See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/319935784

A forward dynamic optimization strategy under contango storage arbitrage with

frictions

Article  in  Journal of Energy Markets · September 2017

DOI: 10.21314/JEM.2017.166

CITATIONS READS
2 943

2 authors:

Behzad Ghafouri Matt Davison

The University of Western Ontario The University of Western Ontario
2 PUBLICATIONS   3 CITATIONS    112 PUBLICATIONS   1,791 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A Forward Dynamic Optimization Strategy for Contango Storage Arbitrage View project

Mathematical analysis and numerical solution of fractional reaction-diffusion problems with: View project

All content following this page was uploaded by Behzad Ghafouri on 21 September 2017.

The user has requested enhancement of the downloaded file.

 A forward dynamic optimization strategy under
contango storage arbitrage with friction

Behzad Ghafouri, Matt Davison

Abstract
The goal is to explain and improve the off-shore oil storage trade observed in a contango market using a
forward dynamic optimization strategy. The strategy formulation is developed in terms of trades in forward
contracts and contrasted with the literature. By simulating forward prices based on realistic May 2009
market conditions, the NPV of the strategy compared to selling the oil on the spot in a base case scenario
is $8.99/barrel, which is comprised of $6.19 generated by the initial forward maximization, and $2.80
achieved by the subsequent trades.
The impact of the forward curve dynamics is studied by examining the trading decisions based on the
realized slope and mean level of forward curve. The path of realized slope and stepwise changes in the
slope is found to be able to explain most of the trading decisions. The effect of initial conditions, frequency
of readjusting the position, and storage cost are also examined. The optimal frequency depends on the
storage cost. Friction is introduced in the problem by penalizing the refund of the storage cost when a
decision to advance the short position in time is made, where its influence, which depends on the storage
cost, is at most $1.40.

Keywords
Oil storage, Contango arbitrage, Forward curve model, Tanker Rentals, Real options, Dynamic optimization

Publication DOI
10.21314/JEM.2017.166

1
1. Introduction
The shape of the futures curve contains important information about the relation between spot and forward
prices and the corresponding trading opportunities in the financial markets. The futures market is connected
with the market for the underlying physical commodity via storage. Oil inventories reflect fundamental
supply-demand factors, which can be exploited in discovering arbitrage opportunities (Ye and Karali 2016).
The decision to sell the oil at the spot price, or to store it and use futures contract to deliver at a later point
in time is a simple yet important question for the participants in the supply chain of oil and oil products
including producers and speculators. When the crude oil futures curve is upward sloping, a condition known
as contango, they may present the profitable trading opportunity of buying oil on the spot at a low price,
storing it, and taking a short position in a longer term futures contract at a high price. If the storage,
pumping, and other related costs are outweighed by the difference between the spot and forward prices, this
strategy can be profitable.
This type of activity was prevalent during the super contango of late 2008 to early 2009 when oil prices hit
a low point. For instance, on Feb 12, 2009, there was a very steep 12-month contango between March-2009
and March-2010 futures contracts, trading respectively at $33.98/barrel and $55.95/barrel. Let F(t,T)
denotes the futures price at time “t” for delivery at time “T”. Figure 1 illustrates historical futures one-year
spread based on WTI crude oil contracts, where the near-end of the spread is the front-month contract. The
figure shows a positive spread in the given time frame. The forward curve slope,
( ( , ) − ( , ))⁄( − ), measures change in value between two delivery points in time. Like
storage cost, it has units of dollar per year per barrel. If the slope is greater than the storage cost, there might
be an opportunity in going long the near-end (buy spot) and shorting the far-end (sell forward). When this
type of trade is profitable, most of the cheaper on-shore storage capacities are exhausted, and traders resort
to “floating storage” using off-shore oil tankers. Very Large Crude Carriers (VLCC), with capacity of
around 2,000,000 barrels, are popular for this purpose. Figure 1 shows the historical VLCC one-year Time-
Charter rates (BIMCO 2016; Charles R. Weber Company, Inc. n.d.), which are annualized on a per barrel
basis. The sharp drop in the VLCC charter prices in the late-2008 to 2009 period, combined with large
futures spreads, triggered the wide use of off-shore storage trades in this period. The slope has stayed
positive since the beginning of 2015, which reflects the new era of abundant crude oil production. Tanker
rates declined from $9.31 to $5.47 in the first half of 2016, another positive signal for the storage trade. In
this paper, we extend the static storage trade into a dynamic trading strategy by allowing subsequent trades.
We show an increase in the added value of the dynamic compared to the static trade, and investigate the
impact of underlying factors on the added value.

2. Existing Literature
The commodity storage literature presents strong evidence of the relation between forward curve shape and
storage (Considine and Larson 2001; Kaldor 1939; Litzenberger and Rabinowitz 1995; Working 1948). By
the no-arbitrage principle, forward price is (0, ) = [( + − ) ], where is the spot price,
is the interest rate, is the storage cost, and is known as the convenience yield. In the case of crude oil,
Geman and Ohana (2009) verified that the slope of the forward curve can be a good proxy for inventory
levels. They documented the relationship among interest-adjusted spread, defined as [ (0, ) −
( )]/ , and inventory levels. When inventory levels are high, convenience yield ( ) is low because
any change in demand can be absorbed through the inventories. As a result, + − > 0, and forward
curve is upward sloping. Under these conditions, they showed that interest-adjusted spread is high, which

2
 Figure 1. Historical One-year Crude Oil Futures Spread and VLCC Time-Charter Rates

The spread data is based on CME NYMEX WTI contracts, while the VLCC data is from The Baltic and International
Maritime Council (BIMCO 2016) for Time-Charter Rates 1, and Charles Weber Company (Charles R. Weber
Company, Inc. n.d.) for Time-Charter Rates 2 (per barrel per year).

inspires a cash-and-carry strategy, i.e. long the spot and short the forward. The cointegration between spot
and forward prices has been studied with the conclusion that the efficiency of crude oil market, as
characterized by the no-arbitrage rule, is influenced by the existence of structural breaks (Chen et al. 2014;
Chinn et al. 2005; Maslyuk and Smyth 2009). For instance, Chen et al. (2014) found that there is evidence
against market efficiency for subsamples 1986-2012 and 1986-2004 within WTI data. So, the question turns
into how to exploit any potential arbitrage opportunities.
To profit from an upward sloping forward curve, Jafarizadeh and Bratvold (2013) explored a static trade
where the trader buys oil and simultaneously sells it using the profit-maximizing forward contract. They
also presented the dynamic version of this trade, Forward Dynamic Optimization (FDO), where the trader
tries to profit from favorable shifts in the forward curve by readjusting his (net) short position at subsequent
points in time. However, there are some issues with their approach, as discussed later in this paper. Other
studies concerned with FDO are related to natural gas storage valuation, where it is also known as Rolling
Intrinsic valuation (Gray and Khandelwal 2004; Maragos 2002; Parsons 2013). Intrinsic value is the value
that can be locked in by hedging a forward position using storage once. On the other hand, extrinsic value
is derived from the flexibility to readjust the position in response to forward curve realizations in the future.
Using FDO, all of the intrinsic value and part of the extrinsic value is captured. In capturing the full extrinsic
value, there is a risk versus reward trade-off, where a higher value comes with the risk of larger variations.
In comparison to other valuation method, FDO is very intuitive and is favored by practitioners (Secomandi
et al. 2015; Ware 2013). This is a result of transparency, and consistency with the methods used by
companies to monetize the flexibility of their storage assets, because it is explicitly based on trading of
futures contracts. Because the subsequent adjustments are risk-free, as trades happen only if profitable, the
owner is exposed to a very limited downside risk. In less liquid markets, this method is very attractive
because one can consider transactions costs in deciding whether to enter a trade once a seemingly profitable
opportunity is identified. On the other hand, the full real option value of storage asset flexibility is less
transparent and more dependent upon the underlying models and assumptions, which makes them harder
to communicate with the management.

3
Prices can be simulated by different varieties of processes, or may be based on historical ones, as utilized
for back testing purposes. However, in the present research, simulated prices are used due to the versatility
they provide in examining different aspect of the problem. It has been shown that single factor models are
not able to completely capture the oil price dynamics, particularly either the changes in the futures curve
(slope and curvature), or the difference in volatility among different maturities, and therefore multiple factor
models have been introduced (Hilliard and Reis 1998; Schwartz 1997). Oil prices may exhibit a stationary
or non-stationary behavior depending on the time period, which supports the use of a two-factor model,
where one factor is mean-reverting (stationary) and the other factor is non-stationary (Hahn et al. 2014) .
Bhattacharya (1978); Dixit and Pindyck (1994), and Dias (2005) reviewed different mean-reverting
processes within the context of modeling oil prices. One main class of the two factor models involves a
mean-reverting process, modelling convenience yield, nested within the drift term of a geometric Brownian
motion (GBM) for the price (Gibson and Schwartz 1990; Ribeiro et al. 2004; Schwartz 1997). In the second
class of the two factor models, the long-term component is modeled with a GBM process, and is combined
with the short-term deviations that are modeled with a mean-reverting process (Pindyck 1999; Schwartz
and Smith 2000). The second approach is computationally easier and provides a more direct way to separate
the two factors. Other more complicated models have more than two factors or jump processes (Casassus
and Collin-Dufresne 2005; Schwartz 1997), however, two factor models are the most common ones from
a capability versus practicality point of view.
In the present research, Schwartz and Smith (2000) model is used, where the first state variable represents
the short-term mean-reverting factor ( ), and the second state variable ( ) describes the long-term
equilibrium price level, by which the spot price is defined in Eq. 1.
ln(S ) = χ + ξ Eq. 1

Under the risk-neutral probability measure, the two processes have the following representation.
= (− − ) + ( ) Eq. 2

=( − ) + ( ) Eq. 3

= Eq. 4

Computing ( , ) is done under the risk-neutral measure. However, the goal here is to evaluate
performance of the proposed trading strategy and the extent of its profit or loss, which falls in the risk
management realm, and thus requires simulating state variables under the physical (historical) measure.
Luckily, the futures curve ( , ) is fully explained by a closed-form solution in terms of the only two state
variables ( , ). When the price has a closed form solution, we can simulate the underlying risk drivers,
( , ), under the physical measure, and feed them into ( , ) to compute future exposures (Schoftner
2008). In contrast to the present research, simulation under the physical measure has not been considered
in previous studies aiming at evaluation of FDO (Jafarizadeh and Bratvold 2013), or storage asset valuation
(Lai et al. 2010).

3. The Model
3.1. Assumptions and Formulation
Assume a trader owns a barrel of crude oil stored in a tanker, and must deliver it by time T. Refilling the
inventory is not permitted. The net exposure of all contracts must be “short one unit” of oil to hedge the
“long one unit” initial position. The time horizon is discretized into N periods = 0, , , … , = .
Pumping Cost denotes all costs due at delivery, such as pumping cost and any location discount to WTI
futures. Holding Cost represents all of the costs associated with operating the tanker. If the trader delivers

4
the oil before the end of the tanker rental agreement, he will be refunded the unused portion of the rent. We
later relax this refund assumption.
At = 0, the trader faces the following alternatives. One is to sell the oil on the spot for payoff − .
Given the contango condition, this will not be an optimal choice. Alternatively, the trader chooses an
optimal forward contract by “forward maximization”; given the vector of forward prices, ( ) =
[ ( , ), ( , ), ( , ), … , ( , )], assuming that the cost of storage has been paid up to ,
( ( ), , , ) computes the NPV of the profit from selling the oil by selecting the most profitable
contract.

( ( ), , , ) = max { , − − − , 0} Eq. 5

At times , , … , = , the trader must perform a “Dynamic Optimization” to readjust his position given
the most recently realized prices. This readjustment captures any favorable shift in the futures curve, which
can only add positive value. According to Jafarizadeh and Bratvold (2013), following Eydeland and
Wolyniec (2003), the FDO algorithm is formulated as
= ( ( ), , , )+ Eq. 6

max{ ( ( ), , , )− ( ( ), , , ), 0}

As well as trivial concerns with time discounting, the term ( ( ), , , )−

( ( ), , , ) does not correspond to an explicit trade in futures contracts to capture
favorable changes. Assuming all terms in the max operator are positive, Eq. 6 sums to an unrealistic
( ( ), , , ) answer.
Assume that the trader reaches time , while holding a short position in a contract that matures at time
≤ . Figure 2 highlights the two new contracts being considered assuming that the new optimal
maturity time is ∗ .

Figure 2. New positions to capture any potential gains

Existing Short: + ( , ) New Short: + ( , ∗ )

Offsetting Long: - ( , )

An offsetting long and a new short must be entered into simultaneously if any trade is to happen at .

The trader solves Eq. 7, where ( ) represents the NPV of the potential trades at .

( ) = max ( )( ( , )− )− .
( , )− Eq. 7

− . − ,0

5
Assuming = 0, Eq. 7 can be written as Eq. 8, which indicates the trader will search for any maturity <
(postponing) with slope of the forward curve greater than HC, or similarly for any advancing maturity
< .

( , )− , Eq. 8
− ,0 , >
−
′( ) = −
, − ( , )
− − ,0 , <
−

The total value derived from the original position and any subsequent trades is
Eq. 9
( )
= [ ( )]

3.2. Parameter Estimates and Simulation

Parameters of Schwartz and Smith (2000) model as estimated by Hahn et al (2014) are listed in Table 1.
The parameters used by Jafarizadeh and Bratvold (2013) were not based on actual estimation of the model
and only represented a hypothetical assumption of a favorable condition for this trade. Hahn et al (2014)
estimated the parameters by employing the Kalman filter (Kalman 1960) method using 1990–2013 WTI
futures data. This data included several major developments in crude oil markets, compared to Schwartz
and Smith (2000) that only included 1990–1996 data.

Table 1. Parameters estimated by Hahn et al (2014)

0.3116
0.2053
1.0880
0.0823
0.3733
∗
= − -0.0252
0.1070
0.0818
These parameters are used in simulating futures prices according to Eq. 1 to Eq. 4.

We generated m price paths by simulating the state variables with Δ = 1⁄10080 year. The number
10,080 = 2 ×3 ×5×7 is chosen to allow perfect divisibility to several other coarser number of periods to
be exacted from the underlying finer price simulation, while keeping the above Δ constant in all
simulations for consistency. The value of trading period is often set to Δ =0.25 (quarterly), which is
equivalent to N=4 periods based on the T=1 year. The case with N=4 time periods, m=10,000 simulations,
storage cost of HC=$6.57/(barrel.year), and initial condition of ( , )=(-0.639,4.637) will be referred to
as the base case, and will serve as a basis in all of the investigations, as listed in Table 2. The chosen initial
condition and storage cost represents the corresponding values on May-3-2009, as an example of a
favorable time for this type of trade. Also, ( , )=(-0.639,4.637) can be translated into =$54.45/barrel
and a long-term equilibrium price of $103.19/barrel based on the price model. A tanker rent of
HC=$6.57/(barrel.year) is based on the actual time charter prices of a 2 million barrel VLCC around May
2009 in Figure 1. Also, we consider a Pumping Cost of PC=$3.75/barrel. If focus on the impact of a

6
particular parameter, where an alternative value relative to the base case is employed, the specific range of
the parameter will be provided accordingly.

Table 2. Problem parameters as specified by the base case.

Description Parameter Base Case Value
Number of periods for dynamic optimization N 4
Storage cost HC $6.57/barrel.year
Pumping cost PC $3.75/barrel
Number of simulated path m 10,000
Initial condition of state variables ( , ) (-0.639,4.637)
Initial Spot Price = exp( + ) $54.45
Time Horizon (constraint) T 1 year
These parameters are used in the FDO algorithm as specified by Eq. 7 and Eq. 9.

4. Results
Different aspects of the results are studied using the following metrics of value; Added Value ($) of a
strategy is defined as the gain relative to selling the oil on the spot price, and Added Value (%) is useful
when comparing the added value under two different spot prices or initial conditions.
($) Eq. 10
Added Value ($) ≔ −( − ), Added Value (%) ≔

In the base case, the 95% confidence interval for Added Value ($) is $8.94-$9.05. For a two-million-barrel
capacity VLCC tanker, this is equivalent to $17.88-$18.10 million. The Added Value ($) of $8.99 can be
broken down into two parts; (i) $6.19 generated by selling the oil forward using F(0,1), and (ii) $2.80
obtained by all of the subsequent trades. Figure 3 shows, while the strategy guarantees $6.19, payoffs are
significantly skewed to the upside. At t=0, part (i) is known (certain), whereas part (ii) is uncertain.
Although most of the value is captured by the first part, the subsequent trades are necessary to capture the
remaining 31% of the Added Value ($). The contribution of the subsequent trades to Added Value ($) will
increase if the initial forward curve is less steep, a less favorable environment to start the trading. So, a
trader may still have an incentive to start although the certain part of the added value is not very high.

4.1. Decision-making analysis at the path level

To focus on the impact of the forward curve slope, payoffs are presented in terms of slopes at the current
and the preceding time steps. Realization of slopes and its consequences on the decisions and payoffs at
each time step are tracked. Figure 4 shows the payoff at t=0.25, where all paths have the same history
because the algorithm starts by shorting F(0,1) at t=0 in the base case. As listed in Figure 4, there are four
selected new contracts, which correspond to four different decision regions labeled A through D, depending
on the slope realized at t=0.25. Region A is where the realized slope is too high, and F(0,1) is held as the
optimal contract because there is no profitable choice. Region B is where the slope is slightly lower, and
F(0.25,0.75) is selected accordingly. As the slope falls further, F(0.25,0.5) is selected in regions C. If the
slope decreases even further to fall within the lowest part of the range, region D, F(0.25,0.25) is chosen as
optimal. As the slope of forward curve decreases more, the maturity date of the new contract is moved from
t=1 (hold existing contract) to t=0.25 (sell on the spot), to generate a larger payoff. Note that the intervals
identifying the regions have a small overlap since the slope of the forward curve is computed as the slope
of the fitted line going through the available contracts. This excellent approximation is not completely
accurate due to the inherent curvature of the forward curve. When the algorithm searches for a potential

7
 Figure 3. Histogram of Added Value ($).

Histogram of strategy payoffs; base case parameters.

contract to short, it considers each spread trade (long one and short another contract) individually, leading
to a slightly different slope from the fitted line. The dashed lines, which illustrate the concept by dividing
the regions, are based on the midpoint of the overlap area.

Figure 4. Payoff and selected contract at t=0.25 as a function of slope at t=0 and t=0.25.

The trader advances the net short position one, two, or three periods, or hold F(0,1) based on the realized slope at
t=0.25, where all simulation parameters are according to the base case as specified in Table 1 and Table 2.

Subsequently, the existing contracts at t=0.5 fall into one the three types; F(0,1), F(0.25,0.5), and
F(0.25,0.75). Figure 5 shows the payoff at t=0.5 as a function of the slope at t=0.25 and t=0.5, where seven
regions are identified, and labeled A through F. For each region, the existing contracts are determined based

8
on the corresponding slope at t=0.25, and are listed in the figure. Conditional on an existing contract, the
realized slope at t=0.5 will determine the new optimal contract. In regions A and B, F(0,1) is the existing
contract. If the slope realized at t=0.5 is very high, F(0,1) will be held (region A). Otherwise, F(0.5,0.75)
or F(0.5,0.5) will be chosen (region B) depending on how much the slope decreases; if the slope decreases
to a large extent, F(0.5,0.5) is preferred to F(0.5,0.75). In regions C, D and E, the existing contract is
F(0.25,0.75). If there is a sufficiently large increase in the slope (region C), the new contract will be F(0.5,1).
If there is a sufficiently large decrease in the slope (region E), the new contract will be F(0.5,0.5).
Alternatively, if the change in the slope is neither large- nor small-enough (region D), the existing contract,
F(0.25,0.75), will be held. In regions F and G, F(0.25,0.5) is the existing contract. If the slope does not
increase enough, the existing contract, F(0.25,0.5), will be held (region G), and the trade will conclude with
delivery of the oil. If the slope increases largely, the net short position will be postponed by choosing
F(0.5,0.75) or F(0.5,1) (region F) depending on how much the slope increases, where; F(0.5,1) is preferred
to F(0.5,0.75) if the slope increases very much.

Figure 5. Payoff and selected contract at t=0.5 as a function of slope at t=0.25 and slope at t=0.5.

Starting with an existing contract based on the slope at t=0.25, the trader acts by choosing a new contract based
on the realized slope at t=0.5, where all simulation parameters are according to the base case as specified in Table
1 and Table 2.

Considering the paths with a zero payoff in region G, the selected optimal maturity at t=0.25 was the next
time step, t=0.5. So, regardless of how much the slope changed by t=0.5, it is not possible to advance the
trade any further. Thus, a small gain at t=0.25 deprived these paths from a larger potential gain at t=0.5.
Similarly for the paths in region A with an existing F(0,1) contract, regardless of how large the increase in
the slope at t=0.5 is, it is not possible to take advantage of the opportunities. To conclude, it might have
been better to make less profit (or incur some loss) at t=0.25 by avoiding the F(0.25,0.5) (or F(0,1)) contract.
Avoiding the time boundaries leaves a two-sided opportunity open to make a potentially large gain at t=0.5.
At t=0.75, existing contracts are comprised of four different types; F(0.25,0.75), F(0.5,0.75), F(0.5,1), and
F(0,1) but may be categorized into just two maturity date classes, t=0.75 and t=1, sufficient for tracking
subsequent actions. Figure 5 shows that if the slope at t=0.5 exceeds 8.25, the selected maturity time is 1,
and it is 0.75 otherwise, which is reflected in Figure 6. So, in regions A and B the existing contracts mature
at t=1. If the slope realized at t=0.75 is greater than a threshold, the existing contract, F(0.5,1) or F(0,1),

9
will be held (region A), however, if it is less than the threshold (6.97), F(0.75,0.75) will be chosen (region
B). In regions C and D, the existing maturity date is t=0.75. If the slope realized at t=0.75 is greater than
6.97, F(0.75,1) will be chosen (region C) to effectively postpone the net short position while a smaller slope
at t=0.75 implies the existing contract, F(0.25,0.75) or F(0.5,0.75), will be held (region D), and the trade
will terminate without a profit.

Figure 6. Payoff and selected contract at t=0.75 as a function of slope at t=0.5 and slope at t=0.75.

Starting with an existing contract based on the slope at t=0.5, the trader acts by choosing a new contract based on
the realized slope at t=0.75, where all simulation parameters are according to the base case as specified in Table 1
and Table 2.

Figure 4, Figure 5, and Figure 6 signify the critical role of the forward curve slope in explaining the rationale
for selecting new contracts; the realized slope is constantly compared against the storage cost to detect any
sufficiently large spread between the two, which triggers a postponing/advancing decision. Accordingly,
we show that there is a threshold of around HC=$6.57/barrel.year, which governs the decisions. In Figure
4, Figure 5, and Figure 6, the slope marked by the dashed line between regions A and B, i.e. hold F(0,1) or
advance respectively, is 9.46, 8.25, and 6.97. Although there is variation among the three values, they get
closer to HC=$6.57, which is due to the slight difference between the slope based on which the trades are
made and the slope of the line fitted to the forward curve. With the passage of time, the length of the forward
curve decrease due to a lower number of available contracts to a point, where at t=0.75, the forward curve
is comprised of only two contracts, and the curve coincide with the fitted line. This is why the values, 9.46,
8.25, and 6.97, get closer to HC=6.57. The remaining discrepancy is due to the fact that the slope does not
consider the time value of pumping cost, or discounting the forward prices, while both of which are reflected
in the decision-making.
In Figure 7, the payoff levels at t=0.25, 0.5, and 0.75 are overlaid on a plot whose axes show the change in
the slope of forward curve and change in the mean forward prices through each period. A decrease
(increase) in the slope is usually accompanied by an increase (decrease) in the mean forward prices such

10
 Figure 7. Payoff levels at (a) t=0.25, (b) t=0.5, and (c) t=0.75 in terms of change in the slope and mean forward
prices.
(a)

(b)

(c)

Payoffs magnitude is divided into five different levels at each time, and for clarity, two plots are provided in parts
(b) and (c), where all simulation parameters are according to the base case as specified in Table 1 and Table 2.

11
that over the three time steps, 62% of the cases falls in the second quadrant; low spot prices are connected
with high slopes, and high spot prices with lower (or even negative) slopes. The magnitude of the change
(dispersion around origin) increases with the passage of time. Simulations initial conditions were
=$54.45/barrel and a long-term equilibrium price of $103.19/barrel (a steep positive slope). As prices
approach their long-term levels the forward curve slope flattens.
Figure 7.a shows at t=0.25 the larger the change in the slope, the higher the payoff, whereas the change in
mean forward prices has no effect on the result. Payoff levels at t=0.5 and 0.75, exhibits a weaker
dependence than t=0.25. At t=0.5 and t=0.75, unlike t=0.25, both positive and negative changes in the slope
may lead to non-zero payoffs. In parts (b) and (c), positive-payoff points are spread around the Y-axis
asymmetrically. The cluster of the points on the right hand side generate value by postponing, and the one
on the left by advancing. The left cluster is more populated than the right one, indicating the higher
likelihood of advancing.

4.2. Initial Conditions

In this section, ( , ) pairs are chosen from −0.7 ≤ ≤ −0.3 and 3.8 ≤ ≤ 5.2, which generates
441 different initial conditions. The results are shown in Figure 8.a, where the ( , ) rectangular region
is mapped into the corresponding “price” region in terms of = exp( + ξ ) and long-term price =
exp(ξ ) transformations. Some of the historical values of ( , ) from the period between August 2008
and October 2011 are extracted from Figure 4 in Hahn et al (2014), which estimated the evolution of prices
in terms of ( , ). These are the 21 black squares shown on the figure, 15 of which lead to a positive
Added Value (%). The figure shows that for instance, at =$60, the long-term price should be greater than
$90 in order to have a steep-enough initial slope and a positive profit.

Figure 8. Added Value (%) as a function of the initial prices, and Certain Added Value ($) vs Uncertain Added Value
($) colored by the initial slope
(a) (b)

(a) Transformation of different ( , ) to the prices lead to $22.20 ≤ ≤ $134.29 and long-term price in the
range $44.7-$181.27. (b) Each ( , ) implies an initial slope and a unique decomposition of Added Value ($) into
certain and uncertain parts (Parameters other than the initial conditions are according to the base case as
specified in Table 1 and Table 2).

12
Figure 8.b shows the relation between Certain Added Value ($) and Uncertain Added Value ($),
respectively generated by the initial forward maximization and the subsequent trades, which are obtained
under each of the initial conditions. The figure is colored by the initial slope of the forward curve implied
by the initial conditions. A decrease in χ , i.e. a larger initial reduction in the spot price, or an increase in
ξ , i.e. higher long-term price, increases the initial slope of the forward curve. As seen on the plot, the initial
slope is a good determinant of Certain Added Value ($). The solid black lines represent X+Y=constant,
where the constant is Added Value ($) equal to $5, $10, $15 and $20. As the initial condition becomes
profitable, the uncertain portion is higher than the certain portion, in the area under the Y=X line. However,
the situation reverses as we move into the higher Added Value ($) regions. Under very unfavorable
conditions the algorithm does not start due to lack of a profitable trade at t=0. However, as soon as an
opportunity for the initial trade presents itself, it allows the trading process to begin, where the uncertain
portion of profits plays a prominent role. So, the trader may consider to break even on the very first trade
and be more concerned about the opportunities that may arise through future trades by letting the game to
initiate.

4.3. Number of Periods

Different values of N, selected for their divisibility in the number of time steps simulated, are chosen to
study the impact of the frequency of trading on the results; 2 through 10, 12, 14, 16, 18, 20, 24, 32, 36, 63,
96, and 120. In addition to the base case storage cost, the analysis is repeated for two other values of storage
cost; 6.57×2=13.14, and 6.57/2=3.285.
Figure 9.a represents return and risk by Added Value ($) and its standard deviation respectively; both
decreasing as N increases, although at a slowing rate. However, the trend of the HC=13.14 is different,
where N=2 or 3 are inferior to higher choices of N from a risk-return perspective. Figure 9.b illustrates the
Sharpe ratio defined as Added Value ($) divided by its standard deviation. Observing that Sharpe ratio
increases with N, it might be said that the decrease in the risk more than justifies the decrease in the returns.
Also, an increase in N benefits Sharpe ratio more at a lower storage cost. Ultimately, the choice of N will
depend on the risk-return preferences of the trader.
Figure 9. Impact of N on risk-return characteristics and Sharpe ratio
(a) (b)

(a) Added Value ($) and its standard deviation as a function of N for three different HC. (b) Sharpe ratio as a function
of N for three different HC. (Parameters other than N are according to the base case as specified in Table 1 and Table
2).

13
The figure suggests that extra trading hurts the profit, unless the costs are very high, where the changes are
small (except for N=2 and 3). The reason is frequent trading causes earlier termination (delivery) of the
paths, while waiting would have been more profitable on average because likelihood of having a negative
slope increases with the passage of time. Because the simulation started with =$54.45/barrel and a long-
term equilibrium price of $103.19/barrel, the prices increase toward the long-term level, and the initial
upward sloping forward curve tilts downward. Therefore, a path would capture larger gains in the future if
stays alive longer. Essentially the algorithm performs shortsightedly by opting out for smaller gains sooner
in the process. However, frequent trading has tremendous risk reduction effect even by switching from N=2
to N=4.

4.4. Storage Cost

The impact of HC is studied by considering different values of HC from $0 to $20, and the analysis is
repeated for N= 2, 4, 8, and 16. Figure 10.a shows the Added Value ($) versus HC colored by maturity of
the initial contract, where Added Value ($) decreases as HC increases for all values of N. As N increases,
the number of jumps increases but their size shrinks, to the point at which the graphs seem almost smooth
at N=16. The lower the N, the sooner Added Value ($) touches zero due to lack of a contract with suitable
maturity at t=0. For N=2, three regimes are evident in Figure 10.a; (i) HC<10, (ii) 10≤HC<16, and (iii)
16≤HC. They correspond to an initial short position in F(0,1), F(0,0.5), and F(0,0), respectively. As HC
increases, the longer term maturity will become unprofitable, where the quick drop at HC around 10 is seen.
Under regime (iii), the problem finishes trivially by selling the oil at t=0.

Figure 10. Impact of storage cost (HC) on value, initial maturity and risk-return characteristics
(a) (b)

(a) Impact of HC on Added Value ($) colored by maturity of the initial contract. (b) Added Value ($) versus its
standard deviation for different values of HC with labels, where only certain HC labels are shown for clarity
(parameters other than HC and N are per the base case as specified in Table 1 and Table 2).

In Figure 10.b, Added Value ($) is graphed versus the standard deviation, labeled with HC values. The
general trend for N=4, 8, and 16 is that as return increases, risk increases too. However, for N=2, the graph
has two disjoint segments; in the right segment corresponding to 0 ≤ < 10, as return increases, risk
decreases eccentrically. The left segment corresponding to 10 ≤ ≤ 20 has a behavior consistent with

14
the general trend. Change of HC from 9.5 to 10 has a profound impact on the performance of N=2 case,
where it suddenly drops much below the other N. Further investigations revealed that it is due to a sharp
decline in Uncertain Added Value ($); as HC increases above 9.5, the selected maturity becomes 6 months
shorter. So, the period for which a potential refund of HC exists shortens considerably, which translates
into a substantial loss for advancing trades, whereas the gains of postponing trades do not increase
sufficiently to offset the loss. In Figure 10.a, most of the difference in results among different values of N
originates from the difference in Uncertain Added Value ($) as the variation in the Certain Added Value is
limited (<$0.50). Under favorable storage cost ( < 10), F(0,T1) is selected initially, where T1 is equal
or very close to one. Under these conditions, the profit is mainly made by advancing the trade, when the
forward curve becomes less steep or downward sloping, and the payoff comprises of a “refund” of storage
cost plus/minus any spread such that the net is positive. The price dynamics moves in general towards a
better spread, and that is why waiting is on average rewarding. Thus a smaller N leads to a higher profit
due to a longer wait time, as long as an initial F(0,1) position can be achieved. Under very unfavorable
conditions (15 ≤ ), the trading process starts by shorting F(0,T1), where T1<<1. In this case, most of the
value is generated by the subsequent trading, and this is when a larger N is beneficial because it increases
the opportunity of subsequent trading. For moderate values of HC (10 ≤ < 15), the results are mixed
as the two forces compete with each other.
To shed further light on the subject, Figure 11.a shows the contribution of Advancing and Postponing trades
to Uncertain Added Value ($) separately for different values of refund ratio, defined as the proportion of
the prepaid storage cost that the trader would be refunded if he selects an advancing trade. The refund ratio
introduces an asymmetric friction by penalizing the refund transaction, and it takes five values in the range
of [0,1], no-refund to full-refund. The four regimes based on the initially selected forward contract, F(0,1)
to F(0,0.25), can be seen visibly. Regardless of the Refund Ratio, advancing trades contribute more than

Figure 11. Separate contribution of advancing or postponing trades to value, and impact of Refund Ratio.
(a) (b)

(a) Uncertain Added Value ($) generated by Advancing (A) or Postponing (P) trades for different values of Refund
Ratio versus storage cost. (b) Added Value ($) as a function of storage cost (HC) for different values of Refund Ratio
(Parameters other than HC and Refund Ratio are per the base case as specified in Table 1 and Table 2).

15
the postponing ones when ≤ 10.5. Only after 11 ≤ do the contributions of postponing trades begin
to overcome the advancing ones with a complete dominance in the 14.5 ≤ range. If the Refund Ratio
<1, the trader receives an imperfect refund. However, the trader does not consider this fact when initially
deciding to postpone the position, with a detrimental effect on the value depending on HC as seen in Figure
11.b.
Considering the zero-refund case for simplicity, any trade is solely triggered by a profitable spread. Given
the initial condition and the environment dynamics, the spread is usually in favor of the advancing trades.
If the initial contract is F(0,1), the majority of profits are due to advancing. So, the advancing contributions
are almost constant under zero refund, and increase faster with HC as Refund Ratio increases. Also, an
opportunity for postponing trades can only exist after an advancing trade has executed because the F(0,1)
position cannot be postponed further. This is reflected in the trend of the postponing trades, which mimics
the trend of the advancing trades (at any Refund Ratio).
The second regime starts with F(0,0.75), where one-period postponing opportunities are added relative to
F(0,1), while advancing opportunities are reduced by one period. This is why there is a step-like change in
both types. Within the regime, postponing is increasingly more expensive as HC rises, and thus it shows a
declining trend. There is a link between postponing and advancing trades; as the number of one type of the
trades increases (decreases), the number of the other type increases (decreases) as well because each type
shifts the net short position in the opposite direction of the other type through time, which creates potential
trading opportunities. However, a large enough refund advantages advancing trades. Accordingly,
advancing trades exhibit a declining trend for low refund ratios and an increasing one for high refund ratios.
In the third regime, the conditions are similar to the second one, except that postponing trades are more
important than before due to the shorter initial maturity (0.5 instead of 0.75). In the fourth and last regime,
postponing trade is the dominant type because the initial contract maturity is only 0.25; as the number and
contribution of postponing trades decreases, the value generated by the advancing trades decrease too, even
under the full refund. This represents an unfavorable environment, where a small number of trades happens
at marginally small profits due to extremely high storage costs. Across all no-refund cases both
contributions decrease as HC rises.

5. Conclusion
We explore the trading opportunities that arise from an upward sloping forward curve using FDO algorithm.
The unique continuous hedging provided by FDO makes it attractive for risk-averse sellers. The
deficiencies in the existing formulations are addressed and an explicit trading method using forwards is
suggested. Given the pivotal role of forward curve dynamic in the storage trade, the results are interpreted
in terms of the change in mean forward prices (shifting), in forward curve slope (tilting), and in curvature
(bending). Mean forward prices does not contribute in explaining the payoff level or selected maturity
because the trades are based on the spreads rather than the absolute values. The change in the realized slope
through each period, computed as the slope of the line fitted to the curve, well explains the behavior of the
algorithm. The decision making process at each time can be summarized based on the slope realized at that
time and the slope at the preceding time step. The impact of all previous slopes is embedded in the maturity
of the existing contract, shaping possible future choices. This path-dependent behavior is an interesting
avenue of further research to increase added value using conditional expectations even on a stepwise basis.
Given the initial conditions, the forward curve change through time is mainly characterized by an increase
in price levels and a decrease in the slope. Initial condition of the state variables is found to have a significant
effect on the added value. We decompose the total added value into two parts, certain and uncertain, where
it is verified that the initial slope can determine the certain part precisely. The Added Value ($) is maximized
at the highest long-term price and the maximum reduction in the spot price from it. However, the Uncertain

16
Added Value ($) is maximized at the highest long-term price but a moderate deviation of spot from it
because a moderately-sloped forward curve can offer more future trading opportunities.
The impact of number of trading periods on the added value is intertwined with the storage cost
assumptions. If the storage cost is low enough that the maximum maturity contract is selected initially,
additional trading hurts due to opting for small gains sooner as opposed to waiting for larger gains later. If
the storage cost is extremely high, a larger N performs better because it offers a better ability in extracting
uncertain added value when the initial maturity is short. However, for the HC in the mid-range, there is not
a clear best performer and the difference among the results is smaller generally, except N=2, which is much
below the rest. It is found that changes in N strongly influences standard deviation-based risk measures. In
conclusion, to choose an appropriate N, one should consider the regime based on the prevailing storage
cost, and risk-return preferences.
The contribution of advancing and postponing trades to Uncertain Added Value ($) is studied under
different Refund Ratios. Generally, the dominant trade influences the other one through the interlink
between the two trades by creating opportunities. Which trade is dominant is determined based on the
regime, or maturity of the initial contract. Within a fixed regime (expect regime one), contribution of both
types of trade decreases with HC, unless the refund ratio is 0.75 or 1. However, comparing the levels of
postponing trades among different regimes shows a non-decreasing trend moving from regime one to four,
while comparing the trend of advancing trades exhibits a decreasing trend, which is due to change in the
opportunity set. To measure the impact of the Refund Ratio on the total Added Value ($), the difference
between full- and no-refund cases was studied; it is around $1.4 at maximum, which occurs at HC=10.5.

Declaration of Interest:
The authors report no conflicts of interest. The authors alone are responsible for the content and writing of
the paper.

References:
Bhattacharya, S. (1978). Project valuation with mean-reverting cash flow streams. The Journal of Finance.
33 (5): 1317–31.

BIMCO (2016). Tanker shipping: still a strong market as demand stays high. Baltic and International
Maritime Council (BIMCO).
https://www.bimco.org/Reports/Market_Analysis/2016/0125_TankerSMOO_2016-01.aspx (accessed 1
November 2016).

Casassus, J. and Collin-Dufresne, P. (2005). Stochastic convenience yield implied from commodity futures
and interest rates. The Journal of Finance. 60 (5): 2283–331.

Charles R. Weber Company, Inc. (n.d.). http://www.crweber.com/index.php (accessed1 November 2016).

Chen, P.-F., Lee, C.-C. and Zeng, J.-H. (2014). The relationship between spot and futures oil prices: Do
structural breaks matter? Energy Economics. 43: 206–17.

Chinn, M. D., LeBlanc, M. and Coibion, O. (2005). The Predictive content of energy futures: An update on
petroleum, natural gas, heating oil and gasoline. National Bureau of Economic Research.
http://www.nber.org/papers/w11033 (accessed31 October 2016).

17
Considine, T. J. and Larson, D. F. (2001). Uncertainty and the convenience yield in crude oil price
backwardations. Energy Economics. 23 (5): 533–48.

Dias, M. (2005). Real options with petroleum applications. Doctoral thesis, Industrial engineering
department, PUC-Rio, Rio de Janeiro, Brazil.

Dixit, A. K. and Pindyck, R. S. (1994). Investment under uncertainty. Princeton University Press.

Eydeland, A. and Wolyniec, K. (2003). Energy and power risk management: New developments in
modeling, pricing, and hedging. John Wiley & Sons.

Geman, H. and Ohana, S. (2009). Forward curves, scarcity and price volatility in oil and natural gas markets.
Energy Economics. 31 (4): 576–85.

Gibson, R. and Schwartz, E. S. (1990). Stochastic convenience yield and the pricing of oil contingent
claims. The Journal of Finance. 45 (3): 959–76.

Gray, J. and Khandelwal, P. (2004). Towards a realistic gas storage model. Commodities Now. (June): 1–
4.

Hahn, W. J., DiLellio, J. A. and Dyer, J. S. (2014). What do market-calibrated stochastic processes indicate
about the long-term price of crude oil? Energy Economics. 44: 212–21.

Hilliard, J. E. and Reis, J. (1998). Valuation of commodity futures and options under stochastic convenience
yields, interest rates, and jump diffusions in the spot. Journal of Financial and Quantitative Analysis. 33
(1): 61–86.

Jafarizadeh, B. and Bratvold, R. (2013). Sell spot or sell forward? Analysis of oil-trading decisions with
the two-factor price model and simulation. SPE Economics & Management. 5 (3): 80–8.

Kaldor, N. (1939). Speculation and economic stability. The Review of Economic Studies. 7 (1): 1–27.
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic
Engineering. 82 (1): 35–45.

Lai, G., Margot, F. and Secomandi, N. (2010). An approximate dynamic programming approach to
benchmark practice-based heuristics for natural gas storage valuation. Operations Research. 58 (3): 564–
82.

Litzenberger, R. H. and Rabinowitz, N. (1995). Backwardation in oil futures markets: Theory and empirical
evidence. The Journal of Finance. 50 (5): 1517–45.

Maragos, S. (2002). Valuation of the operational flexibility of natural gas storage reservoirs. In: E. Ronn
(ed.). Real options and energy management. London: Risk Books, pp. 431–456.

Maslyuk, S. and Smyth, R. (2009). Cointegration between oil spot and future prices of the same and
different grades in the presence of structural change. Energy Policy. 37 (5): 1687–93.

Parsons, C. (2013). Quantifying natural gas storage optionality: a two-factor tree model. The Journal of
Energy Markets. 6 (1): 95–VI.

Pindyck, R. S. (1999). The Long-Run Evolution of Energy Prices. The Energy Journal. 20 (2): 1–27.

18
View publication stats

Ribeiro, R., Diana and Hodges, S. D. (2004). A Two-Factor Model for Commodity Prices and Futures
Valuation. Basel, Switzerland: Social Science Research Network. http://papers.ssrn.com/abstract=498802
(accessed 12 September 2016).

Schoftner, R. (2008). On the estimation of credit exposures using regression-based Monte Carlo simulation.
Journal of Credit Risk. 4 (4): 37–63.

Schwartz, E. S. (1997). The stochastic behavior of commodity prices: implications for valuation and
hedging. The Journal of Finance. 52 (3): 923–73.

Schwartz, E. and Smith, J. E. (2000). Short-term variations and long-term dynamics in commodity prices.
Management Science. 46 (7): 893–911.

Secomandi, N., Lai, G., Margot, F., Scheller-Wolf, A. and Seppi, D. J. (2015). Merchant commodity storage
and term-structure model error. Manufacturing & Service Operations Management. 17 (3): 302–20.

Ware, A. (2013). Accurate semi-lagrangian time stepping for stochastic optimal control problems with
application to the valuation of natural gas storage. SIAM Journal on Financial Mathematics. 4 (1): 427–51.

Working, H. (1948). Theory of the inverse carrying charge in futures markets. Journal of Farm Economics.
30 (1): 1–28.

Ye, S. and Karali, B. (2016). The informational content of inventory announcements: Intraday evidence
from crude oil futures market. Energy Economics. 59: 349–64.

19