The Engineering Economist

A Journal Devoted to the Problems of Capital Investment

ISSN: 0013-791X (Print) 1547-2701 (Online) Journal homepage: http://www.tandfonline.com/loi/utee20

Cash-Flow at Risk valuation of mining project using

Monte-Carlo simulations with stochastic processes
calibrated on historical data

Mathieu Sauvageau & Mustafa Kumral

To cite this article: Mathieu Sauvageau & Mustafa Kumral (2018): Cash-Flow at Risk valuation
of mining project using Monte-Carlo simulations with stochastic processes calibrated on historical
data, The Engineering Economist, DOI: 10.1080/0013791X.2017.1413150

To link to this article: https://doi.org/10.1080/0013791X.2017.1413150

Accepted author version posted online: 05

Jan 2018.

Submit your article to this journal

Article views: 25

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=utee20
ACCEPTED MANUSCRIPT

1 Cash-Flow at Risk valuation of mining project using

2 Monte-Carlo simulations with stochastic processes
3 calibrated on historical data
4 Mathieu Sauvageau1,, Mustafa Kumral1,
5 McGill University, Department of Mining and Materials Engineering, 3450 University,
6 Montreal, Quebec, Canada, H3A 0E8

7 Abstract

8 Mining projects are subject to multiple sources of market uncertainties

9 such as metal price, exchange rates and their volatilities. Assessing a mining
10 project exposure to market risk usually requires Monte-Carlo simulations
11 to capture a range of probable outcomes. The probability of a major loss
12 is extracted from the probability density function of simulated prices at a
13 given time into the future. This paper proposes an approach to calibrate the
14 stochastic process to be used in Monte-Carlo simulations. The simulations
15 are then used for measuring the cash-flow at risk of a mining project. To
16 assess the performance of the proposed approach, a case study is conducted
17 on a mining project. The results show that the calibration approach is robust
18 and apt at fitting various stochastic processes to historical observations.
19 Keywords: stochastic process, Kalman filter, commodity prices, Market
20 risk, particle swarm optimization

Email address: mathieu.sauvageau@mail.mcgill.ca (Mathieu Sauvageau)

Preprint submitted to The Engineering Economist November 27, 2017

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

21 1. Introduction

22 Mining projects are exposed to significant financial risks. Since the pro-
23 duced commodity is traded in markets that are affected by supply and de-
24 mand dynamics, mining companies are subject to market risk. Commodity
25 price and its volatility are among the most important sources of risk. For
26 example, gold price, which was approximately 1,800 $/ounce in 2011, is ap-
27 proximately 1,150 $/ounce at the end of 2016. The price mechanisms of
28 commodities are complex because many commodities and their derivatives
29 are traded in spot and futures markets. They are also indirectly related to
30 capital markets because mining companies usually use equity markets for
31 financing purposes.
32 Because commodities are generally traded in US dollars, exchange rates
33 fluctuations also affect mining projects operating outside of the US. The joint
34 effect of price and exchange rates dynamics can affect the net present value
35 (NPV) of a mining project (Topal et al., 2016). With the assumption that
36 financial markets are efficient, exogenous information constantly changes the
37 expectations of parties involved in the transactions. These market partici-
38 pants can be short hedgers such as mining companies who seek to sell their
39 products in advance to prevent losses from sudden drop in the price of the
40 metal they produce. The other party of the transaction can be a long hedger
41 who seeks to fix a ceiling on its production inputs. Other stakeholders such as
42 arbitrageurs and speculators try to profit from market inefficiencies. When
43 the market is unfavorable for short hedgers, they need to pay a premium (sell
44 their production at a higher discount). Because this can be very expensive,
45 mining companies cannot hedge the entirety of their production. They need

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

46 to correctly measure their risk exposure to market fluctuations and hedge

47 the right amount of their production depending on the risk they are willing
48 to take.
49 In the recent literature, the Kalman Filter (KF) has been used to value
50 real options in engineering projects. Real options generally add value to en-
51 gineering projects but since they are complex to evaluate and interpret, they
52 are not widely used (Block, 2007). Real options are generally used before the
53 project actually starts, to verify that the added flexibility increases the NPV
54 of a project. This paper proposes a different use of the KF calibration work-
55 flow for market risk management. Three different stochastic processes are
56 used to model dynamics of gold, silver, platinum and crude oil prices. These
57 processes are the geometric Brownian motion (GBM), the Merton jump diffu-
58 sion (MJD) and the Heston Stochastic volatility (SV) models. The GBM and
59 MJD processes are calibrated using a maximum likelihood estimation rou-
60 tine. The Heston SV is calibrated using the Unscented Kalman filter. To find
61 the optimal parameters of the Heston SV model, a sub-optimal solution is
62 first found using particle swarm optimization metaheuristics approach (Eber-
63 hart and Kennedy, 1995), from which the optimal solution is found using a
64 gradient-based method. Using calibrated parameters from historical obser-
65 vations, simulations of gold prices are realized over a three-month period to
66 calculate the Cash-Flow at Risk (CFaR) of a gold-mining project. Results
67 show that the MJD and Heston SV models are better suited to reproduce
68 tail behaviour of commodity returns.
69 The originality of this paper lies on the derivation of a robust calibration
70 framework using a hybrid metaheuristics optimization approach and the KF

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

71 to perform CFaR analysis of a mining project. In this research, price volatility

72 in Heston SV model and jump component in MJD model are calibrated
73 through a combined approach using particle swarm optimization and Kalman
74 filter. The most significant advantage of the paper is that unlike real options
75 valuation, the proposed methodology can be used at any stage of a mining
76 project to assess how cash flows are exposed to market risk in the near future.
77 This gives the opportunity to better assess the exposure to excessive cash-
78 flows drawdown and thus, the ability to implement hedging strategies when
79 markets are stable and prices of derivatives relatively low.

80 2. Literature review

81 The Value at Risk has been used by risk managers over the last decade
82 (Jorion, 2001). The approach consists of determining the distribution of
83 profit and losses for an investment over a given time horizon. Then, the
84 lower tail of the distribution is analyzed. The threshold for the lower-tail can
85 be set, for example, to 5%. The value corresponding to the 5th percentile of
86 the profit and losses distribution and can be interpreted as the minimum loss
87 that an investment will suffer 5% of the time. The Value at Risk approach
88 is useful to infer risk measures on a portfolio of assets. In the case of mining
89 investments, another approach is to use the cash-flow at risk measure of risk
90 (Alesii, 2003). Because the fixed capital costs of mining projects are sunk and
91 therefore very illiquid, the CFaR provides a more useful metric for valuing
92 the exposure to market risk. This is because mining companies suffer when
93 cash flows become negative after a sudden drop in commodity prices. This
94 approach can also be implemented in a Real Options valuation framework.

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

95 The Monte-Carlo simulations approach can be used to calculate the CFaR.

96 This approach is based on the simulation of multiple price paths using a
97 stochastic process such as the geometric Brownian motion (GBM). GBM is
98 the most widely used stochastic process to model price dynamics of stocks,
99 commodities or options contracts. Because of its simplicity, it is generally
100 used as a reference to benchmark other stochastic processes.
101 The logarithm of returns on commodity futures are often positively skewed
102 with a significant amount of excess kurtosis (Gorton and Rouwenhorst, 2006).
103 This implies that the distribution of commodity logarithm of returns can be
104 better described with a fat-tailed distribution. Schöne (2014) described two
105 mechanisms contributing to the fat-tailed behaviour of commodity returns.
106 The first mechanism is SV and the second is the inclusion of jumps in prices.
107 The Heston SV (Heston, 1993) model can be used to describe behaviour of
108 volatility clusters in periods of high and low volatility. The MJD model (Mer-
109 ton, 1976) adds a jump component to the stochastic process. Schöne (2014)
110 studied how the choice of the stochastic process in a real options valuation
111 framework can influence the NPV of a mining project. The author calibrated
112 the GBM, SV and MJD models to commodity spot prices by minimizing the
113 square difference between the probability density function (PDF) of models
114 and the kernel density estimated PDF of the observations. Hammond and
115 Bickel (2013) studied how the choice of the stochastic process can affect the
116 NPV rankings of mining projects with embedded real options.
117 One of the main tasks of stochastic volatility modeling is to calibrate the
118 model with actual market observations. Since they were primarily developed
119 for the pricing of options on financial assets, such models are calibrated on

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

120 observed option prices (Gatheral, 2011). In this framework, the goal is to
121 adjust parameters of the model to fit the actual implied volatility surface of
122 all traded options for a given security. However, when modeling stochastic
123 volatility in commodity markets, historical options on commodity spot or
124 futures may be hard to find. Another class of models instead uses historical
125 future contracts observations with the Kalman filter to calibrate the model
126 parameters. For example, Schwartz (1997) used the KF to describe the
127 evolution of commodities spot price subject to a stochastic convenience yield.
128 A two-factor commodity model consisting of a short-term price deviation
129 and a long-term trend was later developed (Schwartz and Smith, 2000). The
130 authors used the KF to adjust model parameters on historical observations
131 of commodity future prices. Hammond and Bickel (2013) showed how the
132 choice of the stochastic process can affect the NPV ranking of oil investments.
133 The authors consider a simple case with no options and the case where real
134 options are embedded in the NPV analysis.
135 To use the KF, the stochastic process is first derived in a state-space
136 form. One of the main assumptions of the KF is that the equations describ-
137 ing the transition between different observations are linear. Javaheri et al.
138 (2003) showed that the Unscented Kalman filter and the particle filter are
139 better suited when the transition equation of a stochastic process such as
140 the Heston SV model is non-linear. Schwartz and Smith (2000) used a quasi
141 maximum-likelihood routine to find the optimal parameters of the model.
142 The routine uses a gradient-based optimization method to derive the opti-
143 mal parameters of the model. When the objective (or cost) function have
144 several parameters to adjust, gradient-based optimization may quickly get

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

145 stuck in a local maxima (Sauvageau and Kumral, 2016). Alternate initial
146 value vectors can converge onto different optimized parameters. Schwartz
147 and Smith (2000) observed a similar issue in the two-factor model. The au-
148 thors used several different initial values to prevent the optimization routine
149 from not getting stuck in a local minimum. A better approach is to use
150 metaheuristic algorithms to ensure a robust convergence when problems are
151 non-linear (Subulan et al., 2017). The next section presents in more details
152 the stochastic processes, and the calibration routine used in this paper.

153 3. Methodology

154 3.1. Merton Jump-Diffusion models

155 Merton (1976) extended the GBM to account for random price discon-
156 tinuities or jumps. The frequency of such jumps is modeled with a Poisson
157 process which is independent from the GBM diffusion process. Using the no-
158 tation of Remillard (2013), the MJD can be represented with the following
159 equation where S0 is the spot price at initiation, and S(t) is the expected
160 price at any given point in time:

S(t) = S0 eX(t) (1)

with
N (t)
σ2 X
X(t) = (µ − λκ − )t + σW (t) + ξj (2)
2 j=1

161 where µ is the drift parameter, λ is the intensity of the Poisson process, κ is
162 the expected value of the jump, σ is the annualized standard deviation, W (t)

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

N
P (t)
163 is a standard Brownian motion and ξj is the compound Poisson process
j=1
164 with ξj ∼ N (γ, δ 2 ). When the jump process is Gaussian, the κ parameter is:

δ2
γ+
κ=e 2 −1 (3)

165 These equations can be used to simulate the MJD model. However, in the
166 valuation or risk management workflows, the model needs to be calibrated
167 with historical observations. The maximum likelihood (ML) method can be
168 used to perform such calibration (Remillard, 2013). The probability den-
169 sity function (PDF) of price returns can be represented with the following
170 equation:

(r−a−kγ)2
∞ 1 σ 2 h+kδ 2
X
−λh (λh)
k
e− 2
fRi (r) = e p (4)
k=0
k! 2π(σ 2 h + kδ 2 )
171 To perform the ML optimization of the parameters, the experimental
172 PDF is approximated with a kernel density estimator. Then, the parameter
173 set θ = (µ, σ, λ, γ, δ) is optimized through a gradient based minimization rou-
174 tine in MATLAB. For example, the fminunc (find minimum of unconstrained
175 multivariable function) in MATLAB can be used.

176 3.2. Stochastic Volatility models

177 Another class of stochastic models capable of reproducing fat tails is

178 the Heston SV model. The evolution of spot price in this model can be
179 represented with the following set of equations (Gatheral, 2011):

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

p
dS(t) = µS(t)dt + V (t)S(t)dW1
p
dV (t) = κ(θ − V (t))dt + η V (t)dW2 (5)
E[dW1 dW2 ] = ρdt,
180 where µ is the drift, κ is the speed of mean reversion, θ is the long-term
181 mean of volatility, η is the volatility of the volatility parameter and ρ is the
182 correlation between the two standard Brownian motions dW1 and dW2 .
183 The calibration of the SV model to historical observations is a non-trivial
184 task. Javaheri et al. (2003) proposed the use of the Unscented Kalman filter
185 to perform the calibration. This algorithm is suitable when the transition
186 equation is non-linear as in the SV model. A Gaussian approximation is
187 made using points in a probability distribution having the same mean and
188 covariance as the underlying distribution. These points are called sigma
189 points. The weight assigned to each sigma point is defined as:

(m) λ
W0 = (6)
na + λ
and
(c) λ
W0 = + (1 − α2 + β) (7)
na + λ
190 and for i = 1...2na

(m) (c) λ
Wi = Wi = (8)
2(na + λ)
191 where λ, α, β are tuning parameters for the sigma points and na is the
192 dimension of the augmented state. In the algorithm, λ = α2 na − na . In this
193 paper, the parameters are α = 0.001, β = 2 which are optimal for a Gaussian
194 approximation. Then, the sigma points are passed through the non-linear

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

195 transition equation of the KF and a Gaussian distribution is approximated

196 through the transformed points. The rest of the Unscented Kalman filter
197 algorithm is presented in the Appendix.
198 The parameters of the Unscented Kalman filter α = (µ, κ, θ, η, ρ) are
199 estimated through the quasi maximum likelihood (QML) method. The QML
200 estimate is given by:

N
X zk − mzk
L1:N = ln(Pzk zk ) + (9)
k=1
Pzk zk
201 where zk and Pzk zk are defined in the Appendix.
To calibrate the Heston SV model, a first sub-optimal solution is found
using the particle swarm optimization routine (Eberhart and Kennedy, 1995).
Instead of working with a single vector of initial parameters, the particle
swarm optimization tests a population of vectors. Each vector corresponds
to a single particle which is allowed to move on the search space at a given
speed. Then, at each iteration, the algorithm searches for a better solution
without using the partial derivatives of the cost function. The particles are
allowed to move at a given velocity given by:

i
vk+1 = ωk vki + c1 r1 (Pki − xik ) + c2 r2 (PkGlobal − xik ) (10)

202 where ω is the inertia weight of the particles, r1 and r2 are random draws
203 from the uniform distribution, and (c1 , c2 ) are the acceleration coefficients of
204 the particles. The algorithm is stopped after a number of iterations selected
205 for convergence criteria and the sub-optimal solution is used as the starting
206 values in the gradient-based optimization routine. The routine is presented
207 in Algorithm 1. Then, as with the MJD model, the set of parameters is

10

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

208 optimized through the gradient-based algorithm in MATLAB.

Algorithm 1 PSO
1: Initialize a particle swarm with random values positions and velocities

from D dimensions in the search space

2: while number of iterations k < N do

3: for Each particle i do

4: Evaluate objective function f (xik , yki )
5: Update the particle best solution Pki
6: Update the swarm best solution PkGlobal
7: Update the velocity with Equation 10
8: Update the position of all particles using xik+1 = xik + vk+1
i

9: end for
10: end while

209 3.3. Estimation of CFaR

210 Unlike financial institutions, where Value at Risk measure is the norm to
211 infer market risk, mining companies own physical assets and are dependent on
212 commodity prices. In general, their investment cannot be liquidated without
213 any major loss. For this reason, the risk metric studied in this paper is
214 the CFaR. It can be interpreted as the minimum cash-flow from operations
215 (CFO) that the mining project can earn for a given period. When the risk of
216 negative cash flows is high, mining companies seek ways to hedge the market
217 risk by using commodity derivatives such as futures contracts.
218 Different approaches can be used to calculate CFaR. The historical ap-
219 proach consists of fitting a statistical distribution to historical observations

11

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

220 and determining the 5th percentile of the cumulative density function. The
221 drawback of this approach is that it is unlikely the past will repeat itself and
222 the true risk could be underestimated. Another approach supposes that the
223 distribution of price returns can be approximated with a normal assumption.
224 The CFaR is extracted from a PDF fitted to the distribution. However, if the
225 production is hedged with derivatives contracts, the CFaR distribution be-
226 comes non-linear and becomes very difficult to fit with an analytical method.
227 In this paper, Monte-Carlo simulations are used to calculate the CFaR. The
228 parameters inferred from calibrated stochastic processes are used to perform
229 MC simulations of price paths. The simulation parameters are calibrated on
230 historical data but they can be modified to account for econometric predic-
231 tions or can be stress-tested to estimate the error on the CFaR predicted
232 value. The stochastic processes of interest are the GBM, the MJD and the
233 Heston SV.
234 To calculate the CFO, the following function is used, with stochastic gold
235 prices from the GBM, MJD and Heston SV models:

CF O =[throughput ∗ price ∗ grade ∗ recovery]−

[mining costs ∗ (throughput + waste production)]− (11)

[processing costs ∗ throughput] − F ixed costs

236 Then, the different CFO are aggregated on daily, weekly, monthly and
237 quarterly periods, and ranked. The CFO corresponding to the 5th percentile
238 is taken as the calculated CFaR. When the CFaR is below a threshold for
239 a given quarter, the company can completely hedge its production by using
240 options on gold futures contracts. Since the gold company is a producer,

12

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

241 it is considered long in the underlying commodity. The proper action to

242 hedge against a decrease in gold prices is to buy put contracts on the futures
243 contracts. This effectively puts a floor price on the underlying commodity in
244 exchange for a premium.

245 4. Case study

246 4.1. Data

247 The dataset used to perform the analysis consists of gold, silver and plat-
248 inum spot prices as well as a continuous series of the CME group crude
249 oil nearest maturing contracts. Historical observations of gold and silver
250 spot prices are from the London Bullion Market Association (LBMA). The
251 LBMA fixes gold prices twice a day by matching bid and ask prices of buy-
252 ers and sellers of the commodity. The spot platinum dataset is from The
253 Johnson Matthey Base Prices dataset consisting of the company’s quoted
254 selling prices. The crude oil contracts are taken from the Chicago Mercantile
255 Exchange (CME) group database and consists of raw assembled contracts
256 without adjustment. Figure 1 presents the time series for the four datasets
257 ranging between 8 February 1993 and 8 November 2016. The time period
258 includes the 2009 financial crisis. The dataset was split in two different pe-
259 riods. The first period corresponds to the pre-crisis era while the second
260 corresponds to the price dynamics after the crisis.
Insert Figure 1 about here
261 To calibrate the models, price needs to be converted to logarithm of
262 returns. Figure 2 represents respectively log returns and spot price of gold in
263 two different periods. The first period ranges between 8 February 1993 and

13

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

264 27 December 2004. The second period ranges between 29 December 2004
265 and 8 November 2016. It is worth noticing how volatility tends to cluster
266 on the first plot of Figure 2. This characteristic can be observed in both
267 periods, but the volatility tends to be higher in the second period. The same
268 phenomenon can be observed for the rest of the dataset.
Insert Figure 2 about here
269 In Table 1 descriptive statistics of the distribution of commodity returns
270 in the three different periods are presented. The returns are calculated using
271 daily observations but the mean and variance are annualized so they are
272 easier to interpret. In all cases, the distribution of commodities price returns
273 are characterized by an excess kurtosis which is a characteristic of fat-tailed
274 distributions. The positive skewness of gold and platinum spot returns in
275 Period 1 implies that there is more weight in the right tail that the left tail.
276 However, in Period 2, this characteristic is inverted and both distributions
277 become negatively skewed. Finally, the minimum and maximum values are
278 generally higher in the second period.
279 Another way to detect extreme values or analyze the dispersion of the
280 distribution is to use a box plot. Figure 3 shows box plots for the four
281 commodities in the two different periods. The central part of the box plot
282 is limited by the median and the 25th and 75th percentiles. The whiskers
283 extend to the 10th and 90th percentiles. The data points that are outside the
284 whisker range are considered as outliers. Both plots share the same y-axis.
285 It is easy to see that both periods contain a considerable amount of outlier
286 points and that Period 2 has more outliers than Period 1. Again, Figure 3
287 shows that the silver spot price returns dataset contains the most extensive

14

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

288 outliers.

Table 1: Descriptive statistics for the gold, silver and platinum spot price returns as well as
the crude oil futures contracts returns. The period 1 covers the weeks between 8 February
1993 and 27 December 2004. The second period covers the weeks between 10 January
2005 and 5 December 2016. The total period covers the whole time range.
counts ann. mean ann. std skewness kurtosis min max
gold 2913 0.0191 0.1271 0.6277 8.1926 -0.054 0.0701
Period 1

silver 2945 0.0376 0.2536 -0.0829 4.6821 -0.112 0.0991

platinum 3026 0.0511 0.1952 0.4949 9.011 -0.0741 0.1393
crude oil 2895 0.0687 0.3579 -0.3326 3.7277 -0.1654 0.1423
gold 2907 0.0389 0.1929 -0.4234 4.8751 -0.096 0.0684
Period 2

silver 2940 0.0417 0.3609 -0.5557 8.8629 -0.1869 0.1828

platinum 2878 -0.0161 0.2435 -0.7157 9.9826 -0.1554 0.1264
crude oil 2928 0.0256 0.3944 0.0669 5.2571 -0.1576 0.1641
gold 5820 0.029 0.1633 -0.1959 6.7039 -0.096 0.0701
Total period

silver 5885 0.0396 0.3118 -0.4522 9.3112 -0.1869 0.1828

platinum 5904 0.0184 0.22 -0.3018 10.3185 -0.1554 0.1393
crude oil 5823 0.0471 0.3767 -0.1042 4.7122 -0.1654 0.1641

Insert Figure 3 about here

289 Figure 4 shows the normalized histograms of silver price returns in both
290 periods. The dashed line corresponds to the normal distribution with the
291 same mean and standard deviation presented in Table 1. The continuous
292 blue line shows a probability density function estimated with a kernel density
293 estimator (KDE). The KDE is very useful to fit any distribution to a set of
294 observations. Only the Gaussian kernel bandwidth parameter needs to be

15

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

295 adjusted. Given that the normal distribution is the appropriate function to
296 describe silver spot returns, both lines should overlay. In Period 2, silver
297 spot returns are also negatively skewed when compared to Period 1. The
298 normal distribution is symmetric and should have a skewness of 0. Finally,
299 the Jarque-Bera (Jarque and Bera, 1980) statistic test was performed on
300 the commodity spot price returns (Table 2). This test statistic follows a
301 Chi-square distribution with two degrees of freedom and is reliable when the
302 sample contains at least 2,000 observations. The null hypothesis is that the
303 observations are normally distributed. This hypothesis cannot be rejected
304 when the test statistics are below the critical value. The results presented in
305 Table 2 shows that the normality hypothesis is rejected at the 99% level of
306 confidence.
Insert Figure 4 about here

Table 2: Jarque-Bera statistics for normality on the 99% level of confidence for the four
different commodities spot price returns.
Commodity spot price Test statistic Critical value (99%
returns conf. level)

Gold 8,440.23 9.21

Silver 2,697.13 9.21
Platinum 10,353.91 9.21
Crude oil futures 1,573.46 9.21

307 4.2. Model Calibration

308 The calibrated parameters and the estimation of their error for the GBM,
309 the SV and the MJD models are presented in Table 3. To reduce the likeli-

16

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

310 hood of getting stuck in a local minimum, several starting parameters have
311 been tested using particle swarm optimization. After 15 iterations, parti-
312 cle swarm optimization is stopped and the sub-optimal best solution in the
313 population becomes the starting values for a gradient-based optimization ap-
314 proach. Also, to ensure the calibrated parameters are realistic, a change of
315 variable was performed in the minimization routine. For example, volatil-
316 ity parameters need to be positive and correlation needs to be between -1
317 and 1. For volatility parameters, the logarithm of volatility is parsed in the
318 minimization routine and converted back taking its exponential before eval-
319 uating the objective function. This ensures that the search space will always
320 contain positive values for volatility. A similar approach is used with the
321 correlation parameter using a tangent and its inverse function. Then, the
322 numerical Jacobian method is used to calculate the error on calibrated pa-
323 rameters (Remillard, 2013). This ensures the Fisher information matrix is
324 always estimated with a positive-definite Hessian matrix. Finally, the Feller
325 condition (2κθ > η) is applied on the SV model to ensure the process never
326 reaches 0 (Albrecher et al., 2006). The MJD converged to the solution in 33
327 steps and the optimization algorithm stopped because the objective function
328 was below the step size tolerance. The results presented in Table 3.

17

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

Table 3: Calibrated parameters of the GBM, Heston SV and MJD on gold, silver platinum
and crude oil between 15 February 1993 and 5 December 2016.

Gold Silver Platinum Crude oil

Param est. std est. std est. std est. std

µ 0.0540 0.0079 0.0729 0.0065 0.0475 0.0046 0.0790 0.0079

GBM

σ 0.1656 0.0020 0.3138 0.0012 0.2223 0.0008 0.3831 0.0020

µ 0.1412 0.0332 0.0563 0.0574 0.0783 0.0361 0.0509 0.0632

Heston SV

κ 1.5008 0.1609 1.7396 0.0245 1.467 0.0787 1.435 0.1584

θ 0.031 0.0120 0.0947 0.0100 0.0506 0.0141 0.1452 0.0224
η 0.3184 0.0529 0.3248 0.0707 0.4665 0.0424 0.3729 0.0361
ρ 0.2075 0.0350 0.2791 0.0520 0.0431 0.0200 -0.4232 0.0346

µ 0.0603 0.0659 0.0957 0.0629 0.059 0.0445 0.1151 0.0789

σ 0.0764 0.0025 0.2052 0.0040 0.0028 0.0076 0.2793 0.0055
MJD

λ 155.87 11.51 56.45 5.6626 59.04 5.6782 45.33 5.97

γ -0.0002 0.0002 -0.0018 0.0011 -0.0006 0.0007 -0.0026 0.0016
δ 0.0114 0.0003 0.0305 0.0012 0.0215 0.0008 0.0386 0.0019

18

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

329 The GBM is the simplest model having only two parameters to adjust.
330 The drift µ can be interpreted as the annual continuously compound return
331 when investing in the commodity. All commodities exhibit positive drift
332 parameters. Crude oil futures exhibit the highest level of volatility while
333 gold prices are less volatile. The volatility of the MJD model is lower than in
334 the GBM but the real volatility of the process is the sum of the GBM and the
335 Poisson process. The jump mean η is negative for the four commodities, but
336 very close to zero. This means that the jump processes are centred around
337 zero and do not significantly affect the skewness of the returns distribution.
338 The λ affects the kurtosis of the distribution by increasing the occurrence of
339 returns in the tails of the distribution.
340 In the Heston SV model, the η parameter determines the volatility of the
341 stochastic volatility process while the long-term trend θ represents its mean
342 reverting level. When the volatility level is above the long-term mean, it is
343 expected to drift to lower values with a speed of mean reversion κ. The initial
344 volatility plays a crucial role on the outcome of simulated results. When this
345 value is far from the mean reversion level, simulated paths will exhibit a
346 transition from the initial volatility to the mean-reverting level. This value
347 was estimated using a rolling standard deviation window on the whole sample
348 and taking the standard deviation corresponding to the average of the first
349 10% of observations. The correlation determines the relationship between
350 price and volatility levels. When the correlation coefficient is positive, a
351 higher level of volatility will cause price returns to move in the same direction.
352 Correlation is positive for gold and silver, neutral for platinum and strongly
353 negative for crude oil futures.

19

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

354 To ensure the calibrated parameters fit the past observations and are also
355 useful for out-of-sample predictions, the root mean square errors have been
356 calculated for the GBM and Heston SV models. The values are presented
357 in Table 4. The dataset was split into two different partitions. The first
358 partition is used for the calibration of the KF parameters and consists of
359 80% of the observations. The second partition is out-of-sample and consists
360 of the remaining 20% observations. Out-of-sample values are systematically
361 lower than in-sample values. This is due to the nature of the data, which
362 contains many more outliers and volatility clusters in the in-sample period.
363 The root-mean-square error (RMSE) is almost identical for the GBM and
364 Heston SV models. The Unscented Kalman filter performs a linear correction
365 on a non-linear problem. In this case study, it does not improve the RMSE
366 when compared to the GBM.

Table 4: In-sample and out-of-sample RMSE for the different commodity datasets. The
KF is calibrated using 80% of the first observations. The out-of-sample RMSE uses the
calibrated parameters on the remaining 20% of the dataset.
GBM Heston SV
in-sample out-of-sample in-sample out-of-sample

Gold 0.0104 0.0105 0.0104 0.0105

Silver 0.0203 0.0169 0.0204 0.0169
Platinum 0.0144 0.0124 0.0144 0.0124
Crude oil 0.0246 0.0223 0.0245 0.0224

20

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

367 4.3. CFaR analysis

368 This paper assesses the CFaR of an open-pit gold mine. The weekly
369 production of the mine is presented in Table 5. To obtain the CFaR value,
370 it is necessary to first calculate the CFO for each trading period. Since the
371 objective of this paper is to assess the effect of the stochastic process on
372 CFaR calculations, the only random variable is the gold price.

Table 5: Characteristics of the open-pit gold mine. These parameters are used for the
CFaR calculations.
Parameter Value

Mine type: Open pit

Processing method: Grinding and flotation
Processing capacity: 180,000 tonne/week
Throughput: 140,000 tonne/week
Waste production: 180,000 tonne/week
Average grade: 1.2 g/tonne
Mining cost: 3.2 $/tonne
Processing cost: 12.5 $/tonne
Gold price: Simulated $/gr
Gold recovery: 84%
Fixed costs: 1,500,00 $/week

373 Using the parameters of Table 5 and Equation 11, the CFO with a gold
374 price of 42$/g is:

21

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

CF O =140, 000 ∗ 42 ∗ 2.6 ∗ 0.84−

[3.2 ∗ (140, 000 + 180, 000)]− (12)

[12.5 ∗ 140, 000] − 1, 500, 000 = $8, 567, 920

375 The daily simulations of gold prices are performed using the GBM, the
376 MJD and the Heston SV models calibrated with historical data. The CFaRs
377 are calculated for a daily, a weekly, a monthly and a quarterly interval. First,
378 5,000 simulations of gold prices are performed using each model, and then
379 the prices are used as an input in Equation 11. The calculated CFaR for each
380 period are presented in Table 6. The results show the simulated cash-flows
381 corresponding to the 5th percentile of the simulated cash-flow distribution.
382 The first thing to notice is that the GBM model systematically undervalues
383 CFaR for each time period. Also, the longer the studied timeframe is, the
384 bigger the difference between CFaR calculated with each different methodol-
385 ogy will be. For the quarterly CFaR calculation, the difference between the
386 MJD and the GBM can vary between 6% and 50.5% when calculated on a
387 quarterly basis while the differences range between 1.9% and 6.1% on a daily
388 basis. Although CFaR values are all positive, they may not meet the hurdle
389 rate required to satisfy shareholders of the company operating the project.
390 If these cash flows were negative, the company might even need to dilute its
391 equity with a secondary offering at an unfavorable market price of shares to
392 prevent a default payment.

22

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

Table 6: CFaR of the gold-mining project for quarterly, monthly, weekly and daily time
intervals
Quarterly CFaR Monthly CFaR Weekly CFaR Daily CFaR

GBM $26,465,238 $10,247,064 $2,189,013 $544,234

MJD $17,577,417 $ 8,671,698 $2,070,152 $512,755
Heston SV $24,739,856 $9,819,090 $2,152,277 $535,176

393 In the last example, the production of the mine was sold at spot price, as
394 soon as it was available. In general, there is a delay between the production
395 of the material and the time when it is sold into markets. This leaves the
396 mining company with an exposure to market fluctuations. In the next case
397 study, two scenarios are studied. In the first scenario, the production is
398 sold at the end of the quarter and is determined with the final price. In the
399 second scenario, the company has the ability to buy put contracts to limit the
400 downside risk of the position. The premium for the options is determined
401 using the Black-Scholes model and the volatility used is the one from the
402 calibrated GBM process. The options are held until expiration and their
403 value at expiration is calculated as the strike price minus the spot price.
404 The option premium is also deducted from the price. The result for the
405 simulations are presented in Table 7.

23

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

Table 7: Quarterly CFaR, mean CFO and standard deviation for 5000 simulations of gold
prices. The stochastic processes for the simulations are the GBM, the SV and the MJD.

5% CFaR Mean CFO Std. dev.

Unhedged 20,496,953 28,947,882 5,482,412

GBM

Hedged 20,912,819 25,892,825 3,146,817

Unhedged 12,751,253 29,283,365 10,928,560

MJD

Hedged 17,451,788 26,180,726 5,871,156

Unhedged 19,206,946 29,141,730 7,004,871

SV

Hedged 19,955,108 25,926,346 4,004,212

406 There is a tradeoff between the mean quarterly CFO and the CFaR.
407 Implementing a hedging programme using options requires paying a premium
408 and the option will often expire worthless. The expected quarterly CFO
409 using the three different stochastic processes are roughly the same, but the
410 MJD and SV unhedged models are more volatile. In the case of the GBM,
411 the hedging programme only increases the CFaR by about 0.5 M$ while
412 decreasing the expected quarterly CFO by 3 M$. So, by using the GBM it
413 may not be clear that the hedging programme will be beneficial. On the other
414 hand, implementing the same hedging programme using the MJD increases
415 the CFaR by 5 M$ and decreases the mean quarterly CFO by 3 M$. The
416 standard deviation decreases from 10.9 M$ to 5.9 M$.

417 5. Discussion

418 Most mining companies use derivative instruments on commodities or

419 exchange rates to reduce their exposure to market risk (Armstrong et al.,

24

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

420 2009). To correctly assess the exposure to market risk, stochastic models
421 are used to describe the price dynamics of commodities. Model risk emerges
422 when the model used to describe commodity price dynamics is misused. This
423 can be due to a violation of the assumptions of the model or a bad calibration.
424 The GBM assumes that commodity price returns are normally distributed.
425 It is clear that commodity price returns are leptokurtic and the model used
426 to simulate price paths calibrated on historical observations should take into
427 account stochastic volatility or jump components. To reduce model risk,
428 different stochastic processes have been calibrated on historical data. Using
429 the KF workflow, the models are calibrated in-sample and out-of-sample. It
430 is important to test the model on an out-of-sample set to be sure the model
431 is not overfitting past observations. In the KF framework, or using ML, the
432 model parameters are calibrated on a proportion of the sample and tested
433 out-of-sample on the rest of the observations.
434 The MJD and Heston SV models are more complex and are harder to
435 calibrate than the GBM. In a robust valuation workflow, different kinds of
436 models should be tested and benchmarked against the GBM. Considering
437 multiple scenarios and using different models can help the risk management
438 team of mining companies to better assess the exposure to market risk. Each
439 model has its own limitations and it is important, from a management point
440 of view, to use multiple models. For example, the GBM model yields reliable
441 results when markets are stable while the MJD and Heston SV models are
442 better suited when there is volatility. By tracking the potential loss when
443 markets become turbulent to the stable market, it is possible to better design
444 hedging programmes. Moreover, it is far less expensive to implement a hedge

25

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

445 on commodity prices when volatility is low, because speculators who take the
446 other position of the hedge will demand a lower premium.
447 Different strategies may be used to calibrate models, depending on their
448 complexity. The MJD model is a GBM with an independent jump process.
449 The PDF of such model is straightforward and can be calibrated using ML.
450 Another good calibration strategy could have been to model a GBM with
451 outlier correction and then model the outliers in the jump component of the
452 model. In the SV model, the diffusion process of both price and volatility are
453 correlated. The model has five parameters to calibrate. The topology of the
454 cost function is highly non-linear and the use of a metaheuristics approach
455 for the optimization helps to find a first good approximation of the solution
456 (Donkor and Duffey, 2013). Then, the sub-optimal solution of the particle
457 swarm optimization is used as an input in the gradient-based optimization,
458 the initial value is very close to the global minimum. The Jacobian and
459 Hessian matrices may not exist in the entire domain of the cost function but
460 it can be approximated in the neighbourhood of the sub-optimal solution. As
461 a result, the gradient-based optimization requires fewer iterations to find the
462 optimal solution. The main advantage of the gradient-based optimization
463 is that it yields an estimate of the parameters error, which can be used to
464 construct confidence intervals or assess if the parameters are reliable.
465 Because of the diversity of stochastic processes that can be used to model
466 commodity price dynamics, there is an inherent lack of generality in the pa-
467 rameter calibration process. To solve this problem, Schöne (2014) developed
468 a general calibration method based on the minimization of a theoretical and
469 a kernel based PDF. In this paper, a different approach is used, and the

26

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

470 criteria to assess how well each model compared to each other are in-sample
471 and out-of-sample root mean-square error measures. This procedure ensures
472 that no matter how the stochastic process is calibrated, it is effective at
473 fitting past observations and is also useful for making predictions. Histor-
474 ical calibration converges onto the estimation of parameters that may not
475 be forward-looking. If the calibration is performed in the recent past, the
476 calibrated parameters will reflect market dynamics of the past period. How-
477 ever, it is very unlikely that the past will repeat itself indefinitely and the
478 estimates should be readjusted to reflect the expectation of future events.
479 This is particularly important when valuing market risk exposure.
480 Considering the CFaR is important when there is a delay between the
481 production of the material and selling it to the market. If no derivatives
482 instruments are used, the mining company is exposed to drops in the com-
483 modity price. Using the CFaR calculations, the company may use derivatives
484 instruments to hedge the downside risk of gold prices using options on fu-
485 tures. For example, if the minimum cash-flow a company needs on a quarterly
486 basis to meet the hurdle rate is 20 million dollars, the CFaR calculated with
487 the GBM will exceed the CFaR threshold but the one calculated with the
488 MJD will breach it. The company decides to hedge against falling prices by
489 buying a call options on gold futures with an expiration at the end of the
490 quarter. This gives the option to enter into a futures contract where gold
491 will be delivered at the strike price given in the contract. Since the hedge is
492 implemented when the option strike price is far out of the money, it will be
493 cheaper to implement than if the options price was at the money.
494 The CFaR was calculated considering only one stochastic parameter. In

27

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

495 general, mining projects will be affected by more than one risk factor. For
496 example, CFaR could be calculated with a stochastic exchange rate level, or
497 with varying production parameters such as the grade of the ore. The CFaR
498 of a project incorporating all of these uncertainties would be lower than the
499 individual CFaR. For example, lower commodity prices could be countered
500 by a more advantageous exchange rate level. Another important task would
501 be to verify that the calculated CFaR actually reflects past observations.
502 If the CFaR is calculated using the historical method, the CFaR will be
503 exactly the one corresponding to historical observations. On another hand,
504 the CFaR calculated using the workflow presented in this paper relies on
505 MC simulations and should be back tested. One drawback of the CFaR
506 methodology is that it yields an upper limit of the minimum cash-flow that
507 a project will produce at a given time. In reality, the actual minimum cash-
508 flow could be much lower. Another approach to better quantify the actual
509 loss is the Conditional Value at Risk (Acerbi and Tasche, 2002). The risk
510 metric produced with this approach can be interpreted as the expected loss
511 assuming that the Value at Risk threshold has been triggered. This can be
512 extended to CFaR calculations.

513 6. Conclusion

514 Stochastic processes are widely used to analyze risks associated with com-
515 modity prices through Monte-Carlo simulations. The important question is
516 how to find or calibrate appropriate values for the parameters of stochastic
517 processes. This paper proposed a new hybrid approach to calibrate the pa-
518 rameters to be used in Heston SV and MJD models. For calibration, KF were

28

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

519 used. Since KF requires an initial sub-optimal solution, particle swarm op-
520 timization was used to generate an initial solution. This paper investigates
521 how the choice of the stochastic process can affect the CFaR of a mining
522 project. The risk analysis is implemented using real world probabilities. De-
523 pending on the complexity of the model, different calibration strategies may
524 be used to fit historical observations. The KF calibration workflow proves
525 to be a very robust. This paper uses the Unscented Kalman filter to model
526 the SV in a filtering framework. The model parameters are calibrated using
527 a hybrid metaheuristics approach using particle swarm optimization and a
528 gradient based optimization method. This approach linearizes a non-linear
529 problem using a Gaussian approximation. As future work, another algorithm
530 called the particle filter could be tested. However, since the particle filter re-
531 lies on individual particles that are each transformed using the state-space
532 transition equation, it is much more computationally intensive. The choice
533 of the stochastic process affects the calculated CFaR risk measure. This is
534 because more complex models such as the MJD and the Heston SV models
535 are better suited for reproducing fat tails observed in empirical observations.

536 Appendix

537 The Unscented Kalman filter algorithm is taken from Javaheri et al.
538 (2003). First, the mean and covariance are initiated:

x̂0 = E[x0 ]
(13)
P0 = E[(x0 − x̂0 )(x0 − x̂0 )T ]
539 The state vector is concatenated with the system noise and the observa-
540 tion noise:

29

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

 
x
 k−1 
xak−1 =  wk−1 (14)
 

 
uk−1
541 Therefore:
 
x̂k−1
 
x̂ak−1 a
= E[xk−1 |zk ] =  (15)
 
0 
 
0
542 and

 
Pk−1 Pxw (k − 1|k − 1) 0
 
a
Pk−1 =  Pxw (k − 1|k − 1) Pww (k − 1|k − 1) (16)
 
0 
 
0 0 Puu (k − 1|k − 1)

For i = 1...na , the sigma points are calculated:

q 
χak−1 (i) = x̂ak−1 + (na + a
λ)Pk−1 (17)
i

And for i = na + 1...2na :

q 
χak−1 (i) = x̂ak−1 − a
(na + λ)Pk−1 (18)
i−na

χk|k−1 (i) = f (χχk−1 (i), χW

k−1 (i)) (19)

543 for i = 0...2na + 1 so:

2na
X (m)
x̂−
k = Wi χk|k−1 (i) (20)
i=0

30

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

544 and

2na
X (c)
Pk− = Wi (χk|k−1 (i) − x̂− − T
k )(χk|k−1 (i) − x̂k ) (21)
i=0

545 The innovation is:

zk|k−1 (i) = h(χk−1 (i), χuk−1 (i)) (22)

546 then:

2na
X (m)
ẑk− = Wi zk|k−1 (i) (23)
i=0

and:
vk = zk − ẑk− (24)

547 For the measurement update:

2na
X (c)
Pzk zk = Wi (zk|k−1 (i) − ẑk− )(zk|k−1 (i) − ẑk− )T (25)
i=0

and
2na
X (c)
P xk z k = Wi (χk|k−1 (i) − x̂− − T
k )(χk|k−1 (i) − x̂k ) (26)
i=0

548 The Kalman gain is:

Kk = Pxk zk Pz−1
k zk
(27)

The next estimate is:

x̂k = x̂−
k + Kk vk (28)

549 and the covariance is:

31

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

Pk = Pk− − Kk Pzk zk KkT (29)

32

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

550 7. References

551 Acerbi, C. and Tasche, D. (2002). Expected Shortfall: A Natural Coherent

552 Alternative to Value at Risk. Economic Notes, 31(2):379–388.

553 Albrecher, H., Mayer, P., Schoutens, W., and Tistaert, J. (2006). The Little
554 Heston Trap.

555 Alesii, G. (2003). Controlling CFaR with real options. A univariate case
556 study. Technical report, Working Paper.

557 Armstrong, M., Galli, A., Lautier, D., et al. (2009). A reality check of
558 hedging practices in the mining industry. Technical report, Paris Dauphine
559 University.

560 Block, S. (2007). Are Real Options Actually Used in the Real World? The
561 Engineering Economist, 52(3):255–267.

562 Donkor, E. A. and Duffey, M. (2013). Optimal Capital Structure and Fi-
563 nancial Risk of Project Finance Investments: A Simulation Optimization
564 Model With Chance Constraints. The Engineering Economist, 58(1):19–
565 34.

566 Eberhart, R. C. and Kennedy, J. (1995). A new optimizer using particle

567 swarm theory. In Proceedings of the sixth international symposium on
568 micro machine and human science, volume 1, pages 39–43. New York,
569 NY.

570 Gatheral, J. (2011). The volatility surface: a practitioner’s guide, volume

571 357. John Wiley & Sons.

33

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

572 Gorton, G. and Rouwenhorst, K. G. (2006). Facts and fantasies about com-
573 modity futures. Financial Analysts Journal.

574 Hammond, R. K. and Bickel, J. E. (2013). On the Decision Relevance of

575 Stochastic Oil Price Models: A Case Study. The Engineering Economist,
576 58(3):209–230.

577 Heston, S. L. (1993). A closed-form solution for options with stochastic

578 volatility with applications to bond and currency options. Review of Fi-
579 nancial Studies, 6(2):327–343.

580 Jarque, C. M. and Bera, A. K. (1980). Efficient tests for normality, ho-
581 moscedasticity and serial independence of regression residuals. Economics
582 letters, 6(3):255–259.

583 Javaheri, A., Lautier, D., and Galli, A. (2003). Filtering in finance. Wilmott,
584 3:57–83.

585 Jorion, P. (2001). Value at risk: the new benchmark for managing financial
586 risk. NY: McGraw-Hill Professional.

587 Merton, R. C. (1976). Option pricing when underlying stock returns are
588 discontinuous. Journal of Financial Economics, 3(1):125–144.

589 Remillard, B. (2013). Statistical Methods for Financial Engineering. CRC

590 Press.

591 Sauvageau, M. and Kumral, M. (2016). Genetic algorithms for the optimi-
592 sation of the schwartz–smith two-factor model: a case study on a copper

34

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

593 deposit. International Journal of Mining, Reclamation and Environment,

594 0(0):1–19.

595 Schöne, M. (2014). Real Options Valuation: The Importance of Stochastic

596 Process Choice in Commodity Price Modelling.

597 Schwartz, E. and Smith, J. E. (2000). Short-term variations and long-term

598 dynamics in commodity prices. Management Science, 46(7):893–911.

599 Schwartz, E. S. (1997). The stochastic behavior of commodity prices: Impli-

600 cations for valuation and hedging. The Journal of Finance.

601 Subulan, K., Baykasoğlu, A., Akyol, D. E., and Yildiz, G. (2017).
602 Metaheuristic-based simulation optimization approach to network revenue
603 management with an improved self-adjusting bid price function. The En-
604 gineering Economist, 0(0):1–30.

605 Topal, E., Haque, M. A., and Lilford, E. (2016). Evaluation of a mining
606 project under the joint effect of commodity price and exchange rate uncer-
607 tainties using real options valuation. The Engineering Economist, 0(0):1–
608 22.

35

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

609 Graphics captions

Figure 1: Price levels for gold, silver, platinum (in $USD/ounce) and crude oil futures (in
$USD/barrel) during 8 February 1993 and 8 November 2016.

Figure 2: Logarithm of returns of gold spot prices during 8 February 1993 and 8 November
2016. The first period ranges between 8 February 1993 and 27 December 2004. The second
period ranges between 29 December 2004 and 8 November 2016.

Figure 3: Boxplots for gold, silver, platinum and crude oil futures price returns. The
central part of the plots covers the median, the 25th and the 75th percentiles. The whiskers
extends to the 10th and the 90th percentiles and points are considered as outliers when
the distribution is assumed to be normal.

36

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

Figure 4: Histogram of spot gold price returns. The dashed PDF corresponds to a normal
distribution with the same mean and variance as observed in the empirical dataset. The
continuous PDF is calculated with a kernel density estimator.

37

ACCEPTED MANUSCRIPT