.sup.B], with r ranging between 0.8043 and 0.7810. Results in this table illustrate again that the theoretical contributions of this article are important in practice.

Table 6 presents correlations between these price risk indexes based on multivariate GARCH including monthly dummies in the first-step univariate GARCH models. Results are similar to the previous table. Here, correlations of [VP.sup.B] with the first four indexes range from only 0.6819 to 0.6663.

Conclusion

This article develops index number methods for aggregating price risk over commodities in production. The standard economic theory approach to construction of index numbers is extended here to price risk, and the resulting index numbers are closely related to modified Tornqvist and Fisher output quantity indexes. Given the multicommodity nature of price risk, these index number methods should have important empirical applications.

The methodology is applied to annual market price data for the six major Manitoba crops over 1950-2002 and to monthly price data for five crops over 1990-2005. Four index number formulas suggested by this article are compared with a variance of a Tornqvist price index, a common but in principle inferior measure of aggregate price risk. Commodity-level price risk is measured both from a naive expectations model and a two-step multivariate GARCH model assuming constant conditional correlations. Although we advocate modified Tornqvist and Fisher output quantity indexes in constructing our measures of aggregate price risk, data limitations required use of standard Tornqvist and Fisher output quantity indexes. Even so, results indicated that our four measures of aggregate price risk are highly correlated with each other and less highly correlated with the variance of the Tornqvist price index. These results illustrate the potential empirical importance of the index number methods for aggregating price risk over commodities.

Moreover, the economic index number approach to aggregation of price risk over commodities developed here has empirical importance in cases where the standard index number theory does not. Although the standard index number theory for aggregation of prices or quantities over commodities is in principle superior to Laspeyres or Paasche indexes, in practice correlations can be quite high (e.g., Diewert 1976). Similarly in this study, Tornqvist and Fisher output quantity indexes are very highly correlated with a Laspeyres index (table 2). Nevertheless, the aggregate indexes of price risk proposed here are not highly correlated with the standard index of price risk, i.e., a variance of an aggregate price index. In this sense, the extensions of index number theory to aggregation of price risk should be particularly important in practice.

Appendix

Proof of Proposition 1: Nonjoint CRTS technologies imply that the total cost function over all outputs can be expressed as

(A1) C(w,y) = [[summation].sub.i=1.,.m] [C.sup.i](w,[y.sub.i])

and marginal cost equals average cost

(A2) [partial derivative][C.sup.i](w, [y.sub.i])/ [partial derivative][y.sub.i] = [C.sup.i] (w, [y.sub.i])/[y.sub.i] i = 1,., m.

Define the following Tornqvist-like output quantity index [Y.sub.1]/[Y.sub.0]:

(A3)

log[Y.sub.1]/[Y.sub.0]) = [[summation].sub.i=1,.,m] {([theta].sub.i1] + [[theta].sub.i0])/2} log([y.sub.i1]/[y.sub.i0])

where [[theta].sub.it] [equivalent to] [C.sub.it]/[C.sub.t] is the share of output i in total cost at time t. This is similar to a standard Tornqvist input quantity index or output quantity index.

Proceeding similarly to Diewert (1976), we can show that this is a superlative output quantity index corresponding to a Translog cost function C(w, y) and nonjoint CRTS technologies, irrespective of risk aversion and output price risk or uncertainty. Since a Translog cost function C(w, y) is quadratic in logarithms, the quadratic lemma of Diewert (1976) implies

(A4)

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

by Shephard's lemma (assuming all inputs are in static equilibrium) and (A1)-(A2)

= log([W.sub.1]/[W.sub.0]) + log([Y.sub.1]/[Y.sub.0])

where [S.sub.wj] [equivalent to] {([w.sub.j1][x.sub.j1]/[C.sub.1]) + ([w.sub.j0][x.sub.j0]/[C.sub.0])}/2 and [S.sub.yi] [equivalent to] ([[theta].sub.i1] + [[theta].sub.i0])/2 is defined as for (A3). The two sums in the second equality of (A4) are designated as log([W.sub.1][W.sub.0]) and log([Y.sub.1]/[Y.sub.0]), respectively. The index [W.sub.1]/[W.sub.0] can be interpreted as an index for the ratio of an aggregate average cost [AC.sub.1]/[AC.sub.0] under our assumptions, by arguments similar to the standard case (Diewert 1976, 1980a). Then (A4) implies

(A5) [Y.sub.1]/[Y.sub.0] = ([C.sub.1]/[C.sub.0])/ ([AC.sub.1]/[AC.sub.0])

i.e., [Y.sub.1]/[Y.sub.0] is equal to ratio of total cost divided by an average cost index. This implies that [Y.sub.1]/[Y.sub.0] is an exact output quantity index, and it is superlative.

Proof of Proposition 2: Assuming a Translog multioutput cost function C(w, y) and proceeding similarly to (A4),

(A6) log([C.sub.1]/[C.sub.0]) = log([W.sub.1]/[W.sub.0]) +log([Y.sub.1]/ [Y.sub.0])

by Shep