ISSN 1795-5300
 MTT Discussion Papers 5  2006MTT Discussion Papers 5  2006
 Investment and the Dynamic 
Cost of Income Uncertainty: 
the Case of Diminishing 
Expectations in Agriculture
 Tiina Heikkinen & Kyösti Pietola
Investment and the Dynamic Cost of Income
 Uncertainty: the Case of Diminishing
 Expectations in Agriculture ∗
 T. Heikkinen †
 MTT Economic Research
 Luutnantintie 13, 00410 Helsinki FINLAND
 email:tiina.heikkinen@mtt.fi
 and K. Pietola, MTT Economic Research
 October 10, 2006
 Abstract
 This paper studies optimal investment and the dynamic cost of income
 uncertainty, applying a stochastic programming approach. The motiva-
 tion is given by a case study in Finnish agriculture. Investment decision
 is modelled as a Markov decision process, extended to account for risk.
 A numerical framework for studying the dynamic uncertainty cost is pre-
 sented, modifying the classical expected value of perfect information to a
 dynamic setting. The uncertainty cost depends on the volatility of income;
 e.g. with stationary income, the dynamic uncertainty cost corresponds to
 a dynamic option value of postponing investment. The numerical invest-
 ment model also yields the optimal investment behavior of a representative
 farm. The model can be applied e.g. in planning investment subsidies for
 maintaining target investments. In the case study, the investment decision
 is sensitive to risk.
 Keywords: investment analysis, real options, OR in agriculture, stochastic
 programming
 ∗An earlier version of this paper was presented at ERSA 2006, Volos, Greece.
 †Useful discussions with A.-M. Heikkila¨ are gratefully acknowledged.
 1
1 Introduction
 Currently, common Agricultural Policy (CAP) after EU enlargement implies
 many uncertainties regarding future agricultural income in Northern Europe.
 This paper addresses the dynamic cost of income uncertainty in investment
 analysis, focusing on a case study in Finnish agriculture. In the case study, the
 income subsidy level, being determined by a political process, can be stochastic
 at the time the investment decision is made. Applying a stochastic programming
 approach [Birge and Louveaux1997, Prekopa1995], a numerical framework for
 studying dynamic uncertainty cost is presented, modifying the classical expected
 value of perfect information to a dynamic setting. Even though motivated by
 the Finnish case, the framework is applicable to investment analysis in similar
 situations where policy uncertainty both at the national and EU level may make
 the rate of return on investment uncertain 1. Furthermore, the framework is in
 general applicable to other dynamic optimization problems similar to the invest-
 ment problem (i.e. optimal stopping problems) where the dynamic uncertainty
 cost is relevant.
 Optimal investment is studied applying real investment options, see e.g.
 [Dixit and Pindyck1994, Keswani and Shackleton2006]. Investment options typ-
 ically involve three parameters: the initial and accumulated costs, the flexibil-
 ity in timing the investment and the uncertainty regarding the future rewards.
 [Dixit and Pindyck1994] studies the optimal investment decision as a Markov
 decision process (MDP) defined in continuous time (Ito process) and with a
 continuous state space. To simplify numerical analysis, this paper applies a
 discrete time MDP with discretized state space to study optimal investment.
 The optimization model is based on assuming a risk-neutral representa-
 tive decision maker. Risk implies an additional cost of uncertainty 2. To
 study the effect of risk on investment, the model is extended to explicitly ac-
 counting for risk, based on the stochastic programming approach introduced in
 [Levitt and Ben-Israel2001] (previously with applications to inventory control
 and the maintenance problem).
 The main results can be summarized as follows:
 • A framework, consisting of numerical models, for quantifying the dynamic
 uncertainty cost is presented, modifying the expected value of perfect in-
 formation (EVPI) [Birge and Louveaux1997] to a dynamic setting. The
 framework can be applied to study the value of information as function of
 the frequency of income uncertainty. The investment model can also be
 used to study the uncertainty-investment relation.
 • Case study examples suggest that the cost of annual income uncertainty
 could be significant: 15%, or even more, of the expected value of invest-
 1Income variability is a persistent problem in agriculture; for simplicity other sources of
 uncertainty are ignored in this paper.
 2For an empirical approach to measuring the cost of risk (representing the variance-
 covariance structure of firm’s income), see e.g. [Amegbeto and Featherstone1992].
 2
ment. A connection between the option value of postponing investment
 and the dynamic uncertainty cost is observed: the two are equivalent in
 the special case of stationary income.
 • The dynamic investment decision is sensitive to risk.
 The lack of complete information causes inefficiency, cf. [Lagerkvist2005]; a
 high uncertainty cost deteriorates the efficiency of investment subsidies. That
 risk matters to optimal investment is a result reached in this paper within a
 dynamic model. Related work [Alvarez and Stenbacka2004] applies an analyt-
 ical continuous time model, whereas in this paper similar results are obtained
 via simulation in a discrete time setting. On the other hand, previous work
 discussed in [Lagerkvist2005] (with reference to [Knapp and Olson1996]) sug-
 gests risk aversion to be of less importance in a dynamic model than in a static
 setting.
 Related work in [Vercammen2003, Vercammen2006] applies a stochastic dy-
 namic programming model to study optimal farmland investment assuming a
 decoupled direct payment. For simplicity, the farm income is assumed inde-
 pendently drawn from a stationary distribution. This paper makes the more
 realistic assumption that income is time-correlated, with a gradual reversion
 to a long-term mean. Mean reverting processes are frequently used in real
 option models; this paper considers the special case of a mean reverting in-
 come process with a non-increasing expected value. Another simplification in
 [Vercammen2003, Vercammen2006] is to assume that the farmer faces a binary
 investment each time period: the farmer is constrained to invest at most one
 unit on capital each period. This paper focuses on the alternative case of a
 lumpy (irreversible) investment decision made at most once, considering the
 case with period-specific financial constraints as a special case.
 In general both the growth in the value of investment and uncertainty affect
 the optimal investment decision [Dixit and Pindyck1994]. Recent econometric
 evidence supports a nonlinear uncertainty-investment relation: for low levels
 of uncertainty an increase in uncertainty has a positive effect on investment,
 while for high levels of uncertainty an increase in uncertainty lowers investment
 [Bo and Lensink2005]. The investment model presented in this paper can be
 used to study the effect of uncertainty (as modelled by the variability in in-
 come) on optimal investment behavior, assuming a risk-neutral or a risk averse
 decision-maker. It should be noted that strategic interactions are not consid-
 ered in this paper. Recently, [Mason and Weeds2005] suggest that the effect of
 uncertainty on investment is ambiguous in a duopoly 3.
 The organization of the paper is as follows. Section 2 introduces the in-
 vestment problem of an optimizing representative agent. Section 3 introduces
 the case study and the stochastic programming model for the expected value
 of perfect information (”uncertainty cost”). Section 4 extends the uncertainty
 cost to a dynamic setting and presents numerical examples. Section 5 considers
 the case of a risk-averse agent. Examples suggest that taking risk into account
 3For an overview on strategic investment under uncertainty, see e.g. [Huisman et al.2003].
 3
in general affects the optimal investment decision. It remains a topic for future
 work to obtain the subjective probability distributions in the Markov model 4,
 e.g. by conducting a survey similar to that in [Lagerkvist2005] using a visual
 impact method [Hardaker et al.1997]. Then the model can be applied e.g. in
 planning investment subsidies.
 2 Optimizing Investment
 The flexibility in timing the investment affects the value of investment [Dixit and Pindyck1994,
 Keswani and Shackleton2006]. In this paper the decision-maker is assumed to
 have full flexibility in timing the investment. At each time t = 1, ..., T the firm
 decides the investment at t, It. Denoting by I the total available budget for the
 investment at t, the decision It at t satisfies
 It ∈ {0, I}, t = 1, ..., T. (1)
 Let rt denote the internal return on investment (%) per time period at time t.
 Due to variability in the rate of return, the future value of investment is ran-
 dom. Letting Ia denote the aggregate budget, the aggregate budget constraint
 requires:
 T∑
 t=1
 It ≤ Ia. (2)
 Letting ρ denote the internal rate of return, the discount factor b ∈ (0, 1] is
 defined as b = 1/(1+ρ). The dynamic optimization problem of a representative
 firm can be written as:
 max E[
 ∞∑
 t=1
 bt(rt +
 ∞∑
 k=t+1
 bkrk)It − It] (3)
 where E denotes the expectation operator.
 Two versions of problem (3) subject to (1)-(2) are studied in what follows;
 in the first model, it is assumed that rt is observable when the investment
 decision is made at time t; in the second model only rt−1 is observable at time
 t. Since the future values of the investment are unknown in both models, there
 is an opportunity cost to making the investment decision at the beginning of
 the time horizon [Dixit and Pindyck1994]; the firm has the option to postpone
 the investment decision. A time-correlated income process is assumed, to study
 optimal investment under decreasing income expectations, allowing the firm to
 make the investment decision at any time t = 1, ..., T . The optimal investment
 rule will be threshold-based, with time-dependent thresholds.
 4Examples suggest that the uncertainty cost can be sensitive to underlying probabilities.
 4
3 Investment with Time-Correlated Income
 In this paper the income process is assumed to be non-increasing, reflecting
 decreasing expectations regarding income subsidies. The motivation is given by
 a case study from Finnish agriculture, summarized in what follows applying a
 discrete time Markov model. A stochastic programming model for measuring
 the dynamic uncertainty cost is introduced, based on two optimization models.
 In the first model (Model 1), the value of investment is observable at the time
 the investment decision is made; in the second model (Model 2), the value of
 investment is unobservable.
 A Case Study
 Milk production is the most important production line in Finnish agriculture
 [Lehtonen2004]. Table 1 summarizes the expected profitability of investment
 in milk production in Finland, based on [Uusitalo et al.2004]. For example,
 assuming the investment subsidy grows by 20 % (or by 50 %, depending on
 the type of the production unit) and assuming the producer price decreases
 by 15 % from 2003 level by 2007, the profitability of a livestock-place is 11 %
 2007, assuming herd size 130. The expected producer price changes reflect ex-
 pected policy changes including the removal of production-based support. For
 details regarding Table 1, see [Uusitalo et al.2004]. Recent survey studies sup-
 port pessimistic expectations regarding future profitability. In a deterministic
 continuous time model postponing investment is not optimal under decreasing
 income expectations [Dixit and Pindyck1994].
 Markov Model
 In Table 1 the states of future rate of return depend on the herd size and future
 producer price. Assume the possible Markov states are defined in terms of the
 future rate of return, corresponding to different scenarios regarding producer
 price change (for a given herd size). Denote the matrix of transition probabilities
 by A. Let rit denote the rate of return at time t in state i. The expected return
 E(rt) at time t is defined as:
 E(rt) =
 ∑
 i
 Pitrit (4)
 where Pit is the probability that the rate of return is determined by state i at
 time t. The probabilities Pt = {Pit} associated with the different states rit at
 time t are determined from:
 P′t = P′0At, (5)
 where P0 denotes the vector of initial probabilities of the different return rates
 and corresponding subsidy levels.
 5
Table 1: Return on investment (ROI %) in milk production (2007 -10 % means
 2007 ROI (including subsidy) when producer price decreases by 10 % from 2003
 and investment subsidy increases by 20 % or 50 % depending on production
 unit type
 ROI % ROI %
 herd size 60 herd size 130
 2003 24 30
 2007 -10 % 10 16
 2007 -12 % 7 14
 2007 -15 % 4 11
 2007 -17 % 2 8
 2007 -20 % 0 5
 In general, the transition probabilities at time t can be defined as function
 of the investment decision at time t. Formally, letting rt denote the state at
 time t and It denote the investment decision at time t, the state transition
 probability is given as a function P (rt|rt−1, It). A Markov Decision Process
 (MDP) is a Markov Model with the above modification, i.e. the transition
 probability matrix depends on the action taken in each stage 5. For example,
 investment may increase productivity: this can be modelled by an MDP with a
 more advantageous transition matrix whenever investment takes place.
 Expected Value of Information
 In stochastic programming literature [Birge and Louveaux1997], the expected
 value of perfect information measures the maximum amount a decision maker
 would be willing to pay for complete information:
 Definition 1 Let f(x) denote the objective function to be maximized with re-
 spect to decision variable x. Let z denote a random variable. The expected value
 of perfect information (EVPI) can be measured as the difference [Birge and Louveaux1997]
 EV PI = E[max f(x, z)]−max E[f(x, z)]. (6)
 5Furthermore, in general the transition probabilities depend on the timing of the investment
 (e.g. due to fixed term investment subsidy programs).
 6
Table 2: Transition Probabilities between States (ROI %)
 0.3 0.16 0.11 0.05 0
 0.3 0.01 0.3 0.4 0.28 0.01
 0.16 0.01 0.8 0.1 0.09 0
 0.11 0.01 0.05 0.7 0.15 0.09
 0.05 0.01 0.01 0.08 0.8 0.1
 0 0 0 0.05 0.15 0.8
 The first term in equation (6) corresponds to a ”wait-and-see” solution and
 the second term to an expected value maximizing solution. EVPI can be used to
 measure the cost of imperfect information due to uncertain income and subsidies.
 Before extending EVPI to a dynamic setting, two optimization models of in-
 vestment are introduced: a ”wait-and-see” model and an expected value model,
 respectively.
 Model 1: Investment in a Wait-and-See Model
 In Model 1, like in [Dixit and Pindyck1994], the value of investment is observ-
 able at any time but the future values are random. The future value of in-
 vestment is assumed to follow the same mean-reverting Markov process as the
 producer price does; with high probability, the value of investment remains un-
 changed. Specifically, at time t, the rate of return rt is observed, and future
 return rate rt+1 is determined by a transition matrix modelling non-increasing
 income expectations (Table 2). The hypothetical Markov model as summarized
 in Table 2 is a simplified model of a mean-reverting income process, with a long
 run mean return rate 0.06. With high probability, the value of investment re-
 mains unchanged (except for the highest rate of return r = 0.3 that is assumed
 to decrease with probability 99 % in Table 2).
 Formally, let yt denote the value of investment in terms of income obtained
 when investing It at t, assuming infinite time horizon:
 yt(It) = rtIt + E[
 ∞∑
 k=t+1
 bkrk]It. (7)
 At each time t = 1, ..., T the firm makes a decision on the level of investment
 It subject to constraint (1)-(2), with an aggregate budget Ia. The dynamic
 optimization problem subject to constraints (1)-(2) can be stated as
 max E[∑
 t
 bt(yt(It)− It)]. (8)
 7
Problem (8) subject to (1)-(2) can be solved recursively applying Bellman equa-
 tion:
 v(rt) = max
 It
 {(yt − It) + bEv(rt+1)}, (9)
 where v(rt) denotes the value function given state rt. For simplicity, the invest-
 ment model will be formalized as an MDP as follows: Define an additional state
 ra = 0 corresponding to a state where the budget has been used up. After in-
 vestment has been made a new transition probability matrix applies: one where
 each state leads to state ra with probability one.
 Model 2: Expected Value Maximization
 In Model 1, like in [Dixit and Pindyck1994], the value of investment at any time
 t is observable. In the case study summarized above, due to a high degree of
 policy uncertainty, the return rate rt can be uncertain at the beginning of period
 t. Assuming rt is observed at the end of period t, all terms affecting the value
 of investment are random. A risk-neutral decision maker in this case solves the
 Bellman equation:
 v(E(rt)) = max
 It
 {E[yt(rt, It)− It] + bv[E(rt+1)]}. (10)
 According to formulation (10) the decision maker has the flexibility to make
 the investment decision at any time; the expected return can be determined
 based on the return observed previous time period. Assuming a stationary in-
 come process, however, there is no motivation for postponing investment; in this
 case NPV maximization is optimal. Using the terminology in [Keswani and Shackleton2006],
 the special case where the investment decision is made at the beginning of the
 time horizon corresponds to optimizing forward start net present value (NPV).
 EVPI and Option Value
 Consider the case of optimizing forward start NPV, restricting the decision
 maker to choose the level of investment at t=1 in model 2. Define the dynamic
 objective function f({It}, {rt}) as:
 f({It}, {rt}) =
 ∑
 t
 bt(yt(rt, It)− It), (11)
 where yt is formalized in (7), cf. problem (8). Definition 1 for the expected value
 of perfect information (EVPI) can be directly applied to the dynamic objective
 (11), replacing the decision variable x in equation (6) by the sequence {It} and
 replacing z by the random sequence {rt}. Then, the first term on the right hand
 side in (6) corresponds to the expected value of the wait-and-see model (Model
 1), based on assuming the value of investment is observable when the investment
 decision is made; The second term on the right hand side in (6) corresponds to
 maximizing the expected forward start NPV, i.e. to determining the optimal
 8
timing of investment at the beginning of the time horizon. Thus, the classical
 option value of postponing the investment decision [Dixit and Pindyck1994] can
 be seen as equivalent to EVPI in Definition 1.
 4 Optimal Investment and the Dynamic Cost of
 Uncertainty
 Directly applying Definition (1) above to the dynamic investment model restricts
 the decision maker to choose its investment policy at t=1 when solving model
 2. Assuming the investment decision can be made at any time even when the
 value of investment is not fully observable, EVPI can be modified to a dynamic
 uncertainty cost as follows. Let {I∗t } denote the solution to (9) (Model 1) and
 let {I∗∗t } denote the solution to (10) (Model 2). Applying expression (6) to the
 dynamic optimization problem (3), implies a dynamic uncertainty cost, EVPI(t),
 defined for period t as
 EV PI(t) = E[yt(I∗t )− I∗t ]− E[yt(E(rt|rt−1), I∗∗t )− I∗∗t ] (12)
 where the first term corresponds to the expected value obtained at t when
 solving the wait-and-see model (Model 1) and the second term formalizes the
 corresponding expected value when the investment decision at time t is based
 on E(rt), given the observed return rt−1 (Model 2).
 Let Prt denote the probability of investment at time t when the state rt is
 observed, and let Pr′t denote the corresponding probability with expected value
 maximization (Model 2). Assuming Prt > 0, define the unit value of investment
 at time t, v1(t), in Model 1 as
 v1(t) = E[yt(I
 ∗
 t )− I∗t ]
 PrtIa
 and assuming Pr′t > 0 define the unit value in Model 2 as:
 v2(t) = E[yt(E(rt|rt−1), I
 ∗∗
 t )− I∗∗t ]
 Pr′tIa
 .
 If Prt = 0, let v1(t) = 0 and if Pr′t = 0 let v2(t) = 0. An uncertain state in terms
 of ROI % lowers the value of investment in two ways: by lowering the unit value
 of investment given the expected amount of investment and by reducing the
 expected investment for a given unit value. Accordingly, the dynamic EVPI(t)
 in (12) can be decomposed into two components:
 EV PI(t) = (v1(t)− v2(t))PrtIa + v2(t)(Prt − Pr′t)Ia. (13)
 Aggregating over time, the uncertainty cost can be defined as follows:
 9
Definition 2 Let x % denote the percentage of time during which Model 2 ap-
 plies: i.e. the decision maker faces uncertainty with respect to annual income.
 The value of information as function of the frequency of income uncertainty (x
 %) can be defined as:
 EV PI = x
 100
 ∑
 t
 btEV PI(t), (14)
 where EVPI(t) for period t is given in equation (12).
 Instead of considering EVPI according to (14), the focus will be on the annual
 (dynamic) cost of uncertainty, specifying the loss due to uncertainty for each
 time period during which the income from investment is subject to uncertainty.
 To begin with it is assumed that the total available amount for investment
 can be spent at any time; later this simplification is removed by introducing
 period-specific budget constraints. Consider the wait-and-see model, assuming
 the return from future investments is determined by transition probabilities
 in Table 2, where the different states are given in terms of different levels of
 return on investment, following the case study example. The net return when
 investing It at time t is given by expression (7). Letting Ia = 10000, r0 = 0.05
 and b = 0.94, problem (8) subject to (2) is solved numerically with backward
 recursion 10000 times, using Matlab [Fackler]. Figure 1 depicts the probability
 of investment, with mean 1 %. The investment probability decreases over time,
 reflecting decreasing income expectations.
 A modification of EVPI (Definition 1) to a dynamic uncertainty cost relative
 to the expected investment can be formulated as follows:
 Definition 3 For Pr′t > 0, a dynamic relative EVPI, REPVI(t), can be defined
 for time period t as the weighted difference (cf. Definition 1):
 REV PI(t) = E[yt(I
 ∗
 t )− I∗t ]− E[yt(E(rt|rt−1), I∗∗t )− I∗∗t ]
 Pr′tIa
 . (15)
 With a large number of iterations, the expected net value of the investment
 at time t, E[yt(I∗t ) − I∗t ] in (15), can be approximated by the mean net value.
 Figure 2 depicts REVPI(t) in the above example, assuming both expected net
 value terms in (15) are approximated by the corresponding mean net values
 over 100000 iterations. Assuming b = 0.94 the relative EVPI as defined in (15),
 when averaged over time, is 0.17 6. The outcome with r0 = 0.11 (not depicted)
 is similar: the relative EVPI averaged over time is 0.15.
 Income Volatility and Uncertainty Cost
 Consider increasing the stability of the return on investment e.g. according to
 Table 3, increasing the probability of unchanged return to 0.95 (for all r < 0.3).
 6REVPI(t) is depicted whenever defined, i.e. for all t > 1 with Pr′t > 0
 10
0 10 20 30 40 50 60
 0
 0.01
 0.02
 0.03
 0.04
 0.05
 0.06
 probability of investmen
 t
 time index
 Figure 1: Probability of investment with b = 0.94, r0 = 0.05 (dash-dotted
 curve), r0 = 0.11 (solid curve)
 0 10 20 30 40 50 60
 −0.2
 −0.1
 0
 0.1
 0.2
 0.3
 0.4
 0.5
 time index
 REVPI(t
 )
 Figure 2: Relative EVPI(t), b = 0.94, r0 = 0.05 (100000 runs)
 11
Table 3: Transition Probabilities between States (ROI %)
 0.3 0.16 0.11 0.05 0
 0.3 0.01 0.3 0.4 0.28 0.01
 0.16 0.01 0.95 0.03 0.01 0
 0.11 0.01 0.01 0.95 0.02 0.01
 0.05 0.01 0.01 0.02 0.95 0.01
 0 0 0 0.02 0.03 0.95
 E.g. letting r0 = 0.05 and b = 0.94, the mean relative EVPI is 0.064. Decreas-
 ing the volatility of income here decreases the cost of uncertainty: uncertainty
 is less costly when the income is less volatile. In this example, the mean uncer-
 tainty cost depends on the underlying probabilities. Furthermore, decreasing
 the volatility of income increases cumulative investments, by more than 45 %
 in Model 1; the value of investment on average more than doubles. These ob-
 servations suggest that policy stability should be favored.
 Dynamic Relative Option Value
 The value of perfect information depends on the flexibility of decision-making
 in optimizing the expected value of investment. Assuming the investment prob-
 ability is positive using Model 1 (Prt > 0), a dynamic relative option value,
 O(t), can be defined as:
 O(t) = E[yt(I
 ∗
 t )− I∗t ]
 PrtIa
 −
 max{maxt E(yt − It), 0}
 Ia
 , (16)
 where the nominator in the second term models the net value obtained when
 maximizing expected forward start NPV at the beginning of the time horizon.
 E.g. with r0 = 0.05 and b = 0.94 the relative option value (not depicted) varies
 between 40 % and 42 % (the second term in (16) is zero). With r0 = 0.11,
 keeping other parameters unchanged, the mean O(t) is 0.35, even if in this case
 the second term in (16) is positive.
 Assuming a stationary income process, the second term in (12) is based on
 maximizing expected NPV; even with flexibility in decision-making, postponing
 investment is not optimal. In this case dynamic EVPI(t) in (12) corresponds
 to a dynamic option value of postponing investment. However, with stationary
 distribution EVPI(t) can be negative at t = 1, as the investment probability Ptt’
 is either 1 or 0 at t = 1. In this case the first term in (13) could be considered as
 an alternative formalization for the dynamic option value. The motivation is as
 follows: if Pr′1 = 1, the quantity cost due to uncertainty could be considered as
 zero; on the other hand, if Pr′t = 0 ∀t, the first term is equivalent to the value
 12
lost due to uncertainty: EVPI(t)=E[yt(I∗t ) − I∗t ]. Analogously, in the case of
 a stationary distribution, the relative uncertainty cost could be modelled using
 the relative option value (16) (REVPI(t) according to (15) is not defined for
 t > 1 as Pr′t = 0 for t > 1.)
 Fixed Term Policy Programs and Uncertainty Cost
 Agricultural policy programs typically have a fixed duration. For example, as-
 sume the same transition probabilities as in Table 2 apply for states given in
 terms of ROI % defined for 3 time periods (instead of one period as above). This
 modification changes the state space in terms of ROI %, increasing the variance
 between the states 7. Assuming the same parameters as in Figure 2, this mod-
 ification implies the mean REVPI over time is 19.1 % 8. The mean investment
 probability using Model 1 is 2 %, twice the mean investment probability with
 the original definition of the transition probabilities (cf. Figure 1). A longer
 term fixed policy increases the investment probability and value of investment
 in both Model 1 and Model 2; here the policy change increases the value of
 investment relatively more in Model 1. Thus, the uncertainty cost in terms of
 REVPI increases.
 The value of information depends on the the length of the time period with
 a certain income in the wait-and-see model, compared to expected value maxi-
 mization with uncertain income. Consider a special case of stable income where
 the rate of return in the wait-and-see model remains at rt for all future periods
 whenever investment is made at time t. For example, with b = 0.94, the mean
 REVPI in this case is approximately 160 % of the expected investment.
 An Application to Policy Planning
 Previous work based on a sector model of agriculture suggests that decoupling
 direct payments from production weakens the incentive for investment in dairy
 production and causes a temporary but significant slowdown in dairy invest-
 ments [Lehtonen2004]. A key issue in planning an investment subsidy program
 is to ensure a target level of productivity-enhancing investments, despite de-
 creasing expectations regarding future income. Assume the return rate is mod-
 elled as a Markov process that depends on the investment subsidy level. Letting
 b = 0.94 in the wait-and-see model gives Figure 3, depicting the cumulative in-
 vestment probability over the first 5 time periods as function of the investment
 subsidy (% of investment expenditure). E.g. with r0 = 0.05, it can be observed
 that to affect investments, the subsidy must increase from its current level 35 %
 to 45 %: this more than triples cumulative investments during first five periods.
 7Above when addressing the effect of income volatility on uncertainty cost, the state space
 in terms of ROI % remained the same, only the transition probabilities were modified.
 8Further assuming the different states are equally likely, the average REVPI over time is
 more than 21 %.
 13
0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65
 0
 0.1
 0.2
 0.3
 0.4
 0.5
 0.6
 0.7
 0.8
 0.9
 1
 su
 m
 t=
 1
 5
 Pr
 t(I a
 )
 investment subsidy %
 Figure 3: Cumulative investment probability as function of subsidy level, b =
 0.94, r0 = 0.05 (dash-dotted curve), r0 = 0.11 (solid curve)
 Uncertainty deteriorates the efficiency of investment subsidies. For example,
 if the start state is r0 = 0.05, and the investment subsidy is 0.35, the cumulative
 investment probability during first five years in the wait-and-see model is 8%,
 compared to 5% with expected value maximization with unobservable ROI %.
 Thus, policy uncertainty should be avoided.
 Financial Constraints
 Like in [Vercammen2003], assume now that the decision-maker decides at each
 time t on investment with period-specific financial constraints:
 It ∈ {0, I}, t = 1, ..., T. (17)
 With I = 200, b = 0.94 and r0 = 0.05, the investment probability is depicted
 in Figure 4. Period-specific financial constraints modify optimal investment be-
 havior (cf. Figure 1) and may explain the postponement of a large part of the
 investments. Financial constrains not only change the optimal investment be-
 haviour but also alter the dynamic uncertainty cost: period-specific constraints
 reduce the value of information to almost zero.
 14
0 10 20 30 40 50 60
 0
 0.05
 0.1
 0.15
 0.2
 0.25
 0.3
 0.35
 time index
 probability of investmen
 t
 Figure 4: Probability of investment with I = 200, b = 0.94, r0 = 0.05
 5 Investment by a Risk-Averse Agent
 The above investment models are based on assuming the decision-maker is a risk
 neutral. To take risk explicitly into account, a modification of a standard MDP
 is presented, following [Levitt and Ben-Israel2001]. Examples suggest that the
 investment decision is sensitive to risk. Furthermore, the uncertainty-investment
 relation is nonlinear.
 Risk in a Markov Decision Process
 The idea that risk affects decision-making is not new in agricultural economics
 [Hardaker et al.1997]; A traditional approach can be summarized as follows.
 Consider a utility function in exponential form:
 U(x) = 1− e−βx, (18)
 where β is a risk-aversion parameter. The expected value of utility (18) can be
 evaluated as [Hazell and Norton1986]
 E(x)− β
 2
 V ar(x). (19)
 Stochastic programming [Prekopa1995] has been previously applied to decision-
 making in agriculture under uncertainty, see e.g. [Hazell and Norton1986]. A
 15
dynamic objective function accounting for risk is defined next based on a stochas-
 tic programming approach presented in [Levitt and Ben-Israel2001] (with appli-
 cations to inventory control and the maintenance problem).
 A Stochastic Programming Model
 Definition 4 The recourse certainty equivalent (RCE) of a scalar random vari-
 able Z is defined as
 SU (Z) = sup
 z
 {z + EU(Z− z)}.
 where U is a concave function.
 Consider the quadratic utility function:
 u(x) = x− β
 2
 x2, (20)
 where β is a risk parameter. Applying Definition 4 to utility function (20) gives
 the RCE associated with this utility:
 Sβ(X) = E(X)− β2 V ar(X) (21)
 where β is a risk parameter. An agent maximizing the criterion in (21) is risk
 averse if β > 0.
 Definition 5 The quadratic recourse certainty equivalent (RCE) of the random
 sequence X = (X1, ..., XT ) is defined as [Levitt and Ben-Israel2001]
 Sβ1,...,βT (X) =
 T∑
 t=1
 bt−1Sβt(Xt) =
 T∑
 t=1
 bt−1{E(Xt)− βt2 V ar(Xt)} (22)
 where the βt parameters allow to model different risk attitudes in different stages.
 The ”utility” obtained at time t, Sβt is defined as the difference:
 E(Xt)− βt2 V ar(Xt). (23)
 The definition of the period-t RCE in equation (23) is analogous to RCE in
 equation (21). An alternative motivation for the definition of period t objective
 in equation (23) is given in equation (19).
 The wait-and-see model (Model 1) can be modified to take risk into account,
 applying the utility model (23). This implies investment probabilities corre-
 sponding to maximizing quadratic recourse certainty equivalent (Definition 5).
 16
0 10 20 30 40 50 60
 0
 0.005
 0.01
 0.015
 0.02
 0.025
 Figure 5: Investment probability (b = 0.94, r0 = 0.05), first with risk-neutral
 firm as in Fig. 1 (upper curve), second assuming exponential utility with risk
 parameter β = 10−4 (lower curve)
 A numerical example is depicted in Figure 5, with β = 10−4 9, applying the
 exponential utility model in equation (18) (cf. (19) and (21)). In this example
 uncertainty lowers cumulative investment probability by almost 50%, compared
 to the case with observable value depicted in Figure 1 (with r0 = 0.05). The
 investment probability depends on the amount of investment, dropping to zero
 at Ia = 10300. The relation between Ia and cumulative investment probability
 (not depicted) is nonlinear. A positive uncertainty-investment relation was ex-
 emplified in section 4, when addressing fixed term policy programs, assuming
 a risk-neutral decision maker. Taking risk into account in general modifies the
 uncertainty-investment relation.
 In general, transition probabilities depend on the timing on the investment.
 The probabilities may change e.g. due to a potential change in income and/or
 investment subsidies. Consider the special case of optimizing forward start NPV,
 assuming time-varying transition probabilities. Examples (not depicted) suggest
 that a cost associated with risk (variance) can be a source of an option value
 of postponing investment (in addition to period-specific financial constraints),
 9The Arrow-Pratt relative risk aversion (RRA) is defined at Ia as −IaU ′′(Ia)/U ′(Ia).
 Using β = 10−4 the RRA equals 1 at Ia; Arrow’s conjecture that RRA approximately equals
 1 is a common reference point. Recent empirical work considering the case of Turkish farmers
 [Binici et al.2003] suggests the mean estimate for β is 0.1.
 17
even if the income process is non-increasing in time.
 6 Conclusion
 This paper has studied the cost of income uncertainty in agricultural invest-
 ment. Applying a stochastic programming approach, a framework for studying
 the dynamic cost of uncertainty is presented, modifying the classical expected
 value of perfect information. Within the framework, the dynamic cost of un-
 certainty is studied numerically from the point of view of a case study. The
 numerical investment model also yields the optimal investment behavior of the
 representative (risk-neutral) farm. The investment model is extended to ac-
 counting for risk; numerical examples suggest that the investment decision can
 be sensitive to risk.
 The dynamic cost of uncertainty specifies the loss due to lack of information
 for each possible timing of the investment decision. Given the percentage of time
 the income from investment is subject to uncertainty, the dynamic uncertainty
 cost can be aggregated over time. The dynamic cost of uncertainty can be
 decomposed into two components: a quantity cost, due to reduced investments
 and a value cost, due to the lowering of the value of investment. The uncertainty
 cost depends on income volatility; in the special case of stationary income, the
 dynamic uncertainty cost is equivalent to a dynamic option value of postponing
 investment.
 The efficiency of investment subsidy programs is deteriorated by the uncer-
 tainty regarding future income. It remains a topic for future work to conduct
 a survey of the subjective probability distributions. In some case examples,
 the mean uncertainty cost is sensitive to underlying probabilities. In future
 work, the model can be applied to e.g. studying the investment subsidy needed
 to maintain target investments under uncertainty. The numerical optimization
 framework presented in this paper is applicable to other dynamic resource allo-
 cation problems where the dynamic cost of uncertainty is relevant.
 References
 [Alvarez and Stenbacka2004] Alvarez, L. and R. Stenbacka: 2004, ‘Optimal
 Risk Adaption: A Real Options Approach’. Economic Theory 23, 123–147.
 [Amegbeto and Featherstone1992] Amegbeto, K. and A. Featherstone: 1992,
 ‘Risk Costs and the Choice of Market Return Index’. Journal of Agricultural
 and Resource Economics 17(1).
 [Binici et al.2003] Binici, T., A. Koc, C. Zulauf, and A. Bayaner: 2003, ‘Risk
 Attitudes of Farmers in Terms of Risk Aversion: A Case Study of Lower
 Seyhan Plain Farmers in Adana Province, Turkey’. Turk J Agric For 27,
 305–312.
 18
[Birge and Louveaux1997] Birge, J. and F. Louveaux: 1997, Introduction to
 Stochastic Programming. Springer Series in Opereations Research.
 [Bo and Lensink2005] Bo, H. and R. Lensink: 2005, ‘Is the Investment-
 Uncertainty Relationship Nonlinear? An Empirical Analysis for the Nether-
 lands’. Economica 72, 307–331.
 [Dixit and Pindyck1994] Dixit, A. and R. Pindyck: 1994, Investment under
 Uncertainty. Princeton: Princeton University Press.
 [Fackler] Fackler, P., Compecon Toolbox for Matlab. http://www4.ncsu.edu/∼
 pfackler/compecon/toolbox.html:.
 [Hardaker et al.1997] Hardaker, J., R. Huirne, and J. Anderson: 1997, Coping
 with Risk in Agriculture. CAB International.
 [Hazell and Norton1986] Hazell, P. and R. Norton: 1986, Mathematical Pro-
 gramming for Economic Analysis in Agriculture. MacMillian.
 [Huisman et al.2003] Huisman, K., P. Korz, G. Pawlina, and J. Thissen: 2003,
 Strategic Investment under Uncertainty: Merging Real Options with Game
 Theory. CentER Discussion Paper No. 2003-06, Tilburg University, CentER.
 [Keswani and Shackleton2006] Keswani, A. and M. Shackleton: 2006, ‘How real
 option disinvestment flexibility augments project NPV’. European Journal of
 Operational Research 168, 240–252.
 [Knapp and Olson1996] Knapp, K. and L. Olson: 1996, ‘Dynamic resource
 managment: intertemporal substitution and risk aversion’. American Journal
 of Agricultural Economics 78.
 [Lagerkvist2005] Lagerkvist, C. J.: 2005, ‘Agricultural Policy Uncertainty and
 Farm Level Adjustments: the Case of Direct Payments and Incentives for
 Farmland Investment’. European Review of Agricultural Economics 32(1),
 1–23.
 [Lehtonen2004] Lehtonen, H.: 2004, ‘Impact of De-coupling Agricultural Sup-
 port on Dairy Investment and Milk Production in Finland’. Acta Agriculturae
 Scnadinavica Section C pp. 46–62.
 [Levitt and Ben-Israel2001] Levitt, S. and A. Ben-Israel: 2001, ‘On Modelling
 Risk in Markov Decision Processes’. in Optimization and Related Topics, A.
 Rubinov and B. Glover, editors pp. 27–40.
 [Mason and Weeds2005] Mason, R. and H. Weeds: 2005, Can Grater Uncer-
 tainty Hasten Investment. Working Paper, Department of Economics, Uni-
 versity of Southampton.
 [Prekopa1995] Prekopa, A.: 1995, Stochastic Programming. Kluwer.
 19
[Uusitalo et al.2004] Uusitalo, P., A.-M. Heikkila¨, and K. Pietola: 2004, ‘The
 Effect of Decoupling CAP-subsidies on Profitability of Livestock Bilding In-
 vestments’. In ”CAP-uudistus Suomen maataludessa”, H. Lehtonen ed. (in
 Finnish, attachments available upon request).
 [Vercammen2003] Vercammen, J.: 2003, ‘A Stochastic Dynamic Programming
 Model of Direct Subsidy Payments and Agricultural Investment’. Presenta-
 tion at the Joint Annual Meetings of the American Agricultural Economics
 Association and Canadian Agricultural Economics Society. Montreal.
 [Vercammen2006] Vercammen, J.: 2006, ‘Farm Bankruptcy as a Link between
 Direct Payments and Agricultural Investment’. Working Paper 2006-06
 Sauder School of Business, The University of British Columbia.
 20
ISSN 1795-5300
 MTT Discussion Papers 5  2006MTT Discussion Papers 5  2006
 Investment and the Dynamic 
Cost of Income Uncertainty: 
the Case of Diminishing 
Expectations in Agriculture
 Tiina Heikkinen & Kyösti Pietola