Professional Documents
Culture Documents
Model Risk in Stochastic DICE Models
Model Risk in Stochastic DICE Models
by
1
Contents
1 Introduction 3
4 Model Risk 18
4.1 Worst-case framework . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3.1 SDICE compared to SDICE1 . . . . . . . . . . . . . . . 21
4.3.2 SDICE compared to SDICE2 . . . . . . . . . . . . . . . 22
4.3.3 SDICE compared to SDICE3 . . . . . . . . . . . . . . . 22
4.3.4 SDICE compared to SDICE4 . . . . . . . . . . . . . . . 23
5 Discussion 24
5.1 SDICE results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2 SDICE comments . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3 Model risk results . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.4 Model risk comments . . . . . . . . . . . . . . . . . . . . . . . . 27
6 Conclusion 28
7 Bibliography 29
8 Appendix 30
2
Chapter 1
Introduction
Climate Change is one of the most discussed long-term problems society faces.
On the short-term, society will see a limited impact. Nonetheless, in the long
run, we will see the negative impacts (Tol, 2020). To reduce the impact of cli-
mate change emissions of greenhouse gasses should be reduced. However, the
costs for these reductions are hard to quantify. To model the impact of climate
change on society and the economy, Integrated Assessment Models (IAM) are
used. These models describe the effect of certain policies on the interaction
between climate and economy.
The Dynamic Integrated Climate-Economy model (DICE) is one of the
three IAM models used by the United States government to assess the impact
of climate change policies. The first version of this model dates back to 1992
and has been updated over the years. It is based on six state variables (CO2
concentration in the atmosphere, shallow and lower oceans; temperature in
the atmosphere and lower oceans; and capital), three policy variables (invest-
ment, consumption, and emission control rate), which are all evaluated for
time periods of five years. The model also takes several exogenous processes
into account (e.g. level of population and productivity factor). Even though
the model does have various limitations, it is seen as one of the most useful
climate-economy models. (Ackerman, Decanio, Howarth and Sheeran, 2009).
The DICE model reflects a deterministic optimization problem, which can
be solved in Excel or GAMS, a modeling system for mathematical optimiza-
tion. Due to its deterministic nature, the DICE-model does not account for
uncertainty in the world, indicating it is not an accurate reflection of the real
world. In order to get a more realistic model it is possible to incorporate
uncertainties into the model by introducing stochastic shocks. These shocks
represent an unexpected change in the state of the world, for example a higher
emission than predicted. The DICE models with stochastic shocks are called
SDICE.
Research about a stochastic implementation for the DICE model is limited
(Matsui, Myrvoll, Murakami and Shevcenko, 2021). However, some research
is available on the stochastic model with implemented shocks. Cai, Judd and
Lontzek (2012) formulate SDICE as a dynamic programming problem with
stochastic shocks. They use the outdated 2007 DICE model, but their solu-
3
tion approach is at the base of the current model. According to them, solving
SDICE as a dynamic programming problem is an efficient and feasible ap-
proach.
The approach of Cai et al. (2012) is extended by Ikefuji et al. (2020), in
which four stochastic shocks are introduced. Ikefuji et al. (2020) propose a
stochastic economy-climate framework and a solution method to solve SDICE.
In this framework only one of the four stochastic shocks is applied at a time.
(In addition it gives the possibility to incorporate heavy tailed stochasticy,
something beyond the scope for this thesis). Their solution method utilizes
Least Squares Monte Carlo, an approach introduced by Longstaff and Schwarz
(2001), to value American options. This approach is applied to SDICE in order
to get least squares coefficients for the state variables of the model.
Ikefuji et al. (2020) find that applying uncertainty to DICE can have a
substantial impact on the results of the model. This should be taken into ac-
count by users of the model, e.g. policy makers. They note that the results of
the model should not be taken literally. But the interaction between the main
variables and results should be seen as a guideline to assess policy decisions.
This thesis will touch light on the SDICE model of Ikefuji et al. (2020),
in which 4 stochastic shocks are introduced. These shocks are integrated in
the deterministic model, such that only one shock is active at a time. This
change of the deterministic model will be extended in this paper to incorporate
all shocks simultaneously. This thesis aims to explore the validity of this new
model. The research questions we explore are:
• What are the differences between the deterministic DICE model and the
new SDICE model?
• What are the differences between our SDICE model and the SDICE
model from Ikefuji et al. (2020)?
To answer the three questions we compare the results of the new model to the
deterministic model (Nordhaus, 2017) and to the SDICE-model of Ikefuji et
al. (2020).
First of all, our model is run and the results will be compared to the re-
sults from the original deterministic model (Nordhaus, 2017) as well as to the
results from Ikefuji et al. (2020). The comparison will focus on the first 20
time periods, to find the "short term" results of the model. The state variables
and welfare will be compared between the models to find whether our model
gives plausible results.
Secondly, the model risk of the proposed SDICE model will be evaluated.
Model risk in this context is described as the loss of value due to an inaccu-
rate model. This inaccuracy can be caused by e.g. the misuse of the model,
incorrect assumptions or inappropriate use of the model. Model risk is hard to
quantify. However, there are techniques to approach this quantification. One
of these techniques is the worst-case or minmax approach. In this approach,
multiple models are considered and the loss is minimized in the worst-case
4
scenario of each model. The minmax approach has been advocated in decision
theory as early as 1989 by Gilboa and Schmeidler (1989). It is extended to
measure model risk by Cont (2006), in which the worst-case approach is used
to measure model uncertainty in light of derivative pricings.
In this paper, we will apply the worst-case framework of Glasserman and
Xu (2014) to our SDICE model, in order to assess model risk. Their framework
is based on "relative entropy", in which they assess model risk by comparing
quantities of interest of an original model to an alternative model. This alter-
native model is a change from the original model, for example by multiplying
parameters of the original model. In order to assess model risk, the worst-
case scenarios of both models, bounded by a relative entropy constraint, are
determined. If the worst-case scenarios of a model, given a maximum relative
entropy, performs worse than those of a different model, we can conclude that
model risk in the former model is higher.
This paper proceeds as follows. In chapter 2, the stochastic DICE model
is given. In chapter 3 we introduce the discrete-time stochastic programming
problem for SDICE, the algorithm to solve this problem and its results. Next,
in chapter 4 the framework for model risk is laid out and applied to our model.
Then, chapter 5 contains the discussion about our SDICE model and the model
risk. Next, in chapter 6 the conclusion of this paper is given. Finally, the ref-
erences and the appendix are presented.
5
Chapter 2
Here, δ is the depreciation rate of capital and policy variable I the investment.
6
determined outside of the model. Furthermore, the second shock Z2,t is intro-
duced, which puts uncertainty into the CO2 -equivalent-emissions output ratio
σt . It will become Z2,t σt .
Et ∆
Mt+1 = (1 − b0 ) Mt + b1 Mts + (2.4)
3.666
Mts+1 = b0 Mt + (1 − b1 − b3 ) Mts + b2 Mtl (2.5)
Mtl+1 = b3 Mts + (1 − b2 ) Mtl (2.6)
Mts+1 and Mtl+1 are the increases in carbon concentration in shallow and lower
oceans. The values for the constants b0 , b1 , b2 and b3 are given in the appendix.
Here, Htl is the increase of temperature in lower oceans in degrees Celsius from
1900. Ft is the radiative forcing, i.e. a change in energy caused by natural
processes and climate change. Positive radiative forcing means that the earth
receives more energy from the sun than it sends to space. The constants a1, a2
and a3 are found in the appendix.
ωt = ϕt µθt (2.9)
with θ > 1 and ϕt the adjusted cost for backstop in period t. With this
fraction we have the following budget constraint, containing policy variables
consumption Ct and investment It :
Z3,t 1 − ωt − Z4,t ξHt2 Yt = Ct + It (2.10)
Two shocks are incorporated into the budget constraint, Z3,t , which puts
uncertainty
into the
damage-abatement fraction and changes it to
Z3,t 1 − ωt − ξHt2 . The fourth and final shock Z4,t puts uncertainty into the
damage parameter ξ and changes it into Z4,t ξ. These two shocks are then
added to get the above-mentioned budget constraint. All exogenous variables
are determined outside the model and the equations can be found in the ap-
pendix.
7
In the previous equations, the shocks that put uncertainty into the model
are introduced. The shocks Zj,t (j = 1, ..., 4) are specified as follows:
−τ 2
Zj,t = e 2 eτ ϵj,t (2.11)
with parameter ρ the discount rate and final period T = 100. Note that the
total welfare is the sum of discounted welfares per period. The discount rate is
set to ρ = 0.015. Therefore at time T, welfare in the final period is discounted
to approximately 22% of it’s original value. Thus, the model takes welfare far
into the future in account. Furthermore, all initial values and parameters can
be found in the appendix.
With these equations the SDICE model can be simulated. The model is
maximized over total discounted welfare, so we need to develop an solution to
this optimization problem. This solution is presented in the next chapter.
8
Chapter 3
In this chapter, the solving algorithm and results of the SDICE model will
be discussed, based on the algorithm proposed in Ikefuji et al. (2020). In
this paper, there will be two deviations from the previously mentioned model.
The first difference is the fact that all shocks are incorporated at the same
time. The second difference has to do with the forward simulation of the
final results. In Ikefuji et al. (2020), the results of the model are calculated
without taking the shocks into account. We will also integrate stochasticity
during the forward simulation. In order to prevent distorted results, due to
the uncertainty in the model, Monte-Carlo estimation will be used to calculate
the results. This papers focuses on the time in which the emission control rate
is optimized, i.e. the first 20 time periods (100 years).
• The stochastic shocks defined by Z1,t , Z2,t , Z3,t and Z4,t , are clustered
into Zt .
9
state variables, are determined outside of the model and can be found in the
appendix. The remaining variables in the model, Et , ωt , U (x), Wt , W and It
are the result of the previously mentioned variables and parameters.
Even though investment is a policy variable, we rewrite budget constraint
(2.10) into a formula in which It is determined by our model. Investment is
thus given by:
It = Z3,t 1 − ωt − Z4,t ξHt2 Yt − Ct (3.1)
It can readily be seen that the state variables are a Markov process or a
Markov chain, as the value of Xt+1 depends on the value of Xt and not on
any previous states. Since we use time steps of five years, SDICE can be seen
as a discrete-time stochastic controlled Markov process. In this process, the
transition of Xt to Xt+1 depends on the control variables, µt , Ct , and stochastic
shocks of the next period. Now, let us call this transition function f (x). We
have Xt+1 = f (Xt , Rt , Zt+1 ). This signifies that if we know the state variables
at time t, the control variables at time t and the shocks at time t + 1, it is
possible to determine the state variables at time t + 1
As we optimize over welfare, we want this function f to be such that welfare
is maximized in each period. Such a problem can be solved using dynamic
programming. Recursively and backwards in time, a Bellman equation will
be solved. Solving the model like this will break up the programming into
different parts for each period. For each time period the optimal values of
the control variables are calculated and put back into the model to get the
maximum pay-off. In our model, this payoff value is welfare.
As previously mentioned we want to optimize the Bellman equation of this
model with respect to Xt . We take a pay-off value Vt (Xt ) with (1 ≤ t ≤ T ).
Vt being welfare from time t to T . The Bellman equation is derived as follows:
We know that welfare W is maximized in the model. Now, assume we
are in period s, we want to maximize the remaining welfare until final period
T, given current information about the state variables, as these are known at
period t. So at time t, we look at the remaining welfare at [t, T ], adapted
to the filtration Ft , i.e. all information known at time t. Then, we call this
remaining welfare the value function V (X ). If we assume full information
available, Vt (Xt ) is defined as follows:
T
X Ws
Vt ( Xt ) = s
(3.2)
s=t (1 + ρ )
Now we combine 3.2 and 3.3. However, we only have information up until time
10
period t, so we take the expectation of Vt∗ (Xt ) and condition on filtration Ft .
T
Ws
Vt∗ (Xt ) = max E
X
F
s t
(3.4)
µt ,Ct s=t (1 + ρ )
The optimal values (µ∗t , Ct∗ ) for state variables Xt are given by
" #
Wt
(µ∗t (Xt ), Ct∗ (Xt )) = arg max E + V ∗
t+1 ( f ( X t , Rt , Zt+1 Ft
))
(3.8)
µt ,Ct (1 + ρ)t
With these equations, a policy maker can determine the optimal strategy
for the policy variables. Since the shocks for period t + 1 can be simulated
and all other variables at time t are known, the model can be assessed by
backwards induction. First the optimal control variables in the final period
are calculated. Then, these values are put back to evaluate the value function
at time T − 1, by maximizing over the value function with respect to time t
and the discounted welfare at that time period. These steps are repeated until
time period 1.
11
assumed that welfare in time periods [T + 1, ∞) is zero, giving VT∗+1 (XT +1 ) =
0. Therefore we can rewrite the value function at time T into
" #
WT WT
VT∗ (XT ) = max E + V ∗
( X T +1 FT = max
) (3.9)
T T +1
µt ,Ct (1 + ρ) µt ,Ct (1 + ρ)T
From here on, our algorithm works as follows recursively from time period
T to period 1.
• Draw random values for state variables Xt and shocks Zt . The state vari-
ables lie between 0.5 and 2.5 times the optimal state variables from the
deterministic DICE model. The shocks are drawn randomly according
to the errors described in chapter 2.
• From the regression of the previous step, we get the least-squares regres-
sion coefficients, which are denoted by β̂i,t . The approximation of the
value function is given by:
J
X
Vbt ( Xt ) = β̂j,t ψt (Xt ) (3.11)
j =0
These steps are then recursively repeated from time period T until time
period 2. Since we know the current state at time period 1, optimization of µt
and Ct is done over the initial state.
With the β̂i,t ’s obtained from the previous steps, it is possible to simulate
the model forward using Monte-Carlo estimation. Stochasticity is accounted
for in the forward simulation by using another set of shocks Zj,t ∗ . The forward
12
• Use state XtDET of the deterministic DICE model to calculate bounds.
Obtain a lower bound of 0.5XtDET and upper bound of 2.5XtDET +1 . If t =
1 use XtDET ∗
as our current state, otherwise use Xt , obtained from the
previous loop.
• Obtain optimal control variables µ∗t and Ct∗ by running the model with
state variables Xt∗ . From (3.8) we know that the optimal control variables
are the values for which the value function is maximised. Now use the β̂i,t
for the value function at time t + 1. The solution for time t is (Rt∗ , Xt∗ ).
The optimal control variables are:
J
Wt
(µ∗t , Ct∗ ) = arg max E β̂j,t ψt (f (Xt∗ , Rt , Zt+1 )) Ft
X
t
+
µt ,Ct (1 + ρ) j =0
(3.12)
In order to simulate one possible path of the SDICE model, the above-mentioned
steps are executed for all time periods. After the simulations are done, we use
Monte-Carlo to estimate the outcome of the model. If, for example, we want to
look at the temperature increase after 20 time periods H20 . N simulations are
run. In one simulation i (i = 1, ..., N ), we obtain the temperature at t = 20,
H20,i . We are interested in the Monte-Carlo estimator H20 ∗ of all H
20,i .
N
∗ 1 X
H20 = E[H20 ] ≈ H20,i (3.13)
N i=1
This Monte-Carlo approach can be extended to obtain estimates for all vari-
ables of the model.
In this paper, the first part of the algorithm is run for 10000 simulations,
which gives R2 ≈ 0.98 at time period 1 and R2 ≈ 0.999 at time period 20. For
the second part of the algorithm, 1000 simulations of the SDICE model are
obtained from the forward simulation.
The pseudo-code, describing the algorithm, can be found in the appendix.
3.3 Results
As we are interested in the validity of our model, we will review final results,
obtained from running the model in Matlab. These results are then compared
to the deterministic model and the model from Ikefuji et al. (2020). We start
by analyzing the policy variables, followed by the temperature increase. Lastly,
we look at the welfare. The model is simulated for 100 time periods. However,
we are interested in the part of the model in which µ is optimized. Therefore,
the focus of our results will be from time period 2 up to time period 20. In
order to obtain the results, 1000 sample paths of the model are simulated.
13
3.3.1 Emission control rate µt
In figure 3.1a, the simulations of µt are plotted. The initial starting values
of µ2 are between [0.25,0.39]. Most sample paths end at values in the range
of [0.7,1]. In time period 16 and 18 there are sample paths with low outliers.
It can readily be seen that some sample paths converge to the same solution
for certain time periods. An example of this can be seen at time period 12:
approximately at µ12 = 0.65, multiple sample paths flock at the same point.
More about this can be found in the discussion of the results.
By using the Monte-Carlo estimation, explained at the end of section 3.2,
the estimates of µt are calculated. These can be found in figure 3.1b, indicated
by the orange line. In addition to the the Monte-Carlo estimator, the 95%
confidence interval of the estimates along with the emission control rates of
the deterministic model (yellow line) are plotted. We find that the confidence
interval and thus the inaccuracy of our estimates increases over time, but only
with a small amount. The emission control rates of our SDICE model are
substantially larger than in the deterministic model. Thus, according to these
results, more money should be spent on fighting climate change than in the
deterministic model.
In table 3.1 the values from four different time periods of Ikefuji et al.
(2020) and the Monte-Carlo estimation are found. Our model has a higher µ
for all shown periods. It can easily be assumed that for all 20 time periods,
this observation holds, since µ is increasing over time.
Table 3.1: Comparison of µt between Ikefuji et al. (2020) and this paper’s
model
14
3.3.2 Consumption Ct
The trajectories of Ct are found in figure 3.2a. All sample paths follow an
upward trend. There is a substantial difference between all the values at
time period 20. The 2.5% quantile and 97.5% quantile are [519.52,831.04].
Therefore, 95% of the sample paths lie within this interval. On the upper and
lower ends of the sample paths a triangle shaped shape can be found. This
means that if consumption in a period is high or very low, it will be "corrected"
to a more average value next period.
The Monte-Carlo estimator of consumption is plotted in blue in figure 3.2b
and the consumption per period from the Nordhaus model is plotted in orange.
We observe that the values for consumption are very close to each other. From
period 4 up to period 19, the Nordhaus model has a higher consumption. After
period 19, the consumption in our model is higher.
From figure 3.2 it can readily be seen that the consumption in the Ikefuji
et al. (2020) model is higher than in our model. However, the difference is
very small.
Table 3.2: Comparison of Ct between Ikefuji et al. (2020) and our model
3.3.3 Temperature Ht
In figure 3.3a, the sample paths of the temperature are found. 95% of the
paths end up in the interval [3.21,3.34] at time period 20.In the first periods
the difference between simulations is very small, contrary to the later periods,
in which there is more variance between the sample paths.
15
The Monte-Carlo estimator of temperature is plotted in orange in figure
3.3b. The yellow line is the temperature from the Nordhaus model. The grey
area around the Monte-Carlo estimator is the 95% CI of this estimator. There
is a substantial difference between the two results. In the deterministic model
the temperature after 100 years is almost 4.5 ◦ C, while the temperature in
the stochastic model is approximately 3.3 ◦ C. In addition, the increase of
temperature in the former model is not slowing down, contrary to the latter
model, in which the temperature curve flattens.
The temperature in our model is lower than in the model from Ikefuji et
al. (2020). At time period 16 (after 80 years) the difference is almost 0.2 ◦ C.
Table 3.3: Comparison of Ht between Ikefuji et al. (2020) and this papers
model
3.3.4 Welfare Wt
In figure 3.4, the welfare at period t from this papers model is plotted as the
blue line. The welfare of the Nordhaus model is the orange line. Evidently,
the welfare of the deterministic model is higher in the first 14 time periods.
From period 18 forth, the welfare of the stochastic model is higher.
Unfortunately, there is not enough data known from Ikefuji et al. (2020)
to compare the welfare from their model to the welfare of this papers model.
16
Figure 3.4: Welfare per time period
17
Chapter 4
Model Risk
This chapter describes the model risk of our SDICE model. First of all, a
conceptual framework of worst-case model risk will be laid out. Secondly,
this framework will be applied to compare the worst-case scenarios between
our model with optimized control variables and our model with the Nordhaus
control variables. Thereafter, the results of other worst-case comparisons will
be shown. Finally, the results will be discussed.
f˜(x) f˜(x)
Z !
E[m ln(m)] = ln f (x)dx. (4.1)
f (x) f (x)
18
Now, consider a value function V(X), which measures some aspect of a
model. for example, in our SDICE model it can be the temperature increase,
welfare or one of the policy variables. We want to find the bounds of this value
function, with respect to the relative entropy constraint.
These bounds are the range of values E[V (X )] can have within the relative
entropy constraint. Glasserman and Xu (2014), proceed to argue that the
upper bound is defined by
1
inf sup[mV (x) − (m ln(m) − η )] (4.4)
θ>0 m θ
and the lower bound accordingly, with the only difference θ < 0. For the inner
supremum, the optimal m(X ) is given by
exp(θV (X ))
m∗θ (X ) = (4.5)
E[exp(θV (X ))]
i.i.d. This property is used to get the Monte-Carlo estimator of the likelihood
ratio m̂i (X, θ ). With this estimator, it is possible to approximate the relative
entropy evaluated at θ. Using Monte-Carlo, we get the following equations for
the likelihood ratio and relative entropy:
exp(θV (Xi ))
m̂i (X, θ ) = 1 PN
(4.6)
N j =1 exp(θV (Xj ))
N
1 X
η̂ (θ ) = m̂i (X, θ ) ln (m̂i (X, θ )) (4.7)
N i=1
4.2 Implementation
Now, an example of the worst-case framework will be applied to our SDICE
model. For this implementation, the model risk between our model, with
optimized control variables, will be compared to our model with the control
variables from the Nordhaus model. The quantity of interest for this example
is the temperature at t = 20. Note that this implementation can be extended
to implement other quantities of interest or different models.
We will evaluate model risk based on the simulations from subsection 3.3.5,
in which a Monte-Carlo evaluation is described of the sample paths for the de-
terministic control variables, and the simulations of our model. As said before
we are interested in the temperature at period 20. Thus, the value function,
as described in 4.1, becomes V (X ) = H20 . Note that V(X) is a general value
function and not necessarily the same as the value function used in SDICE.
19
Define H20,i as the temperature at period 20 from the i-th forward simula-
tion of the SDICE model. Then, the likelihood ratio of simulation i, dependant
on θ, is given by:
exp(θH20,i )
m̂i (θ ) = 1 PN
(4.8)
N j =1 exp(θH20,j )
From 4.8, N likelihood ratios are acquired for a specific θ. These likeli-
hood ratios are put into the formula of (4.7) and the relative entropy η̂ (θ ) is
obtained. In addition to the relative entropy, we are interested in the bounds
of the value function. Therefore, the expectation of the value function un-
der another model, is calculated by Monte-Carlo estimation: E[m(θ )H20 ] =
1 PN
N i=1 m̂i (θ )H20,i .
Combining the results of the previous paragraphs, we find the relative en-
tropy necessary to move from the original result H20 , to the result under
the alternative model (m(θ )H20 ). However, in order to assess the worst-case
model, we need to evaluate for multiple θ’s. For this specific example we let
θ ∈ [−150, 150].
As we are interested in the model risk between two models, the previous
steps are also done for the simulations of the SDICE model with the Nord-
haus control variables. Both results are plotted and we get the following figure:
The blue line shows the worst-case scenario performance of the SDICE
model with the control variables from the deterministic model. The red line
shows the worst-case scenario performance of our model with optimized control
variables. If θ = 0, there is no relative entropy and the expected temperature
is the Monte-Carlo estimate as discussed in section 3.3. The dotted lines are
the corresponding Monte-Carlo estimator values.
20
Model risk is defined as the distance between the worst-case scenario line
and the dotted line. Even though the starting point of the two models is not
the same, it is obvious that the worst-case scenario of the deterministic control
variables performs worse than the optimized control variables.
4.3 Results
In this section, we will describe several evaluations of model risk. First of all,
we introduce four models in which only one shock is active. These models
are called SDICE1 , SDICE2 , SDICE3 and SDICE4 , in which the subscript
refers to the shock integrated in the model (e.g. SDICE1 refers to SDICE with
only shock Z1 active). These alternative models are the same as our SDICE
model, except for the amount of shocks used.
In the SDICE models, total discounted welfare is maximized. We will
compare the worst-case scenarios of this welfare, in order to assess the model
risk. In order to compare the model risk between all models, the techniques of
the previous chapter are used.
21
4.3.2 SDICE compared to SDICE2
The worst-case scenario’s for SDICE and SDICE2 are plotted in figure 4.3a,
using the same notation as in 4.3.1. The difference in total welfare between
the simulations of SDICE2 is very small. Therefore, figure 4.3a does not give
easily interpretable results. In order to compare the model risk between the
two models, they are laid over each other in figure 4.3b. We observe that
model risk for the SDICE2 model is very small, compared to the model risk
of SDICE.
22
Figure 4.4: Worst-case evaluation of SDICE vs. SDICE3
23
Chapter 5
Discussion
In this chapter, we will discuss the results of our model and the model risk
assessment. First, we examine the SDICE analysis, then we discuss possible
shortcomings of the model. Next, the model risk results are reviewed. Lastly,
we discuss the flaws of our model risk analysis and propose possible future
research.
24
trol variables are put into the model, as described in 3.3.5. This seemed counter
intuitive at first, since welfare is optimized and thus expected to be higher.
However, the solution algorithm optimizes Vt in a way, such that the welfare
at final time period T , is maximized and not the welfare per time period. This
effect can be seen in the model, as welfare of our model is higher after 18
periods. To extend this, in figure 5.1 the welfare per time period from period
20 up to period 40 is plotted for the original Nordhaus model and our SDICE
model. It is apparent that welfare in the SDICE model is higher. Thus we can
conclude that our model performs better than the deterministic model after
20 periods. Since the deterministic DICE model and the alternative SDICE
model are optimized over 500 years, our model is also optimized over the same
time period. If a policy maker wants to apply optimal policies for the short
term, they should rerun the model with a lower final time period.
25
gives inaccurate results for some sample paths. However, the effects of this
inaccuracy are reduced by the Monte-Carlo estimation, which averages over
all simulations. Therefore, the results are still applicable. In future research,
it is recommended to utilize another optimization method.
Taking another look at the simulations of µt , the assumption that it is
always 1 after 20 time periods, can be seen as invalid, as a lot of sample paths
seem to not end up around 1 at period 20. However, from the distribution of
µ20 in figure 5.2, it is obvious that this assumption is plausible. Most sample
paths end up around 1. Furthermore, the model is stochastic, therefore there
will always be sample paths which have outliers.
The simulations of both µt and Ct at upper and lower bounds of the sam-
ple paths in figure 3.1a and figure 3.2a have a triangular shape. At first sight
this can seem wrong. However, the triangular shape can be explained by the
stochasticity of the model. If a shock becomes very large or small, the corre-
sponding optimal control variables can deviate substantially from most other
simulation outcomes in that period. Consider for example shock Z3,t large.
Then, there is an increase in output able to be spent on consumption. As a
result, consumption will increase for that time period. However, in time period
t + 1 this shock is very different and the spendable output is expected to be
lower. Therefore, consumption will be lower in period t + 1 and this specific
sample path will have a triangular shape at time t.
26
the model. They found that applying shock Z3 , corresponding to SDICE3 ,
has almost no impact, meaning that the results are close to the deterministic
results. Furthermore, shocks Z1 and Z4 have a moderate impact. Lastly, they
found that shock Z4 has substantial impact on the model, especially if shocks
become large or small. These results relate closely to our findings about model
risk. In the models with more uncertainty, the model risk is higher.
An interesting result is that of SDICE3 . Namely, the model risk is almost
the same as in our model. Therefore, if a policy maker wants to put uncer-
tainty into the model trough the damage-abatement fraction (i.e. shock Z3 ),
they can choose to incorporate all shocks simultaneously, without increasing
model risk substantially.
27
Chapter 6
Conclusion
28
Chapter 7
Bibliography
Ackerman, F., DeCanio, S. J., Howarth, R. B., and Sheeran, K. (2009). Limi-
tations of integrated assessment models of climate change. Climatic change,
95(3):297–315.
Cai, Y., Judd, K. L., and Lontzek, T. S. (2012). Dsice: a dynamic stochastic
integrated model of climate and economy.
Cont, R. (2006). Model uncertainty and its impact on the pricing of derivative
instruments. Mathematical finance, 16(3):519–547.
Glasserman, P. and Xu, X. (2014). Robust risk measurement and model risk.
Quantitative Finance, 14(1):29–58.
Ikefuji, M., Laeven, R. J., Magnus, J. R., and Muris, C. (2020). Expected
utility and catastrophic risk in a stochastic economy–climate model. Journal
of Econometrics, 214(1):110–129.
Shevchenko, P. V., Murakami, D., Matsui, T., and Myrvoll, T. A. (2021). Im-
pact of covid-19 type events on the economy and climate under the stochastic
dice model. Environmental Economics and Policy Studies, pages 1–18.
29
Chapter 8
Appendix
Parameter values:
30
Initial values at t=1:
31