Public Disclosure Authorized

Po Ry roh l *r,h
WORKING PAPERS

L DebtandIntenmtlonal
Finance

trntemational Department
Economics
TheWorldBank
Januery1992
WPS835
Public Disclosure Authorized

CommodityStabilization
Funds
Public Disclosure Authorized

PatricioArrau
and
Stijn Claessens
Public Disclosure Authorized

The optimal rule for depositsin and withdrawalsfrom a com-

moditystabilizationfund: keep the fundsmall - less than one
month'sexports. For the windfallgain oil exportersreceivedas
a resultof thePersianGulfcrisis-about fourmonthsofaverage
exports- the optionaldepletionperiod is aboutfour years. In
the long run, the exporter's fund shouldbe small, significantly
less than one monthof oil exports.
ThePolicyRese rchWoixngPapensdisseminatethefindingsof workinprogresand encouragetheexchangeofideasamongBankstaff
andall othemintered in developmentissues.Tlhe papersdistibuted bytheResearchAdvisoryStaff,carrythenamesof the authots,
re*lecatolytheirview,andshouldbeoed ndcitedaccordingly.Thendinp,inteapations,andconclusiomsatetheauthorsown.They
shouldnot be atibuted to theWorldBank,its Boardof Dimsors. its management,or any of its membercounties.
 Pdlky R.l.arch

Finance
DebtandInternational

WPS 835

This paper - a product of the Debt and Intemational Finance Division, Intemational Economics
Department - is part of a larger effort in the Department to contribute to a Bankwide work program on
issues related to developing country management of external risk, including currency and exchange rate
risk management, currency reserve management, and commodity price risk n,anagement. Copies of the
paper are available free from the World Bank, 1818 H Street NW, Washington DC 20433. Please contact
Sheilah King-WNatson,room S8-040, extension 31047 (36 pages). January 1992.

Commodity stabilization funds are hard-currency
savings to protect against a fall in income from
commodity exports in the presence of borrowing
constraints.
Arrau and Claessens develop the optimal
rules for deposits in and withdrawals from such a
fund by using a benchmark model of precaution-
ary savings with liquidity constraints.
They show that the optimal stabilization
fund is small. For the Chilean Copper Stabiliza-
tion Fund, they show that the actual accumula-
tion of foreign assets has been much larger than
the benchmark model requires. Over long
periods, the copper fund should contain less than
one month's exports.
They also use the model to find the optimal
depletion of the windfall gain oil exporters
received as a result of the Persian Gulf crisis-
amnountingto about four months of average
exports. They find that such a windfall gain
should be depleted in about four years. In the
long run, an oil exporter should keep a small
fund, significantly less than one month of oil
exportc But higher-than-predicted funds can be
justified if there are externalities associated with
the fund, frictions in the economy, or the bor-
rowing constraint is relaxed.

The Policy Research Working Paper Series disseminates the fmdings of work under way in theBanrk An objectiveof the series
is to get these findings out quickly, even if presentations arc less than fully polished. The fimdings, interpretations, and
conclusions in these papers do not necessarily represent official Bank policy.

Produced by the Policy Research Dissemination Center

 TABLE OF CONTENTS

1. INTRODUCTION . ..
........................................... 2

2. INCOME FROM EXPORTS OF COPPER AND OIL . . .... 4

2.1 Copper and oil Prices.* .. ....... . .. . ... .. . ...... 4

2.2 Exports Revenues .... ... **** *

................................... 6

3. DEATONS PRECAUTIONARY SAVING MODELNODE.. .. .... o.......... . . 7

4. APPLICATIONS... .. ..... . .... . .. .. . ........ 13

4.1 Chile's copper stabilixation rund. . ........ 13

4.2 The 1990-91 oil windfall gain.. .. ... 19

S. EXTENSIONS.
....... .. .. .... .. .. . ................... . .o 20

n6. ...........................
coNCLUsIONS*-*-*-- *........ *.................... vv
-.- 22

Reference-n....... . .... 24

Appendix A .. .......... . ............. 26

Appendix ............................. B 28

I-b e . . . .. . . . . . .. . . . . . . .. . l0 . . .. . . . . . 29

Figures .......... .......................... 33

'We would like to thankRon Duncan, VikramNehru, and Larry Summersfor their commentsand Heinz
Rudo!ph for able research assistance.
 2

1. INTRODUCTION.

Many developingcountrieshavelarge exposuresto commodityprice risk becauseof concentrated

export structures. Two methods are availableto reduce these large exposures and/or smooth income

fluctuations resulting from commodity price movements: 1) self-insure; and 2) transfer risk to

international (capital) markets. The first can be achieved in a variety of ways including: (a)

diversificationof the export structure and, (b) accumulationof foreign assets. The second can be

achievedthrough a combinationof borrowingand lending in internationalcapital markets, and through

commodityprice-linkedhedging instruments. In principle, the second method is usuallyprefe-red: a

developingcountrydoes not have a comparativeadvantagein bearingprice risk, while the international

(capital)markets at large are better able to do so.

However, developingcountries' access to the internationalcapital markets has proven to be far

from full. Borrowingfrom internationalmarkets has often proven to be procyclical rather than (the

desired) countercyclical. Contingent trrowing facilities (such as the IMF's Compensatory and

ContingencyFinancingFacility(CCFF) and the EC's STABEX)have the disadvantagethat they are often

only available ex-post. Lack of creditworthiness,in general, tends to restrict the access of developing

countriesto internationalcapitalmarkets. Accessto comniodity-linkedhedginginstrumentsis, for similar

reasons, limited. In addition, for many of the commoditiesfor which developing countries have

significantexposure, the markets for long-termhedginginstrumentsare still very limited, e.g., only for

metals and energyproducts is there a fairly completespectrumof hedging instrumentsavailable;long-

term markets for tropicalcommoditiesare non-existent.'

The limited accessto internationalborrowingsand the incompletenessof the commodity-linked

'In addition, many developingcountrieshold a large share of world production of some of these
commoditiesand the amountsto be hedgedare often very large comparedto the size of existinghedging
markets (for instance, Mexico's yearly oil exports represent many times the daily turnover on the
exchangefor oil futures).
 3
hedging markets imply that to some extent developingcountries will have to rely on methodsof self-

insurance. Of these methods,exportdiversificationwill often take longerto achieveand may not be the

most efficient since it may run counter to the country's comparative advantage. A self-insurance

mechanismthat may be a relatively efficient alternativeis the accumulationof foreign assets by the

country to act as a commoditystabilizationfund (CSF). A fund like this was establishedin Chile in

1985. During periods of high commodityprices--and high exports earnings-the country would

accumulateforeign assets which it woulddraw down in periods of low commodityprices. The problem

would thus be very similar to that of a liquidity-onstrained individualwho also has a demand for

precautionarysavings.

The purpose of this paper is to estimatethe processesfor incomefrom exportsof copperand oil,

and to derive the optimal rules for a CSF using as a benchmarklDeaton's (1991) precautionarysaving

model. The problem of designinga CSF is find the optimalrules for deposits and withdrawalsgiven a

particular state of nature. The rules will thereftre dependon the stochasticcharacteristicsof the income

process. Few papers exist on this topic. Hausman and Powell (1991), who assume a random walk

process for prices, justify the creation of a fund on adjustmentcosts in the productionprocess of the

commodity. Basch and Engel (1991) combine the predictions of a large-scalecopper model with the

policy functionsderived by Deaton A) when incomeis assumel stationary. Their workdiffers from

ours in that we derive the policy functionsin accordancewith the incomeprocess estimated.

The setup of this paper is as follows. In section two, we determinethe stochasticprocess for

incomefrom copperand oil exports. In sectionthree, we presentthe theoreticalmodel(basedon Deaton,

1991) and determine the optimal saving rule given the estimated income process. Our anaiysis

complementsDeaton's analysisas we study in-deptha special case-an AR(1) incomegrowth process-

which is discussedonly brieflyby Deaton. In sectionfour, we applythe resultsto the situationof Chile-

and its Copper StabilizationFund-and derive the optimaldrawdownrules for the windfallgain recently
 4
receivedby oil expoiters. In sectionfive, we presentsome possil"e extensions. The last sectioncontains

our conci-' ions.

2. INCOME FROM EXPORTS OF COPPER AND OIL

2.1 Copperand oil prices

The stochasticprocess of income from commodityexports dependson the stochasticprocesses

of commodityprices and volumesexported. From the country's point of view, volumesare, however,

subject to considerablyless uncertaintythan commodityprices and, at least in the short run, are more

a choice variable. We therefore first model the process of copper and oil prices and then infer the

process for income by incorporatingthe volumecomponent. We then verify that we have appropriately

modeledthe income process.

We start with stationaritytests on the prices of copper and oil. Table 1 summarizesthe result

of the tests. Tests are performed on prices in logarithmic-levelsand !ogarithmicfirst differences, in

nominal as well as real terms. To obtain a real price of copper, the monthly dollar copper price

(London)for the period 1960-90is deflatedby the US CPI index. For oil prices, we use two nominal

prices: the West-Texas Intermediate(WTI) price and the actual price of Mexicanoil exports. The

former is deflatedby the US CPI indexand the latter by the price indexof importedprivate consumption

goods.2 Table 1 makesclear that the (log) level of real copper and oil prices are likely better modelled

as I(l), i.e., non-stationaryprocesscs, and that the logaritmic difference is stationary. Only when 12

lags are int oduced in the augmentedDickey-Fullertest is there evidencethat the copperprice couiuibe

stationary and revert to some mean. On the contrary, for the WTI price series, the ADF(12)test even

questionsstationarityin logarithmicfirst differences,but the test is probablyof too low a power as the

series is short (83 observations). From Table 1, and giventhe time horizonwe are interestedin and the

2 We also used the price indexesfor total Mexicanimportsand for importedgovernmentconsumption

goodsto deflatethe Mexicooil price and found virtually identicalresults.
 5

difficulties of incorporatingautocorrelationof high order in the model, we conclude that the level of

prices s best modeledas non-stationaryin levels and stationary in first differences.

Next we modelthe logarithmicfirst differenceof prices. The randomwialkhypothesisis a logical

starting point here. However, Williams and Wright (1991), Deaton and Laroque (1991) and Trivedi

(1991) find that typically real commodityprice series do not behave like a randomwalk and that there

is more linear and nonlineardependencein the first differenceof prices thanis consistentwiththe random

walk hypothesis. Part of this behaviorcan be explainedby the competitivestoragemodel (Williamsand

Wright, 1991, and Deaton and Laroque, 1991)which indicatesthat price jumps due to "stock-outs"lead

to skewed price distributionsand serial dependence. Trivedi (1991) conjecturesthat more generalized

versions of the competitivestorage modelimply (non-linear)AR processesfor commodityprices. Given

the applicationwe have in mind, we modelthe commodityprice thereforeas an first order autoregressive

(AR(1))process.3 Consequently,we estimate:

(AP, -,,) -,P(AP.-1 -po) (1

where:
Pt = the log of the (real or nominal)commodityprice,
p = an autoregressiveparameterless than 1,
Apt = Pt - Pt-I,
PIP = unconditionalmean drift of prices.
et = normal, independent,and identicallydistributed (i.i.d) error with a zero mean and a
varianceof a2.

Table 2 summarizesthe estimationof (1) for the copper and oil price, both in nominal and real terms.

The table shows that there is a small (inflaticnary)drift in the nominal prices and no drift in the real

prices. Except for the drift term, deflatinghas virtuallyno effect on the other parameterestimates(both

the autoregressiveterm as wellas the estimatedvarianceof the error are unaffected). This indicatesthat

3 Using an AR process is consistentwith rational expectations(e.g., a no-profitscondition in the

futures markets)because of thleasymmetryintroducedby stockouts.Furthermore, using an AR process
biases our results towards findingprecautionarysavings(see footnote2).
 6
all real price volatility is due to commodityprice volatilityas the deflatorsare very smooth.

2.2 ExportsRevenues

AlthoughpiIce ;movementsexplainmost of the unexpectedmovementsof incomefrom exports,

a full descriptionof the incomeprocessrequiresus to modelthe processfor the volumeexportedas well.

The volume of exports depend on production as well as on the inventory decisions. In principle,

inventoriesare saving in an asset with return that dependson the price next period (net of any storage

cost and depreciation). However, the joint portfolio-savingdecision of production, inventories and

financial assets is beyond the scope of this paper.' We concentratetherefore on the income (foreign

exchange)earned from exports and do not make a distinctionbetweenproductionand exports (abstracting

thus from domestic consumption).

The simplestway to modelthe volumeof commodity(exported)is to assumea deterministictrend

process. Table 3 shows the estimationof such a deterministictrend (using annualdata), and Appendix

B provides plots of all the series. For Chile, we find that a trend variable can explain 90% of the

variance in the volumeof copper exports (and 97% of the variancein production)(see also Figure B.2).
5 For Mexicoand Venezuela,
Only a small part of the varianceof volumeexportedremainsunexplained.

however, modellingthe volumeprocessby a trend only does not seem appropriate. Figures B.4 and B.6

show importantstructural changesin the volume of oil exported. Large discoveriesin Mexico in 1976

led to high volume growth in the followingten years until--in part as a result of fiscal adjustments-

exports stagnatedin the eighties. For Venezuela,exportsin the secondhalf of the seventiesdeclineddue

to OPEC quota's, but stabilizedthereafterat that level.

Since volume variability may influence (offset) price variability, we verified the two-step

' For a model of commodityprice determinationwith storage see Deaton and Laroque (1991) and
Williamsand Wright (1991). See Choe (1991)for an analysisof the precautionarydemand for physical
stocks in the presenceof consumptionor productionshocks.

5 The low Durbin-Watsonsuggeststhat the copper exports could be non-stationary.

 7
derivationof the incomeprocess by directly estimatingan AR(1)process for real (dollar) incomeusing

monthlydata.' The parameter estimates(reportedin Table 4) were quite lifferent from those impliedby

the two-stepderivation:the estimatefor p in caseof the real incomeof copperexports wasnegativewhile

it was positive for the copper price equation; and Mexico's real income itom oil exports resemblesa

random walk while for the real oil price an AR(1)was appropriate. The reason for these differencesis

the month-to-monthbehavior of the volume of exports which exhibits strong negativeautocorrelation.

We suspectthat this is largely the resultof seasonalityand other (physical)customsrelatedto the shipping

of the commodities. We have little reason to believe that next month's volume of exports cannot be

predicted accurately by the countries. We considerthe volume of exports accordinglynot a source of

risk and we use the two-step derivation for the income process. Future work should, ho;vever,

investigatewhy the month-to-monthvariability is much higher than that impliedby the yearly data.

For Chile, an AR(1)for prices and a trend for volumesprovidesthus a good explanationfor the

stochasticprocess for income from exports. The incomeprocess has the same distributionas the price

in (1), but with the additio,

1 al drift term "'jq" in Table 3 added to the (zero) drift of the real price (J,~in

Table 2). For both oil exporters(where we only have a reasonableprocess for price fluctuationsdue to

structural breaks in volumes during the sampleperiod) we simply assume a deterministictrend for the

future volume of exports.

3, DEATON'S PRECAUTIONARY SAVING MODEL

The benchmarkprecautionarysaving model used here is based on Deaton (1991). He derives

the implicationsfor the optimalsaving rule under stationaryas well as non-stationaryincomeprocesses.

In light of the empirical evidencepresented in the previous section, we present an abbreviatedversion

of his model under a non-stationaryAR(1) income process (the interestedreader is referred to Deaton

'Stationaritytest were also performed (not reported), which indicatedthat income is best modelled
as 1(1).
 8
for a full derivation).

The precautionarysavingsmodelis based on the notionthat the individualor countryfacessome

constraintson its borrowingswhich prevent it from smoothingadverseincomefluctuations. This is very

realistic for many developing countries, including Chile, Mexico and Venezuela. Even though these

countrieshave beenreceivingsomeprivate voluntarycapital inflowslately--inparticularMexicanprivate

and public companieshave gained renewedaccessto international(bond)markets--mostof this renewed

access has been project or companyspecific. There is no indicationthat the govern-mentsof these (and

many other) countriesare able to use the internationalcapital .narketsto smoothadverseincomeshocks.

We assume that the governmenthas an objectivefunctionwhich is a functionof a sequenceof

expenditures-governmentconsumption,provisionof publicgoods, a socialprogram, etc.--to be financed

with the income from commodityexports. We assumethat all other expenditures,asset accumulation,

and policy decisions by the government can be separated from the saving and expendituredecisions

related to commodityincome.' We further assume that the governmentis risk averse. Becauseof the

volatile income stream from commodityexports and the borrowing constraint, the governmenthas an

incentiveto set up a fund (CSF) to smooth income fluctuationsoptimally. Formally, the government

maximizes the expected value of a function of the control variable, saving, subject to budget and

borrowingconstraints.

V, = (1+,,6s,IBLFU(G,) I

s.t. F,, 1c(1+r)(F,+Z,-G,) (2)

F,kO

where:

7 Alternatively,we could have assumedthat all unexpecteddeviationsof total incomecan be attri',uted

to commodityprice (income)fluctuationsor that all other expenditures,asset accumulation,and policy
decisions of the government can be separated. As a third option, we could have assumed that the
government acts benevolently on behalf of the country's risk-adverse citizens who are borrowing
constrainedthemselves.
 9
E = the expectationoperator,
u(.) = the objectivefunotion,satisfyingthe usual (concavity)conditions,
G, = sequenceof govermmentexpenditures,
X, = income from exports,
F, = assets in commodityfund,
6 = rate of time preference,
r = internationalinterest rate.

Becausethe income process X, is non-stationary,the model is best normalizedby dividing all variables

by the incomein that period. The iower case letters indicatenormalizedvariables, i.e., x,+, =

g,= Gt/X,; and f, = F,/X,. By specifyingthe utility functionto display constantrelative risk aversion

(CRRA) (i.e., u(g) = gl1 /(l-a), where a is the relative risk aversionparameter),the normalizedmodel

can be expressed as the solutionto the followingtwo equations:

`= max{ (1+f*) (ler)

I x,+, 3) (3)

5(1r) ( 1f,-g,) (4)

5.1~~~t.

Equation3 is the (modified)Euler equation. Becauseof the borrowingconstraint,the expenditure

ratio in the current period has to be less than 1 plus the fund ratio, that is g S (1 + f). If the

borrowingconstraintis not binding,then g, is such thatthe marginalutilitytoday is equalto the expected

marginalutility tomorrow appropriatelydiscounted. Becausemarginalutility is convex, precautionary

saving will result and consumptionwill be less than under perfect certainty. Equation(4) is the law of

motion for the fund ratio.

For a given process ';: x,+1, we can find a policy function p(f, x) which maps the two state

variables f, and x, into the control variable g, = p(f,, xe). From the previous section, we know that

(conditionalon informationin period t) x,+, is lognormallydistr6uted with conditionaldistribution

N(u(l-p)+plog(x.), o2) FollowingDeaton(1991), we approximatethe normal distributionfor log(x,+

1)
 10

with a ten-value, discrete Markov process.' Therefore, we numericallycomputeten policy functions,

g = p(f, i), i = 1,... 10, which map the fund ratio onto the expenditureratio-given that state "i" was

realized. AppendixA describesthe btpnsto obtain the ten policy functions.

Replacingthe policy functionsin (3) and using (4), the (ten)equilibriumconditionsof th' model

can be written as follows

al,, ) axf(1f.-', X-- (1+3) g(5)

where wrjis the probability of state j occurring in the next period giver that state i occurred in this

period. In this formulaLion,x,+ takesone of ten discrete values. We checkedthis approximationto the

income process by simulatingthe discrete process for Chile's copper exports and verified that the

simulatedseries provides estimatesfor the income process very similar to those of Table 2 and 3 (see

furtier Appendix A).

Figure 1 showsthe ten policyfunctionscorrespondingto the aboveincomeprocessand parameter

values et, 8, and r of 2, 0.08, and 0.04 respectively. We start with ci = 2 mainl; because estimations

of the degree of relative risk aversionfor the US economyhave found the value 2 to be in the lower

range. The policy functionsare upward sloped,mainlybecausea higherfundratio implieshigherinterest

earnings on the foreign assets which allows for a higher expenditureratio. The policy functions

corresponding to good states lie above those correspondingto bad states; because of the positive

autocorrelationin the growth of income, a good state indicatesa higher likelihoodof more good states-

and thus a higher (permanent)income-leadingto a higher ratio of consumptionto assetsat hand. A bad

' Note that using a (discrete)ML .ov process allowsfor easy approximationof other, non-symmetric
distributionsas long as they are of first autoregressiveorder.
 11
state, however, results in increasedsaving in anticipationof likely lower future income growth.9

We can see from Figure I that this combinationof parameters leads to little saving. Savings

would only take place if either the worst state or next to worst state is realized and when there is no

moneyin the fund. As soonas the fund reaches5% of incomeall policyfunctionsindicatethat spending

exceeds income (an expenditureratio greater than 1). In this structure, the fund will be depletedover

time.10 We have also used a very high monthlyinterestrate and social rate of time preference (A% and

8% respectively)for reasons that will be spelledout below.

It is clear from Figure 1 that higher saving will only result if most of the ten policy functions

have an expenditureratio below one for low fund ratios. We tried several parameter combinationsto

obtain this. Figure 2 shows the sensitivityof policy functions (for the worst state only) for different

vaiues of the CRRA, interest rate and time preferenceparameters. The figure showscases of a between

0.5 and 0.8 only, which is very low comparedwith estimatesobtained for the US. For high values of

a and reasonable interest rates, we found greater difficulty in obtaining convergenceof the numerical

algorithm. To understand this, we need to resort to the sufficient condition for the existence of a

stationary, stochasticrational expectationequilibriumin this model. As shown by Deaton (1991), this

conditionfor the i.i.d growth case is: E[(1 + r)/(l + 6) x+, 1] < 1. For simplicity,we will discussthe

unconditionalversion of our AR(1)growth case. Using the approximationx = log(l +x), taking logs,

and using the unconditionallognormaldistributionfor xt+1 (AppendixA), we can express the condition

as:

9This countercyclicalbehavior of saving is emphasizedby Deaton (1991) who states that this is a
counterfactualimplicationof his model. However, recent (Latin American) experienceshows often a
consumptionboom and a drop in saving in the beginningof recovery (Mexico in 1989and 1990is the
most recent exan.ple).

10 Strictly speakingwe must also considerthe differencebetweenthe rate of incomegrowth and the
rate of interest as we do below. The statementabove is still accurate,however.
 12

a 1 '(r-6) + CT,__ < ' (6)

2( I ~p2)

The left hand side of equation(6) is the expectedslope of the expenditurefunctionof the unconstrained

problem while the right hand side expressionis the expectedgrowth in income. If the former is lower

thanthe latter, an unconstrainedgovernmentwouldlike to obtaindebt and the borrowingconstraintwould

be binding. Note that the expenditureslope (left hand side of equation (6)) comprisestwo terms. The

first term is the life cycle componentof savings and is the slope of expenditure in the absence of

uncertainty. The term is positivefor r > a and vice-versa. The importanceof this term dependson the

inverse of the CRRAparameter a (the intertemporalelasticityof substitutionin this framework). The

second component of the expectedslope of expenditureis motivatedby precautionarymotives. The

higher the volatilityof income, the higher the precautionarymotive impliedby the convexnature of the

marginalutilityof consumption,and the higherthe expectedslope of expenditureon accountof this term.

Becausethe volatilityof income is so large for commodityexporters-as shown above, a value

of a as high as 2 generates a large desire for precautionarysavings and requires a large differential

between the interest rate and time preference parameter to satisfy condition (6). But, the required

calibrationof the modelwas not alwayspossible. For reasonableinterestrate and time preferencevalues,

no convergencyof the policy functionsresultedfor a = 2. For instance, in the case of (a, 6, r) = (2,

0.12/12, 0.05/12), even the (annualized)time preferencerate which was7 percentagepointshigherthan

the interest rate was not enoughto compensatefor the desire to raise precautionarysavings.

To furtherevaluatethe largeuncertaintywe observe, considerthe followingcomparison. Deaton

(1991) presents a simulationfor US quarterly data with an (AR(1))growth process for income and a

CRRAparameterequalto 2. For his simulation,incomegrowth fluctuatedbetweena maximumof 2.5%

per annum and a minimumof -0.5%. In the caseof copper, (monthly)incomegrowthprocessfluctuates

between-12.0% and 12.7%, a much larger range. The difficultiesfaced in calibratingthe model with
 13
CRRA parameters higher than 1 are therefore predominantlydue to the high volatility encounteredin

commodityprices."'

Another compellingreason to use a low a is evidencefrom Euler estimationsfor these countries

which suggests a lower value than the one typically estimated for the US."2 The availableempirical

evidence,combinedwith the fact that lower valut.zof a still implylarge precautionarysavings,reinforce

3 The net effect of the high volatility for
the notion that a low value of a is appropriatein this context."

commodityincomeand a somewhatlower a can still generatea higher demandfor precautionarysavings

than Deaton's simulationswith US data imply.

In the next section we apply the model to the Chilean copper stabilizationfund and the recent

windfall gain for oil exporters.

4. APPLICATIONS

4.1 Chile's copper stabilization fund

Chile's exports are highly concentratedin copper. Even though, as a result of govermnent

policies, Chile has reduced its dependenceon copper significantlyover the past decade, still more than

" Of course, we are limitedby the time-separableutilityfunctionwe use where high RRAparameters
imply low intertemporalelasticity. Obviously,there would be great benefit from disentanglingthe two
parametersas in the non-expectedutilityframeworkof Epsteinand Zin (1989, 1991),Farmer (1991)and
Weil (1991).

Althoughthe empiricalevidenceis limited in this regard, estimationsfor Chile and Mexico-in the
12
context of a monetary model-suggest a point estimate for a of 0.63 for Chile and 0.35 for Mexico
(Arrau, 1990). These estimatesare not very precise-for instance, a value of 1 for Chile could not be
ruled out.

'3In fact, becausevolatility is so much higher in Latin Americathan in the US, the low estimatefor
axcould very well be the result of a misspecifiedmodelthat cannot accommodatesuch a high volatility,
and therefore results in biased estimatesfor a. Applicationsof the Euler approach to developing
countries with annualdata (Ostry and Reinhart, 1991) result in valuesof a of the order of 2. Because
annual data is less volatile than quarterly or monthlydata, it seems plausible that the estimateof this
parameter might be driven by the volatilityof the series employed in the estimations. Applicationsto
Mexico of the more flexibleframework of the Epstein-Zinmodel (Arrau and van Wijnbergen, 1991)
show that the RRAparameter is around 1, in concordancewith estimationsfor the US (Epsteinand Zin,
1991; Giovaninniand Weil, 1990).
 14
50% of exports comes from this source. Furthermore, since most colpperis produced by state-owned

mines, it is a very importantsource of reverue for the Treasury. Volatilecopperprices can also 'fect

the competivenessof the rest of the Chileaneconomythrough movementsin the real exchangerate.

Chilean authoritieshave long been aware of the importanceof managingthis risk. In 1981they

establishedrules governingthe treatmentof excesscopper revenues. In 1985they establisheda Copper

StabilizationFund (CSF) as part of a StructuralAdjustmentLoan agreementwith the World Bank. The

first deposit into the CSF was subsequentlymade in 1987. The rules of the CSF stipulatethat deposits

(or withdrawals)will be proportionalto the excess of the copper price over trigger prices which are

establishedas two bands (narrow and wide) around a reference price. The reference price is set in real

terms (adjusted for dollar inflation) and cannot exceed a six year moving average of the spot price.

Within the narrow band there would be no deposits or withdrawals;outside the wide band all excess

copper revenues wouldbe deposited(if price is above)or withdrawn(if price is below); and in between

the two bands 50% of the excessshould be deposited(if above)or withdrawn(if below). Furthermore,

withdrawalswere only to be used for "extraordinaryamortizationsof public debt".'4 Table 6 presents

figures correspondingto two differentmeasuresfor the fund. The first measureconsidersas withdrawals

only current expenditures,i.e., the subsidyto grape exportersin 1988-89to cover the damagescaused

by the US ban on grape exports from Chile in the wake of al!egationsabout cyanide poisoning. The

second measurealso includesas withdrawalscapital expenditures,i.e., repurchasesof externaldebt by

the Central Bank in the secondarymarket.

The CSF has clearlyprovideda mechanismfor avoidingunwantedpoliticallygeneratedpressures

14 The latter qualificationhas led to some confusionaboutthe exact nature of the fund-a Treasury
fund or part of foreign reserves-and severaldefinitionsare used to measuredepositsand withdrawalsin
the fund, includingsome coveringthe accountingtransactionsbetweenthe Treasuryand the Central Bank
(e.g., prepaymentof debt by the Treasury to the Central Bank), which consequentlyshows a very low
balance for the fund at the end of 1990. It should also be notedthat there is some lack of clarityas to
the final purposeof the CSF; officialdocumentsindicateits purpose to be "stabilizationof fiscal income
from copper receipts" as well as "stabilizationof the real exchangerate".
 15
to spend funds accumulatedduring periods of unusuallyhigh copperprices. From an economicpoint of

view, however, the mechanicsappear (in addition to being confusing)to lack a solid foundation. For

instance, the reference price is set withouttaking into accountthe unde.rlyingprocess generatingcopper

prices formation. Furthermore, the rules for accumulationand withdrawaldo not considerthe level of

the fund as the model above does."

Two different parameterizationsof the benchmarkmodel are used to evaluatethe performance

of the CSF. In order to make the conclusionsmore robust, we use the two parameterizationsmost

favorable to saving in Figure 2. As mentionedabove, this implies a value for u = 0.5, and a = 0.8

(Cases I and 2 res,-.ctively). In both cases, the annualizedinterestrate is 5%. However, in Case 1, the

time preference is assumedto be equal to the monthlyinterestrate minus 0.05%, which leads to a small

desire for life cycle saving. In Case 2, the rate of interest and socialtime preference are assumedto be

equal. It is importantto emphasizethat the borrowing constraint is never binding in Case 2.16 This

impliesthat this case can only be valid if one believesthat a borrowingconstraintdoes not constitutean

essentialpart of a model for Chile, and its results are biased towardshigher savings.

We first discuss the results of Case 1. Figure 3 depicts the ten policy functions for this

configuration. The very flat policy functionsfor modestly-highfund ratios indicatelittle change in the

expenditureratio as incomeincreases. This is mostlydue to the low interest rate assumed, whichresults

in little interest income and thus a low expenditureratio even for a higher fund ratio. Close to a zero

fund ratio, however, the slopes of the policy functionsincrease sharply as a result of the nonlinearities

introducedby the borrowing constraint. Moreover, the five worst states of nature generate saving for

all fund ratios, while the five best states of nature generate dissavingfor almost all fund ratios. Since

"Independently,Btaschand Engel (1991)make this pointtoo.

'"All ten policy functionsfor Case 2 intersectthe vertical axis below 1 and thereforethe borrowing
constraintis not bindingfor any.
 16
every state of nature has equal probabilityof occurrence, we can see from Figure 3 that for high fund

ratios the unconditional(average) policy function implies dissaving. The fund ratio can therefore not

increase forever but will be mean-reverting.

A typicalsimulationfor 200 periods using these policyfunctionsis shown in Fig Ire 4. We plot

the growth factor "x" (plus one for clarityof presentation),the expenditureratio, and the fund ratio. As

can be seen, the fund ratio is stationary, as for high values of the fund, dissaving occurs

(unconditionally). For this particular simulation,the averagv fund ratio is about 0.30 (30% of one

month's income).

We repeatedthe simulationlike the one underlyingFigure 4 1000times. Table 5 presents some

data based on these 1000replications. The results show that the average fund ratio is higher, 0.66, with

a standarddeviationof 1.344. The averagefund ratio is misleading,becausethe fund is boundedbelow

by zero. We therefore also present the medianfund ratio, 0.383. The importantresult to note is that

both the mean and median fund ratiosare very small, implyingless than one monthof export incomeheld

in foreign assets. So even with a parameter combinationthat leads to a relatively high desire for

precautionarysavings and litde life cycle saving, the optimalfund is relatively very small.

However, a small fund may still result in large smoothingof income. Table 5 shows the mean

and standarddeviationof the drift of the expenditureratio, which can be comparedto the parametersof

the incomeprocess (Table 4). Table 5 showsthat the standarddeviationof expendituresis 40% less than

the standarddeviationof income, indicatinga considerabledegree of smoothingfor such a small fund.

Table 5 also presentsstatisticson the marginalpropensityto spend (consume)(MPC)out of incomefrom

exports (excludingthus, interest income), which is obtainedfrom linear regressionsin levels as well as

in log differencesfor every Monte Carlo replication. The regressionin levels suggestsa MPC equal to

one, which is not optimalin our framework. One wouldexpect improvementin results by using a linear

specificationin levels, since the cointegratednature of the regression would provide more consistent
 17
estimates. However, as the low Durbin-Watsonstatistic suggests,the borrowing constraintintroduces

an asymmetry in the errors (consumptionis not symmetric)that casts doubt that this is a cointegrated

equation. The MPC derived from the regressionin log differences (with well-behavederr.rs) provides

the correct answer. The correct marginalpropensityto spend out of incrementalincome is thus about

0.56, not one.

Finally, we applythe modelto Chile usingthe actualcopperprice realizedsince 1987. We first

classifythe actualprices betweenSeptember 1987and December 1990to one of the ten statesof nature

(details in AppendixA). Usingthe appropriatepolicy function,we then computethe optimalfund ratio.

This is done for both cases describedabove. For Case 1, Figure 5 shows the evolutionof actual income

growth x (plus one for clarity of presentation),and the optimal fund and expenditureratios for every

month. Based on this analysis,the fund shouldoptimallyhave amountedto about 20% of one month's

worth of export income at the end of 1990. Case 2 is depicted in Figure 6. As noted above, this

parameter combinationimpliesthat the borrowingconstraint is irrelevantbecauseof the absenceof the

desire to borrow for life cycle motives and because of the high level of desire to save for precautionary

motives. Consequently,in Case 2 the fund is never depletedcompletelyduring the period 1987-1990.

Even in this high-savings,no-borrowingdesire case, the fund should have reached a level equivalentto

only 2.2 times monthlyexports at the end of 1990.

Table 6 comparesthe actual deposits, acccrdingto the two definitionsof the fund spelledout

above, and the two simulationcases. As we can see, the stock of the fund was between$2.4 and $1.9

billion by 1990, depending on the assumptions made about capitalizationof interest (zero and 2%

quarterly respectively). These levels correspondto about six to seven times the average monthlyincome

in the last quarter of 1990 (and about 25% of annualtotal exports). The third and fourth columnsare

computed by applying the derived optimal monthly saving ratio to the actualmonthly exports from

October 1987to December 1990, and then cumulatingthe monthly saving for each quarter. By using
 18
actualmonthlyexports, we preventthe high month-to-monthvariability in the volumeof exports (which

we kept as certain in the theoreticalmodel) from influencingthe differences betweenthe optimalfund

and the actual fund.

For Case 1, the model suggeststhat the fund should optimallyhave been between $60 and $70

milliondollars at the end of 1990(20% to 22% of monthlyincome). However,the actualfund was six

to seven times monthlyexports, or more than thirty times as large as predictedby the benchmarkmodel

under Case 1.*7 Even in Case 2 (the non-borrowingconstrainedbenchmark),the actual fund was only

three times as large as desired (between $580 and $640 million). Clearly, from the point of desired

consumptionsmoothing, there has been a substantialoveraccumulationof foreign assets in the CSF

through the end of 1990.13

The optimal level at the end of 1990does not necessarilyrepresent the optimal long run fund

level. Equation (4) canbe rewritten, by substitutingthe expectedgrowth rate for income(1 +,) for x,+,

and setting f,+1 equal to f,, as one of two conditionsthat define the stochastic, stationarysteady state.

u =
- 1- ( -r) (7)

The secondconditionfor the steady state is the (unconditional)policy function,p(f) = £, 0.1 *p(f, i), the

state weightedaverage of the equation(5) policy functions. Figure 7 plots these two functionsfor the

parametersof Case 1. The intersectionof the two functionsrepresentsthe steady state value of f. In

Case 1, the steady state ratio of f is about 0.30-close to the median value in our earlier Monte Carlo

17Thisdiffers from
Baschand Engel(1991)who-assuming stationaryprices-find that the fund should
have been between $835 and $1649 millionat the end of 1990 (in constant 1989dollars).

'3 There may be one caveat to this observation. Foreign exchange reserves are not held for
precautionaryreasonsonly: for example,transactiondemand(to cover imports)also gives rise to demand
for rew-ves. If, during the perixd that the Chilean authorities accumulateda too large CSF, other
demandwas reduced, then the overaccumulationwould have been less. However, then the CSF needs
to be defineddifferentlv.
 19

simulations(Table 5). Note that the stationarystochasticsteady state for f is stable as the slope of p(f)

is higher thanthe slope of (7). Thereforefor f valueshigherthan 0.30, the unconditionalpolicyfunction

impliesan expenditureratio higherthan the one compatiblewith a stationaryf and the fund wouldbe run

down (in an expectedvaluesense). For valuesof f lower than the equilibriumvalue, the oppositeis true

and the fund increases(in an expectedvalue sense).

For Case 2 (not shown), the state steady ratio of the fund is much larger, about 16, confirming

the higher desire for savings under this set of parameters.

4.2 The 1990-91oil windfall gain

The increase in oil prices from August 1990to January 1991due to the Gulf Crisis represented

a windfall gain for oil exporters. The income gain for a typical oil exporter can be approximatedas

being equal to four months of normal exports (for example, in the case of Mexico, prices were on

average 80% higher than in the previousyear during roughlyfive months). Both Mexicoand Venezuela

did not spend the windfall, but investedit (Mexicoused part of it to retire public debt and Venezuela

establishedan oil stabilizationfund(alongthe linesof the ChileanCSF)). The questionarises, how much

of the windfallshouldbe investedand how it shouldbe spent over time. As we showed in the previous

section,the optimalfund size for copperamountsto only about0.2 monthsof imports. The oil windfall

is much more than that: four months. This does not mean, however, that it would be optimalto deplete

the fund quickly.

Questionsregardingthe optimalspe"d at which to depletea windfallgain of this magnitudecan

be determinedusing the benchmarkmodel, given the estimatesfor the Mexicooil income process. In

accordancewith the results of section2, the real price is held constant. Since, as was shown earlier, it

is not clear what is the volume drift for Mexico (and Venezuela), we solve the ten policy functions

assumingp = 0.003, or 0.3% export volumeincrease per month. We use values for other parameters

as follows: a = 0.5, 6 = 0.08/12, r = 0.05/12 and the estimatesp = 0.48 and oa = 0.005723 from
 20
Table 2. Note that becausea2 and p are both larger than those for copper,the unconditionalvarianceof

incomefrom oil exports is evenlarger than from copperexports. To partiallyoffset the resultinghigher

desireto save for precautionarymotivesand to obtainconvergenceof the algorithm,we use an annualized

time preference of 8%.

The exerciseuses the unconditionalaverage of the ten policy functions-insteadof using Monte

Carlo simulations,so the results should be interpretedin an expectedvalue sense. Figure 8 shows the

optimal, unconditionalexpenditureratio, g = p(f), and the relation (7). Starting from a fund ratio of

four, the optimalpolicy is to run the fund down by consuming(in an expectedsense) along the function

g=p(f), using equation (4), until the fund ratio reaches its long-run level of about 0.10. The exercise

leads to the followingresults. In the first year, the fund ratio is reduced from 4 to 2.13 months of

exports. In the next two years the fund ratio is fulther reduced.o about0.33 at the end of the third year

and in year four to about 0.13. Therefore, the optimalpolicy is to run the windfallgain down slc vly:

a windfallgain worth four months of exports should be run down over a period of about 48 months.

5. EXTENSIONS

A straightforwardextensionof the model is to considerthe interactionwhich may exist between

commodityprice movementsand interestrate movements. If interestrates are stochastic,then the optimal

saving rule will dependon the varianceof interest rate movementsand the covariancebetween interest

rates and prices. The effect of a stochasticr is best demonstratedby lookingat the sufficient condition

for the existenceof an equilibrium. Equation(8) showsthe condition(which replacesequation(6)) when

the growth process for income is independent,identicallydistributed:

a'I (r -8 r ) + < (8)

where r standsnow for the expectedinterest rate, a2, for the varianceof r, a2' for the varianceof log(x),

and o2,.,,for the covariancebetween log(x)and r. The existenceconditioncan, in a heuristic way, show
 21
the effect of introducingstochasticinterestrates. If a',.. is positive, then introducinga stochasticinterest

rate is similar to increasing the slope of the income path, since an income shock is to some extent

exacerbated by an interest earnings shock (and vice-versa for a negative covariance). Interest rate

volatilityper se (the term a', on the left hand side) acts as an increase in the interest rate whui. is likely

to generate more saving. However, empirically, this modificationof the model is not likely to be

significant. For the period 1975-90,for instance,using the US T-bill rate and the logarithmicdifference

in copperprices, o2,is about 1.08xl 5 and o',.. is about 1.7xlO4 .

We have shown in the previous section that a well designed and well managed commodity

stabilizationfundwill lead to a reasonableamountof smoothingof short-term incomefluctuations. How

doesthis smoothingcompareto the results of using other risk managementmeasures?Short-termmarket-

based, commodityrisk managementinstruments--futuresand options-are one of the possibilitieshere.

Margin requirementsassure that credit risk is not an issue, and, in principle, make these instruments

availableto (entitiesin) all countriesat low costs. Through simulations,Claessensand Varangis (1991)

found that oil futures could rerr:;e up to 85% of price risk for periods one to two months ahead. For

longerperiods, these instrumentsstill reduceduncertaintyby up to 81%. Similar results were foundfor

coffee (approximately70% risk reduction)and canbe found for hedgingmany other commoditieswith

market-basedinstruments. As a short-termhedgingtool, market-basedinstrumentscould thus perform

equally well or better than a CSF in removingshort-periodvolatility at lower costs. Especiallyin the

presenceof spikes and fatter tails in the commodityprice distributionit may be even more efficient to

hedge using market-basedinstruments (especiallydesigned to remove the effects of spikes, such as

options)insteadof holdinglow-yieldingassets in reserves. It is surprisingthat to date so few developing

countrieshave used these instruments.

Becausethesehedginginstrumentshaveshort maturities,they willprovide--evenwith rollingover

 22
of the hedges--limitedhedgingvalue over longer periods."
9 Longer-datedinstrumentswould thus be
desirable. The market for longer-termcommodityhedginginstrumentsis thin, however, and, so far, less

available (or only on unattractive terms) to many developing countries due to institutional,

creditworthiness,and other constraints.' It would appear, therefore, that the best strategy to follow

for a commodity-dependentcountry would be to remove as much short-periodcommodityprice risk as

possiblethroughshort-datedinstrumentsas well as--tothe extent possible--uselonger-termhedgingtools.

The CSF mechanismcan be used for smoothingany remaininglonger-periodrisk.

6. CONCLUSIONS

This paper concludes that given actual commodityprice processes observed, and assuming

reasonable values for other parameters, a risk-averse, credit-constrainedcountry would only wish to

maintaina relativelysmall CSF, amountingpossibly to only a fraction of one month's export income.

UsingDeaton's (1991) precautionarysaving model with liquidityconstraints, we found that the Chilean

Copper StabilizationFund should have held only about 20% of monthly incomefrom copperexports at

the end of 1990-far less than the actual level of six to seven times monthlyincome held at that time.

Even if Chile is assumednot to be a credit-constrained,the fund should only have been abouttwo times

the monthlyexports at the end of 1990. The opportunitycost of maintainingsuch large liquid foreign

reserves to smooth adverseincome shocksis too high, even grantinga high degree of risk aversion.

How can we explain CSFs as large as the Chilean case? One, seeminglyrational explanation,

could be that the fund provides positive externalities,such as an increase in overall confidence,which

justifies keepinga large amountof low-yieldingassets even whenone is credit-constrained. Part of this

"The problemwith rolling-overcoveragebasedon short-termmaturinginstrumentsis that this usually

implies a large exposure in these instruments.This can imply large margin calls and/or large option
premiums, makingthe instrumentsless attractivefor foreign-exchangeconstrainedcountries.

'Instruments such as commodity swaps involve credit risks, compete with other credits to the
borrower, and may thus not be availableto the countryon attractiveterms.
 23
improvedconfidencecouldcome from a stabilizationof the real exchange rate, which may benefitother

sectorsof the economywhen there are costs of adjustment. Baschand Engel (1991) explore this notion

by introducingfrictions in adjustinggovernmentprograms. Other possible benefits of holding a CSF

includereduction in (a) the public pressuresto quicklyspend windfallincomegains, (b) spendingwhich

either takeson a life of its own and cannotbe reversed,and (c) spendingwhich is inefficient. These are,

however, more politically-orientedexplanationsand have little appeal from an economicpoint of view.

The benchmark model was also used to compute the optimal depletion rule for a typical oil

exporterthat receiveda windfallgain of aboutfour times its monthlyexport revenuesduring five months

of high oil prices as occurredduring the Gulf crisis. The paper showthat the typicaloil exporter should

take about four years to depletethis large a windfallgain.

Our paper complementsDeaton (1991) by providing an in-depth study of the AR(1) growth

process for income. We show that, even for risk aversion Parametersless than 1, the much higher

volatilityencounteredin the incomeprocessesof developingcountriescomparedto developedcountries

(becauseof dependingon volatile commodityprices) implieshigh levels of precautionarysavings.

Finally, future work which deals with the use of self-insuranceschemes such as commodity

stabilizationfunds shouldtake into accountthe existenceof commodity-linkedcontingentinstrumentsand

present an integratedanalysis, giventhe possible substitutabilityand complementaritybetween the two.

Especiallythe benefitsof market-basedinstrumentsdesignedto removethe effectsof spikes in commodity

prices, such as options, needs to be explored. Further analysisof the volume and the price components

of incomerisk is anotherfruitful researchavenue.

 24

Rdfirences

Arrau, Patricio (1990), "IntertemporalSubstitutionin a MonetaryFramework:Evidencefrom Chile and

Mexico", PRE WorkingPaper # 549,The World Bank, December.

Arrau, Patricio and Sweder van Wijnbergen(1991), "IntertemporalSubstitution, Risk Aversion and
Private Savingsin Mexico", PRE WorkingPaper # 682, The World Bank, May.

Basch, Migueland EduardoEngel(1991), "Shockstransitoriosy mecanismosde estabilizacion",mimeo,

CIEPLAN,October.

Choe, Boum-Jong(1991), "The PrecautionaryDemand for CommodityStocks", mimeo, IECIT, The

World Bank

Claessens,Stijn and Panos Varangis (1991), "HedgingCrude Oil Imports In DevelopingCountries",

PRE WorkingPaper # 755,The World Bank, August.

Deaton, Angus (1990), "Savingin DevelopingCountries:Theoryand Review", Supplement,The World

Bank EconomicReview(Proceedingsof the World Bank Annual Conference-onDelo2me
Enomics 1829)vol. 4, 61-96.

Deaton, Angus (1991), "Saving and Liquidityconstraints",Econometrica,vol. 59, no. 5, September,

1221-1248.

Deaton,Angusand GuyLaroque (1991), "On the Behaviourof CommodityPrices", Reviewof Economic

Studies, forthcoming.[NBERworkingpaper # 34391

Epstein, L. G. and S. E. Zin (1989), "Substitution,Risk Aversion, and th,e Temporal Behavior of
Consumptionand AssetReturns:A TheoreticalFramework",Econometrica,vol. 57, no. 4, July,
937-969.

Epstein, L. G. and S. E. Zin (1991), "Substitution,Risk Aversion, and the Temporal Behavior of
Consumptionand Asset Returns:An Empirical Analysis", ournalof Political Economy,vol.99,
no. 2, April, 263-286.

Farmer, R. (1990), "R.I.N.C.E. Preferences", Ouarterly Journal of Economics, vol CV, Issue 1,
February, 43-60.

Giovannini,A. and P. Weil (1989), "Risk Aversionand IntertemporalSubstitutionin the Capital Asset
Pricing Model", NBER WorkingPaper # 2824, January.

Hausmann, Ricardo and Andrew Powell (1990), "El fondo de estabilizacion macroeconomica:
lineamientosgenerales,", mimeo, IESA, Caracas, November

Mackinnon,James (1990), "Critical Values for CointegrationTests", working paper, University of

California,San Diego, January.
 25
Ostry, Jonathan and Carmen Reinhart (1991) "The Harberger-Laursen-MetzlerHypothesis: Evidence
from DevelopingCountries",InternationalMonetaryFund, Washington,mimeo, August

Tauchen, G. (1986), "Finite State Markov Chain Approximation to Univariate and Vector
Autoregressions",EconomicLetters, 22, 237-55

Trivedi, P. (1991), "Time Series Behaviorof Some CommodityPrices", mimeo, IECDIT, The World
Bank

Weil, P. (1990), "NonexpectedUtility in Macroeconomics",OuarterlyJournal of Economics, vol CV,

Issue 1, February, 2942.

Williams, Jeffrey and Brian Wright (1991), Storageand CommodityMarkets, Cambridge University
Press.
 26
AppendixA: Computationof policy functions

The first step is to compute the 10-pointdiscrete Markov approximationto the continuous,
autoregressiveprocess for log(x,+,). If log(x,+,)is distributedconditionalon the informationin period
1 ) is unconditionallydistributedas N(ju,u2/(l - p2)).
t as N(j(1 - p) + p log(xJ, a;), then 1,g(x,+

To approximatethe unconditionaldistribution, we follow Deaton (1991) and Tauchen (1986).

First divide the real line R in 10 segmentsof equal area under the unconditionaldensity so that every
segmentof the distributionhas a one tenthprobabilityof occurrence. Let at,... ,ak be the nine valuesthat
divide the 10 segments (from left to right) and let a3 = -mo and a10 = x. Every segment of the
distributionof log(x) is now representedby its within-intervalmean z1 ,...,z1O. (For the case of copper
income distributionzt = -0.120 and z,0 = 0.127).

Because the process is positivelyautocorrelated,given a value z; in the previous period, it will

be more likely that this period realizationwill be close to zi. This can be representedby the Markov
transition matrix with typical element7rj, which stands for the probabilityof state j, given that state i
occurred last period. Unlike Deaton (1991), and more along the lines of Tauchen (1986), we
approximatethis transition matrix by computing
fjfj = Pqa,>log(x 1+1 ) 2 a, 1 I log(x,) =z,]

We use thus the conditionaldistributionof log(x,+,),and the transitionprobabilitieswhich can easily be

computednumerically.

We verify the accuracy of the discrete approximationby simulation.We simulatethe income

process using the parameterizationfor Chile: (u, 02,p) = (0.0032, 0.004436, 0.32). The drift i&= Fp
+ Aqis obtained from the first regressionsin Table 2 and 3 respectively(the estimateof p, in Table 2
is taken as zero) and the other two parameterscome from the first regression in Table 2. We simulate
1000replicationsof 370 observationseach and estimatethe autoregressiveprocess for incomefrom these
1000replications. Table Al comparesthe values from the actualregressionswith the average estimates
from the Monte Carlo replications. We can see that our approximationto the true process is very
accurate as true and simulatedestimatesare very close.
 27

Table Al: Simulationof incomeprocess:

AX = y(1 -p) + p X., + e,

e,is N(0, a2)

True Mean Standarddev.

Parameters Simulation Simulation

'U(l-p) 0.00218 0.00214 0.00332

p 0.320 0.306 0.0599
a2 0.004436 0.004277 0.00027
R2 0.0956 0.0301
D.W. ^ 1.99 0.0326

Note True parameterscome from the first equation in Tables 2 and 3 (a = + &). Columns 2 and
3 come from OLS estimationsof 370 Monte Carlo simulationseach, replicated 1000 times. For each
estimatedparameterand statisticwe presentthe mean and the standarddeviationof the 1000replications.

For the simulationsunderlyingFigures5 and 6 and Table 6, we first classifythe actualrealization

of the logarithmicdifferenceof real copper pricesbetweenSeptember1987and December 1990 (adding
the drift u = 0.0032) so one of the ten intervalsof the unconditionaldistribution.

The ten policy functionsare computedby iterating equation(5) in the text, which as discussed
by Deaton (1991), are contractionmappingsfor some sets of parameters.
 28
Appendix& CommodityProductionand Exports for Chile, Mexicoand Venezuela(volumes)
Cn le: Copper ProductIon Chi le COoper Exports

140
.40~~~~~~~~~~~~~~~~~~~~~~~~~~2

110ico: 011 Production Mexico. Oi l ExC>os22

110 06

*0 *0

600_
,i - .. .

(index,1977=100) (millionpounds)
11 0 /0 0''0

ltexico: Oil Production MexICO. Ofl Exports

16

10

lie~~~~~i
.06

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~0

. 00 M0 4"'i n4M11

(KToil equivalent)
(KCT equivalent) (KT
(KT oil equivalent)
equivalent)

Venezuela:
OilProduction Venezuela: Oil Export~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Source anldNotes: See Table 3.

 29
Table 1: Integrationtests for real commodityprices. Are the series stationary?

DF ADF(4) ADF(12)

C CT C CT C CT #obs

Real Copper Price

log level NO NO NO NO NO YES(*) 372
log. diff. YES YES YES YES YES YES 371
Real Oil Price West-TexasInterm.
log level NO NO NO NO NO NO 83
log. diff. YES YES YES YES NO NO 82
Real Oil Price for Mexico
log level NO NO NO NO NO NO 132
log. diff. YES YES YES YES YES YES 131
Note:Dickey-Fuller (DF)and Augmented Dickey-Fuller(ADF)testsat 5%criticalvalue.Appliedto the logof the
levelof pricesand logarithmicdifferenceof prices. Thenominalcopper(1960.01-1990.12) and theWest-Texas
intermediate oil price(1984.01-1990.12)
are deflatedby the US CPI. An indexof importedprivateconsumption
goodsis usedto deflatetheoil pricefor Mexico(1980.01-1990.12). 'C' meansthat only theconstantis included
in thetest; 'CT' meansconstantas wellas trendareincludedin thetest. Thecriticalvaluesfor thetestscomefrom
Mackinnon (1990).(*)Rejectat 10%criticalvalue.
Sources: Bancode Mexico, lIndicadores Economicos';IMF, International FinancialStatisticsdatabase.
Table 2: Estimationof commodityprice processes
A log(p,) = 1p(1 - p) + p A log(p,.1) + e,
e, is N(0, o2)
R2
A, (1 - ,3) p (02) DW

Real copper price -0.00057 0.320 0.10 1.88

(-0.17) (6.47) (0.004436)
Nominalcopper price 0.00222 0.317 0.10 1.88
(0.64) (6.41) (0.004417)
Real oil price WTI -0.00367 0.356 0.12 1.82
(-0.35) (3.34) (0.008841)
Nominaloil price WTI -0.00160 0.364 0.13 1.82
(-. 15) (3.42) (0.008980)
Real oil price Mexico -0.00323 0.483 0.23 1.78
(-0.48) (6.11) (0.005723)
Nominaloil price Mexico -0.00188 0.493 0.23 1.79
(4.28) (6.26) (0.005727)

Not: t-statisticsin parentheses.

Source:IMF-Intenational FinancialStatisticsdataset;Bancode Mexico,*Indicadores
Economicos".WTIstands
for theWestTexasIntermediate price. Deflatoris the US CPIindexexceptfor Mexicowherethedeflatoris an
indexof the unitvalueof importedprivateconsumption goods.
 30

TIable3: Estimationof deterministictrend for volume of commodity

Log(Q) = intercept + pq t + et

Volume
Index Intercept Jlq R2 DW

Chile
Copper produced 3.844 0.00318 0.97 1.31
1960-90 (191) (34.9)
Copper exported 2.111 0.00371 0.91 1.01
1960-90 (43.5) (16.9)
Mexico
Fuel produced 9.268 0.00835 0.93 0.15
1961-89 (99.1) (18.4)
Fuel exported 9.196 0.00897 0.91 0.13
1965-88 (86.9) (14.5)
Venezuela
Fuel produced 12.251 -0.00280 0.78 0.33
1961-89 (252) (-9.9)
Fuel exported 12.259 -0.00287 0.71 0.31
1965-88 (216) (-7.3)

Note: t-statisticsin parentheses.

Sources: Mexico and Venezuela: International Econoraics Department, The World Bank, Energy data set.
Production and export index is in terms of kilo tons (KT) oil equivalent. Chile: Banco Central de Chile
'Indicadoreseconomicosy sociales1960-88', and 'Monthly Bouletin'. Productionindexfor copperis 100for 1977.
Volumeof copper exports is derived by dividingthe nominaldollar amountof copperexports (in millions)by the
LondonMetal Exchangecopper price (cents/pound).
 31

Table 4 Direct Estimationof IncomeProcess

A log(X.)= t( - p) + p A logQ4.,) + e,
f, is N(O,a)

R2
1, (I- P) p (o2) DW
Real copper exports -0.00016 -0.598 0.36 2.06
(0.01) (8.42) (0.0655)
Nominalcopper exports 0.01043 -0.597 0.36 2.05
(0.46) (8.40) (0.0655)
Real oil exports Mexico 0.00195 -0.108 0.01 1.98
(0.12) (1.23) (0.0320)
Nominaloil exports Mexico 0.00509 -0.102 0.01 1.98
(0.32) (1.15) (0.0321)

Nte: t-statisticsin parentheses.

Source: Dollar copperexports values come from IMF, IntemationalFinancialStatistics. Oil exports values come
from Bancvde Mexico, lIndicadoresEconomicos".Deflator is the US CPI index for copper and an index of the
unit value of importedprivate consumptiongoods for Mexico.

Table 5: Simulationof Full Model for Chile, Case 1

Mean Standarddev.
Simulation Simulation

Fund ratio (f) 0.657 (0.383a) 1.344

A log (G) 0.0015 0.0028
s. d. A log (G) 0.0402 0.00173
M.P.C in level 1.000 0.00672
- Durbin-Watson 1.46 0.189
M.P.C in log diff. 0.559 0.0146
- Durbin-Watson 2.01 0.0540

Note: Mean and standard deviationof 1000 replicationsof the simulatedmodel with 200 observationsin each
replication. M.P.C. is the marginalpropensityto consume(spend)out of income from exports. It is obtainedby
a linear OLS regression between the simulatedexpendituresequenceand income sequence. In the first case the
regressionis in levels and in the second in growthinnovations(log differences). In both cases we reportbelow the
M.P.C the Durbin-Watsonstatisticfrom the OLS regression.
(a) median(midpoint)
 32

Table 6: Comparisonof actualand optimum quarterlydeposits in Copper StabilizationFund for Chile

(millionsof US dollars)

Deposits Deposits
Net of Net of
Expenditure Expenditure
Withdrawals + Capitald Model Model
(grape crisis)0 Withdrawals Case 1 Case 2

1987.IV 26.36 26.36 -2.37 21.76

1988.1 125.43 125.34 21.45 51.24

1988.11 146.95 146.95 -1.71 33.34
1988.111 140.77 140.77 2.63 37.01
1988.IV 82.85 -87.15 -28.65 13.90

1989.1 467.51 467.51 18.11 58.24

1989.11 112.17 112.17 34.31 73.06
1989.III 219.93 219.93 -24.91 26.72
1989.IV 40.40 -53.16 33.92 82.41

1990.1 263.34 263.34 -10.46 25.50

1990.1I 79.65 79.65 3.32 49.81
1990.111 229.32 229.32 -18.77 27.08
1990.1V 239.76 239.76 36.07 76.56

Fund 1 2174.4 1910.9 62.9 576.60

Fund 2a 2419.7 2119.3 69.1 643.40

Fund 1/Expb.
(months) 6.7 6.0 0.20 1.8
Fund 2/Expb.
(months) 7.6 6.6 0.22 2.0

Fund2 usesa
Nots : Fund1 is thestockof assetsat theendof theperiodwithoutinterestcapitalizations.
quaterlynominalinterestof 2%.
b: Theaveragemonthly of 1990wereUS$320million.
copperexportsin thelastquarter
c: Fundswereusedto smoothouttheeffectsof thegrapecrisisbyallowingexportersto purchasedebt
inatrumentsin thesecondarymarket
andthenreseUthemto theCentral Bankata priceclosetopar.
d:Thegovermentrepurchased debtof theCentral
external of $252.3million).Other
Bank(in theamount
"accounting'transactions
betweentheTreasury andtheCentralBankarenotconsidered.Source:Fundfigures:
Ministryof Finance;Copperexports:CentralBank,'MonthlyBouletin-
 33

GAUSS

Figure 1: Policy Functions, g = p(f), for a given state

r = 0.04, d = 0.08, a 2

00
D 0a4 bes statett

C:

x 0

'3 0. 0. 0. 03 0. 0. 0. 07 0. 0. 10

o- ~ ~ ~ ~ ~
.Sensitivity
ofPliyFncina(ostsaeeny ~ ~ ~ ~ rs
>(D 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
fund ratio (f)

GAUSS

Figure 2:
Sensitivity of Policy Functions (worst state only)

(0

x O ~~~~~r =0.05/126 =r-0.000, 2

0)
-~cc

c r =0.05/12. O.O/1. 0.5 6

0)0

oo- .0.05 .1. r . .r. . a - 0 .

°0.0 0.1 0.2 3 4 0 5 0 6 0.7 0.8 0.9 1.0

fund ratio (f)
 34

GAUSS

Figure 3: Policy functions base simuiation case

O . . .

r ~~~~~~~~~~~best
state
.r05/1 2. = r 0.0005. a = 0.5,

04

03
CL

worst state

° 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Fund ratio (f)
GAUSS

Figure 4: Typical simulation for Chile

Ni

income growth foctor + 1 (1 + x)

O expenditure ratio (g)

O fund ratio (f)

0 20 40 eo 80 100 120 140 160 180 200

time
 35

GAUSS

Figure 5:
Actual Simulation for Chile: 1987-1990, Case 1

04

1 + income growth factor ( 1 + x)

expenditure ratio (g)

00 4 8 12 16 20 24 28 32 36 40
months from September 1987 through December 1990

GAUSS

Figure 6:
Actual Simulation for Chile: 1987-1990, Case 2

(N

. 1 + income growth factor (1 + x) g

expenditure ratio (9g)

O / und rotio (f)

0 0 4 8 12 16 20 24 28 32 36 40
months from September 1987 through December 1990
 36

GAUSS

Figure 7: Long-Run Fund Level for Chile, Case 1

O4 g =POt

o0
- o.- -r/ Fr

00
09
co/-
0)

O /

o^ ,I. = 0.003x, r = 0.05/12 = 0.0041 - d

a 0.18 0.22 0.26 0.30 0.34 0.38 0.42

fund ratio (f)

GAUSS

Figure 8:
Depletion Rule of the oil Windfall Gain for Mexico
CN

g p(f

x-

04

4.-~~~~~~2163. .

O)
or g a 0.003 - 0.05/12, 6 =0.08/12
C~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~I
+ a _ o .5.
s - 0 005723.,o p 0.48

°l 0.0 0.4 0. 8 2 .6 20 2.4 28 ;5.2 3. .

fund ratio (f)
 PolleyResearch
WorkingPaoerSeries

Contact
v~ Autbor D fora"

WPS814 Finance,Growth,and PublicPolicy MarkGertler December

1991 W. Pitayatonakarn
AndrewRose 37666

WPS815 Governance
andEconomy:
A ReviewDeborahBrautigam December
1991 Z. Kranzer
37494

WPS816 EconomicConsequencesof German GerhardPohl December

1991 CECSE
Reunification:
12MonthsAftertheBig 37188
Bang

WPS817 HowDoesBrady-Type Commercial MohuaMukherjee December

1991 Y. Arellano
DebtRestructuring
Work? 31379

WPS818 DoRulesControlPower?GATT J. MichaelFinger January1992 N.Artis

ArticlesandArrangements
inthe Sumana Dhar 37947
Uruguay Round

WPS819 FinancialIndicators
andGrowthin a RobertG. King January1992 W. Pitayatonakarn
CrossSectionof Countries RossLevine 37666

WPS820 Taxationin Decentralizing

Socialist Christopher
Heady January1992 D. Sebastian
Economies: TheCaseof China PradeepK.Mitra 80423

WPS821 WagesandUnemployment
in Pcland:FabrizioCoricelli January1992 V. Berthelmes
RecentDevelopments
andPolicy AnaRevenga 39175
Issues

WPS822 Patemalism
andtheAlleviation
of NancyJesurun-ClementsJanuary1992 F. Betancourt
Poverty 18-126

WPS823 HowPrivateEnterpriseOrganized StevenM. Jaffee January1992 C. Spooner

Agricultural
Marketsin Kenya 30464

WPS824 Back-of-the-Envelope
Estimates SergioMargulis January1992 J. Arevalo
of Environmental
Damage Costsin 30745
Mexico

WPS825 TheEmptyOpportunity: LocalControlMarlaineE. Lockhead January1992 D.Eugene

of Secondary
SchoolsandStudent OinghuaZhao 33678
Achievement
in the Philippines

WPS826 DoWorkersin the Informa!

Sector ArielFiszbein January1992 N. Perez
BenefitfromCutsin theMinimum 31947
Wage?

WPS827 FreeTradeAgreements withthe RefikErzan January1992 J. Jacobson

UnitedStates:What'sIn It for Latin AlexanderYeats 33710
America?
 PolicyResearch
WorkingPanerSeries

Contact
] i AJ= Dotor p=r

WPS828 Howthe Macroeconomic

Environment
ArvilVanAdams January1992 V. Charles
AffectsHumanResource RobertGoldfarb 33651
Development TerenceKelly

WPS829 Regulationof SecuritiesMarkets: TerryM.Chuppe January1992 F. Harbottle

SomeRecentTrendsandTheir MichaelAtkin 39616
Implications
for EmergingMarkets

WPS830 FixedParityof the Exchange

Rate IbrahimElbadawi January1992 V. Barthelmes
andEconomic Performance
in the NaderMaid 39175
CFAZone:A Comparative Study

WPS831 RealOvervaluation,
Termsof Trade IbrahimElbadawi January1992 V. Barthelmes
Sht.nks,andtheCostto Agriculture 39175
in Sub-Saharan
Africa

WPS832 Sustainability
andthe Economics RichardB. Norgaard January1992 J. ShinYar.g
of AssuringAssetsfor Future 81418
Generations

WP' 33 Stabilization
andGrowthRecovery DanielF. Oks January1°92 L. Frznchini
in Mexico:Lessons
andDilemmas 38835

WPS834 Scer.arios
forGrowthin the 1990s ShahrokhFardoust January1992 J. Queen
Jian-Ping
Zhou 33740

WPS835 Commodity
Stabilization
Funds PatricioArrau January1992 S. King-Watson
StijnCiaessens 31047

WPS836 Sourcesof IncomeInequality

in Rural RichardH.Adams,Jr. January1992 C. Spooner
Pakistan:A Decomposition
Analysis HaroldAlderman 30464

WPS837 Manpower
Planningin a Market ArvilVanAdams January1992 S. Khan
Economy
witnLaborMarketSignals JohnMiddleton 33651
AdrianZiderman