Abstract
We formulate a probabilistic Markov property in discrete time under a dynamic risk framework with minimal assumptions. This is useful for recursive solutions to risk-sensitive versions of dynamic optimisation problems such as optimal prediction, where at each stage the recursion depends on the whole future. The property holds for standard measures of risk used in practice, and is formulated in several equivalent versions including a representation via acceptance sets, a strong version, and a dual representation.
Similar content being viewed by others
1 Introduction
The Markov property is a main tool used in the dynamic evaluation of risk, for example in the solution of risk-sensitive optimisation problems. In this paper we present a probabilistic formulation of the Markov property under a risk framework with minimal assumptions, which we call dynamic conditional risk mappings, and give applications to optimal prediction, a class of risk-sensitive stochastic optimisation problems.
To fix ideas, let \(X = (X_t)_{t \in {\mathbb {N}}_0}\) be a Markov chain taking values in a measurable space E and let \((\Omega ,{\mathcal {F}},({\mathcal {F}}_t)_{t \in {\mathbb {N}}_0},{\mathbb {P}}^x)\) be its canonical probability space, where \(X_0=x\), \({\mathbb {P}}^x-\)a.s. and \({\mathbb {N}}_0=\{0,1,\ldots \}\). Let \(\varrho =((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) be the family of conditional linear expectations given by
where Z is an arbitrary bounded random variable depending on the whole sample path (that is, Z is measurable with respect to \((\Omega ,{\mathcal {F}})\)). Then \(\varrho \) is Markovian in the sense that
where \(\theta _t\) is the shift operator, and we would like to generalise this property to an appropriately large class of nonlinear (that is, risk-sensitive) families \(\varrho \).
A number of settings have been given for the Markov property under dynamic risk frameworks. Broadly they are formulated either on functions of the state of the Markov process (that is, analytically), or on the canonical probability space (that is, probabilistically). In the linear case the probabilistic and analytic formulations are equivalent, and below we obtain sufficient conditions on \(\varrho \) for their equivalence (Proposition 2.11).
Analytic formulations, which are often based on so-called transition risk mappings (cf. Definition 4.1), are useful in recursive solution techniques which evaluate risk only one step ahead, taking \(Z=f(X_{t+1})\) in (1). Since probabilistic formulations apply to the whole path of X, they are useful for recursions which directly evaluate risk multiple steps ahead, taking \(Z=f(X_{t+1}, X_{t+2},...)\). In optimal prediction problems, for example, the evaluation of risk depends on the evolution of the process after a user-selected stopping time: for instance, the problem of stopping as close as possible to the ultimate maximum of a time-homogeneous Markov chain X taking values in \(E = {\mathbb {R}}\) (cf. Allaart 2010; Yam et al. 2009 in the case of linear expectation):
where \(T \in {\mathbb {N}}_0\), \(X_T^* := \max _{0 \le s \le T} X_s\), and \({\mathscr {T}}_{[0,T]}\) is the set of stopping times taking values in \(\{0,1,\ldots ,T\}\) (see also du Toit and Peskir 2007; Pedersen 2003 for work in continuous time). In the aforementioned studies, explicit solutions have been obtained for this problem by applying the probabilistic Markov property to represent the objective as a function of \(\tau \) and \(X_\tau \), obtaining a function F such that
see e.g. page 1077 in Allaart (2010). The probabilistic Markov property, which is satisfied by the commonly used entropic, mean semi-deviation, \({\text {VaR}}\), \({\text {AVaR}}\) and worst-case risk mappings (see Sect. 3), is applied in Sect. 5.1 to solve (3) recursively.
The evaluation of risk for random variables via so-called sublinear (and therefore convex) functionals goes back to Lebedev (1992, 1993), where results including dual representations are obtained in both the static and conditional settings. To enable dynamic programming for risk-sensitive Markov decision processes, dynamic risk-sensitive frameworks have also been proposed using analytic formulations of the Markov property under the assumption of time consistency. In Ruszczyński (2010) a dynamic setting is introduced in which risk-sensitive Markov decision processes are studied in both finite and infinite time horizon. Also with infinite time horizon, the average risk of controlled Markov processes is studied in Shen et al. (2013) while Çavuş and Ruszczyński (2014) address the undiscounted total risk of transient controlled Markov processes. In Fan and Ruszczyński (2018a, 2018b) a structure for dynamic risk measures is introduced based on a stronger concept of stochastic conditional time consistency. Utility-based (also known as certainty equivalent) frameworks are special cases using the Markov property under linear expectation, see for example Bäuerle and Rieder (2014), Bäuerle and Rieder (2017). Other frameworks are presented in Bartl (2020) using analytic sets and in Pichler and Schlotter (2020) using the Kusuoka representation.
While the dynamic risk-sensitive frameworks above involve a reference probability measure, analytic settings of the Markov property also exist in risk-sensitive frameworks without such a measure. When the state space E is finite, these frameworks include Markov chains under imprecise expectations, which are related to sensitivity analyses under a set of possible transition probabilities for the Markov process \((X_{t})_{t \in {\mathbb {N}}_0}\), see for example de Cooman et al. (2009), Hartfiel (1998), Krak et al. (2017). More generally they include nonlinear expectations which, in Peng (2005) and Nendel (2021), are related to finite-dimensional properties of so-called nonlinear Markov chains. As in the present paper, in Denk et al. (2018) the framework is related to the infinite dimensional path space of Markov processes, although convexity of the nonlinear expectation is then assumed. Also without a reference measure, the risk forms of Dentcheva and Ruszczyński (2020) have been applied to the optimisation of partially observable two-stage systems.
The general study of dynamic conditional risk mappings can also be approached via backward stochastic differential or difference equations, see Cohen and Elliott (2008), Cohen and Elliott (2010). In contrast to the latter setup our risk mappings do not assume time consistency. In the other direction, in Martyr et al. (2022) reflected backward stochastic difference equations are derived from dynamic conditional risk mappings, in the study of non-Markovian optimal switching problems.
In the present work we assume a reference measure and make minimal further assumptions. Time consistency is not assumed, making our formulation applicable to risk mappings including mean semi-deviation and average value at risk (cf. Sect. 3). In the time-consistent case we make the connection to analytic formulations, and provide a recursive solution to the optimal prediction problem.
For convex risk mappings we characterise the Markov property in terms of the dual representation (see for example Artzner et al. 1999; Delbaen 2002; Detlefsen and Scandolo 2005; Frittelli and Rosazza Gianin 2002; Lebedev 1992, 1993). More precisely, we show that a Markovian convex risk mapping can be characterised as a supremum over penalised linear expectations with respect to certain transition kernels, extending the dual representation of transition risk mappings beyond the coherent case studied in Ruszczyński (2010). We also obtain sufficient conditions under which the latter structure implies the probabilistic Markov property.
The paper is structured as follows. Section 2 provides the probabilistic framework, together with equivalences between versions of the Markov property, and a representation in terms of acceptance sets. Section 3 gives examples and Sect. 4 addresses the dual representation, while applications to optimisation problems are given in Sect. 5.
2 A probabilistic Markov property for risk mappings
After presenting the setup and briefly recalling necessary definitions (Sect. 2.1), in Sects. 2.2 and 2.3 we provide our novel probabilistic setting for the Markov property and establish equivalent forms. The Markov property in terms of acceptance sets is studied in Sect. 2.4.
2.1 Setup and notation
Suppose we have an E-valued time-homogeneous Markov process \((X_{t})_{t \in {\mathbb {N}}_0}\) with respect to the filtered probability space \((\Omega ,{\mathcal {F}},{\mathbb {F}},{\mathbb {P}})\), where:
-
E is a Polish space equipped with its Borel \(\sigma \)-algebra \({\mathcal {E}}\),
-
\({\mathbb {N}}_0= \{0,1,2,\ldots \}\) is the discrete time parameter set,
-
\(\Omega \) is the canonical space of trajectories \(\Omega = E^{{\mathbb {N}}_0}\),
-
X is the coordinate mapping, \(X_{t}(\omega ) = \omega (t)\) for \(\omega \in \Omega \) and \(t \in {\mathbb {N}}_0\),
-
\({\mathbb {F}} = ({\mathcal {F}}_{t})_{t \in {\mathbb {N}}_0}\) with \({\mathcal {F}}_{t} = \sigma (\{X_{s} :s \le t\})\) the natural filtration generated by X and \({\mathcal {F}} = \sigma (\bigcup _{t \in {\mathbb {N}}_0} {\mathcal {F}}_{t})\).
Let \({\mathscr {P}}({\mathcal {F}})\) denote the set of probability measures on \((\Omega ,{\mathcal {F}})\). Unless otherwise specified, all inequalities between random variables will be interpreted in the almost sure sense with respect to the appropriate probability measure. We write \({\mathscr {T}}\) for the set of finite-valued stopping times and \({\mathscr {T}}_{[t,T]}\) for the set of stopping times taking values in \(\{t,t+1,\ldots ,T\}\). We denote by \(b{\mathcal {F}}\) the space of bounded random variables on \((\Omega ,{\mathcal {F}})\) and similarly for other \(\sigma \)-algebras. It will also be convenient to define \({\mathcal {F}}_{t,\infty } = \sigma (X_{s} :s \ge t)\) and \({\mathcal {F}}_{t,t} = \sigma (X_t)\).
In the above setup the following objects exist:
-
The law \(\mu ^{X_{0}}\) of \(X_0\) under \({\mathbb {P}}\) and a family of probability measures defined by the measurable mapping \(x \mapsto {\mathbb {P}}^x\) from E to \({\mathscr {P}}({\mathcal {F}})\), which is a disintegration of \({\mathbb {P}}\) with respect to \(X_0\) (see Dellacherie and Meyer 1978, p. 78). To be precise, this family satisfies \({\mathbb {P}}^{x}(X_{0} = x) = 1\) and for every \(F \in {\mathcal {F}}\) we have
$$\begin{aligned} {\mathbb {P}}(F) = \int _{E}{\mathbb {P}}^{x}(F)\,\mu ^{X_{0}}({\textrm{d}}x). \end{aligned}$$ -
A time-homogeneous Markov transition kernel \(q^{X} :{\mathcal {E}} \times E \rightarrow [0,1]\) such that for every \(x \in E\) and \(B \in {\mathcal {E}}\) we have \(q^{X}(B \vert x) = {\mathbb {P}}^{x}\big (X_{1} \in B \big )\),
-
Markov shift operators \(\theta _{t} :\Omega \rightarrow \Omega \), \(t \in {\mathbb {N}}_0\) such that \(\theta _{0}(\omega ) = \omega \), \(\theta _{t} \circ \theta _{s} = \theta _{t+s}\) and \((X_{t} \circ \theta _{s})(\omega ) = X_{t+s}(\omega )\) for each \(\omega \in \Omega \) and \(s,t \in {\mathbb {N}}_0\).
For \(\tau \in {\mathscr {T}}\) define the random shift operator \(\theta _{\tau }\) by
We recall the definitions of risk mapping and conditional risk mapping (which are interchangeable via the mapping \(Z \mapsto \rho (-Z)\) with the monetary conditional risk measures of Föllmer and Schied 2016, Def. 11.1):
Definition 2.1
(Risk mapping) A risk mapping on the probability space \((\Omega ,{\mathcal {F}},{\mathbb {P}}^x)\) is a function \(\rho ^x :b{\mathcal {F}} \rightarrow {\mathbb {R}}\) satisfying
-
Normalisation: \(\rho ^x(0) = 0\),
-
Translation invariance: \(\forall \; Z \in b{\mathcal {F}}\) and \(c \in {\mathbb {R}}\) we have \(\rho ^x(Z + c) = c + \rho ^x(Z)\),
-
Monotonicity: \(\forall \; Z,Z' \in b{\mathcal {F}}\), we have \(Z \le Z' \, {\mathbb {P}}^x\text {-a.s.} \implies \rho ^x(Z) \le \rho ^x(Z')\).
Definition 2.2
(Conditional risk mapping) A conditional risk mapping on the probability space \((\Omega ,{\mathcal {F}},{\mathbb {P}}^x)\) with respect to the \(\sigma \)-algebra \({\mathcal {F}}_t \subseteq {\mathcal {F}}\) is a function \(\rho _t^x :b{\mathcal {F}} \rightarrow b{\mathcal {F}}_t\) satisfying:
-
Normalisation: \(\rho _t^x(0) = 0\) \({\mathbb {P}}^x\)-a.s.,
-
Conditional translation invariance: \(\forall \; Z \in b{\mathcal {F}}\) and \(Z' \in b{\mathcal {F}}_t\),
$$\begin{aligned} \rho _t^x(Z + Z') = Z' + \rho _t^x(Z), \qquad {\mathbb {P}}^x\text {-a.s.} \end{aligned}$$ -
Monotonicity: \(\forall \; Z,Z' \in b{\mathcal {F}}\),
$$\begin{aligned} Z \le Z' \, {\mathbb {P}}^x\text {-a.s.} \implies \rho _t^x(Z) \le \rho _t^x(Z') \,{\mathbb {P}}^x\text {-a.s.} \end{aligned}$$
Conditional risk mappings also satisfy the following property (cf. Cheridito et al. 2006, Prop. 3.3 and Föllmer and Schied 2016, Ex. 11.1.2):
-
Conditional locality: for every Z and \(Z'\) in \(b{\mathcal {F}}\) and \(A \in {\mathcal {F}}_t\), we have \({\mathbb {P}}^x\)-a.s.
$$\begin{aligned} \rho _t^x(\mathbbm {1}_{A}Z + \mathbbm {1}_{A^{c}}Z') = \mathbbm {1}_{A}\rho _t^x(Z) + \mathbbm {1}_{A^{c}}\rho _t^x(Z'). \end{aligned}$$
Definition 2.3
(Dynamic conditional risk mapping) For each \(x \in E\) a dynamic conditional risk mapping on the filtered probability space \((\Omega ,{\mathcal {F}},{\mathbb {F}},{\mathbb {P}}^x)\) is a sequence \((\rho _{t}^x)_{t \in {\mathbb {N}}_0}\) where
-
\(\rho _0^x\) is a risk mapping,
-
for each \(t \ge 1\), \(\rho _{t}^x\) is a conditional risk mapping on \((\Omega ,{\mathcal {F}},{\mathbb {P}}^x)\) with respect to \({\mathcal {F}}_{t}\).
We use the superscript x in \((\rho _{t}^x)_{t \in {\mathbb {N}}_0}\) to indicate a dynamic conditional risk mapping on \((\Omega ,{\mathcal {F}},{\mathbb {F}},{\mathbb {P}}^x)\).
Note that the codomain of \(\rho _0^x\) is \({\mathbb {R}}\) while, for each \(t \ge 1\), the codomain of \(\rho _t^x\) is \(b{{\mathcal {F}}_t}\). This setup is motivated by the fact that any \({\mathcal {F}}_0\)-measurable random variable is \({\mathbb {P}}^x\)-a.s. constant. For example, for each \(x \in E\), the sequence \((\rho _t^x)_{t \in {\mathbb {N}}_0}\) given by (1) is a dynamic conditional risk mapping.
For a finite stopping time \(\tau \) define
noting that \(\rho _{\tau } :b{\mathcal {F}} \rightarrow b{\mathcal {F}}_{\tau }\).
In some results below we will assume continuity.
Definition 2.4
Let \(t\in {\mathbb {N}}_0\), \(x \in E\). We say that \(\rho ^x_{t}\) is continuous from below (resp. from above) if \(\rho ^x_{t}(Y_n) \rightarrow \rho ^x_{t}(Y)\) \({\mathbb {P}}^x\)-a.s. for every increasing (resp. decreasing) sequence \((Y_n)_{n \in {\mathbb {N}}_0}\) in \(b{\mathcal {F}}\) converging \({\mathbb {P}}^x\)-a.s. to \(Y \in b{\mathcal {F}}\).
Note that results for decreasing risk maps (e.g. in Föllmer and Penner 2006) requiring continuity from above can be applied to increasing risk maps of Definitions 2.1–2.3 if continuity from below is assumed.
2.2 Markov property
We begin with measurability with respect to the initial state of the Markov process, referring to this as regularity.
Definition 2.5
(Regularity) A collection of risk mappings \((\rho ^x)_{x \in E}\) is said to be regular if for all \(Z \in b{\mathcal {F}}\) the map \(x \mapsto \rho ^x(Z)\) is bounded and measurable.
Definition 2.6
(Markov property) The family \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) of dynamic conditional risk mappings satisfies the Markov property (for the chain \((X_t)_{t \in {\mathbb {N}}_0}\)) if
-
1.
\((\rho _0^x)_{x \in E}\) is regular,
-
2.
for each \(x \in E\), \(Z \in b{\mathcal {F}}\) and \(t \in {\mathbb {N}}_0\) we have
$$\begin{aligned} \rho ^{x}_{t}(Z \circ \theta _{t}) = \rho ^{X_{t}}(Z) \;\;{\mathbb {P}}^{x}\text {-a.s.}, \end{aligned}$$(4)
where \(\rho ^{X_{t}}(Z)\) is interpreted as the random variable \(\omega \mapsto \rho ^{X_{t}(\omega )}(Z)\).
By construction, if \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) is a family of dynamic conditional risk mappings then \((\rho _0^x)_{x \in E}\) is a collection of risk mappings. For convenience we often write \(\rho ^x\) for \(\rho ^x_0\).
In particular we have
Note that the linear conditional expectation (1) satisfies this Markov property and corresponds to the risk-neutral case. Examples of \(\rho \) which are risk sensitive are presented in Sect. 3.
Remark 2.7
Note that (4) could have been specified differently. For example, by relating all risk mappings \(\rho _t^x\) to the same regular collection \((\rho ^x)_{x \in E}\) in (4) we have imposed a time homogeneity on the measurement of risk. This is not essential, since taking a collection \(\{\rho ^{x,s}: x \in E, s \in {\mathbb {N}}_0\}\) indexed also by time and specifying
the family of dynamic conditional risk mappings may be time-heterogeneous.
A regular collection \((\rho ^x)_{x \in E}\) of risk mappings can also be used to construct a Markovian family \(\varrho =((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) satisfying Definition 2.6, as follows. We use the fact that any bounded \({\mathcal {F}}\)-measurable random variable Z can be represented as \(Z=f(X_0,X_1,\ldots )\) for some measurable and bounded function \(f :E^{{\mathbb {N}}_0} \rightarrow {\mathbb {R}}\), which follows by standard monotone class arguments, see Blumenthal and Getoor (1968), Prop. 0.2.7 or Çinlar (2011), Th. 2.4.4. As it is obtained without reference to any probability measure, the equality \(Z=f(X_0,X_1,\ldots )\) holds for all (rather than almost all) \(\omega \in \Omega \) and therefore the function f is unique.
Proposition 2.8
Let \((\rho ^x)_{x \in E}\) be regular. For each \(x \in E\), \(t \in {\mathbb {N}}_0\) and \(Z=f(X_0,X_1,\ldots )\) let
where
Then for each \(x \in E\), \((\rho ^x_t)_{t \in {\mathbb {N}}_0}\) is a dynamic conditional risk mapping and the family \(\varrho = ((\rho ^x_t)_{t \in {\mathbb {N}}_0})_{x \in E}\) satisfies the Markov property.
Proof
Let \(x \in E\), \(t \in {\mathbb {N}}_0\) and \(\omega \in \Omega \) be arbitrary. For compactness we will write \(X_{0:t}(\omega )\) for \((X_0(\omega ),\ldots , X_t(\omega )) \in E^{t+1}\). Clearly \(\rho _t^x\) is normalised, so we check conditional translation invariance and monotonicity. Taking \(Z = f(X_0,X_1,\ldots ) \in b{\mathcal {F}}\) and \(W = g(X_0,\ldots ,X_t) \in b{\mathcal {F}}_t\), by construction we have
To check monotonicity let \(Z = f(X_0,X_1,\ldots )\) and \(Z' = f'(X_0,X_1,\ldots )\) be two bounded random variables such that \(Z \le Z'\) \({\mathbb {P}}^x\)-a.s. We first show that \(Z_t(X_{0:t}(\omega )) \le Z'_t(X_{0:t}(\omega )), {\mathbb {P}}^{X_t(\omega )}\)-a.s. Writing as usual \({\mathbb {P}}^x(A \vert {\mathcal {F}}_t)\) for \({\mathbb {E}}^x[1_A \vert {\mathcal {F}}_t]\) for each \(A \in {\mathcal {F}}\), and applying conditional locality and the Markov property, for almost all \(\omega \) we have, with a slight abuse of notation
By the monotonicity of \(\rho ^{X_t(\omega )}\) we then have that \({\mathbb {P}}^x\)-a.s.,
Lastly we verify the Markov property for the family \(\varrho \). For \(Z=f(X_0,X_1,\ldots )\) we have by construction and (5) that \({\mathbb {P}}^x\)-a.s.
\(\square \)
2.3 Equivalent forms of the Markov property
Just as for the linear conditional expectation, the Markov property for risk mappings can be stated in several equivalent forms. We begin with the strong Markov property.
Proposition 2.9
(Strong Markov Property) If \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) satisfies the Markov property then for any stopping time \(\tau \in {\mathscr {T}}\) and \(Z \in b{\mathcal {F}}\) we have
Proof
Using \(\{\tau = t\} \in {\mathcal {F}}_{t}\), conditional locality and the Markov property we have \({\mathbb {P}}^{x}\)-a.s.:
\(\square \)
To make a connection to one-step Markov properties we will require time consistency:
Definition 2.10
The family \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) is said to be time consistent if for all \(Y,Z \in b{\mathcal {F}}\), \(t \in {\mathbb {N}}_0\) and \(x \in E\) we have
We say that a regular collection \((\rho ^x)_{x \in E}\) of risk mappings is time consistent if the associated Markovian dynamic conditional risk mapping (constructed in Proposition 2.8) is time consistent.
It is well known (see e.g. Acciaio and Penner 2011, Prop. 1.16) that we then have the following recursive relation: for every \(x \in E\) and \(0 \le s\le t\),
As noted in Föllmer and Schied (2016), Exercise 11.2.2, this relation can be generalised to stopping times: for any bounded stopping times \(\tau _1 \le \tau _2\) one has
Since risk mappings are nonlinear in general, in the next proof we use a non-standard version of the Monotone Class Theorem (see Appendix 1) which, unlike Blumenthal and Getoor (1968), Th. 0.2.3, does not appeal to vector spaces.
Proposition 2.11
Let \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) be a family of dynamic conditional risk mappings such that \((\rho ^x)_{x \in E}\) is regular. Let each \(\rho _t^x\) be continuous from above and below: that is, \(\rho ^x_{t}(Y_n) \rightarrow \rho ^x_{t}(Y)\) a.s. for every \(t\in {\mathbb {N}}_0\), \(x \in E\) and monotone sequence \((Y_n)_{n \in {\mathbb {N}}_0}\) in \(b{\mathcal {F}}\) converging to \(Y \in b{\mathcal {F}}\). Then
-
(i)
\(\varrho \) is Markov iff for all \(k\ge 0\) the k-step Markov property holds:
$$\begin{aligned} \rho _t^x(f(X_{t+1},\ldots ,X_{t+k})) = \rho ^{X_t}(f(X_1,\ldots ,X_k)), \end{aligned}$$(7)\({\mathbb {P}}^x\)-a.s. for all \(t \in {\mathbb {N}}_0\), \(x \in E\) and bounded measurable functions \(f:E^{k} \rightarrow {\mathbb {R}}\).
-
(ii)
If the family \(\varrho \) is time consistent, then \(\varrho \) is Markov iff the one-step Markov property holds: for every \(t\in {\mathbb {N}}_0\), \(x \in E\) and bounded measurable function \(f :E \rightarrow {\mathbb {R}}\) we have
$$\begin{aligned} \rho _t^x(f(X_{t+1})) = \rho ^{X_t}(f(X_1)), \qquad {\mathbb {P}}^x\text {-a.s.}. \end{aligned}$$(8)
Remark 2.12
The assumptions of this proposition simplify in the case of convex risk mappings, for which continuity from above implies continuity from below (see proof of Corollary 4.3).
Proof
For both claims (i) and (ii), the ‘only if’ part is trivial and so it remains to establish the ‘if’ part. We first prove this for claim (ii).
Therefore let the family \(\varrho \) be time consistent and suppose that the one-step Markov property (8) holds. We begin by showing, proceeding by induction on k, that the Markov property (4) holds for the class of simple functions – that is, functions of the form
This is true for \(k = 1\) since this is a special case of the one-step Markov property (8). Suppose it is also true for some \(k \ge 1\). We have
By taking all possible intersections of the sets \(A_{11},\ldots ,A_{1n}\) and their complements, we can define \(N \ge n\) mutually disjoint sets \({\tilde{A}}_{1},\ldots ,{\tilde{A}}_{N}\) belonging to \({\mathcal {E}}\) such that
where \({\tilde{\alpha }}_{\ell j} = \alpha _{j}\) if \(A_{1j} \cap {\tilde{A}}_{\ell } \ne \emptyset \) and \({\tilde{\alpha }}_{\ell j} = 0\) otherwise. Therefore we can rewrite f in (10) as
where the \({\tilde{A}}_{\ell }\) are mutually disjoint and each \(f_{\ell }\) has the form (9). Using the local property and time consistency for \(\rho _t^x\), the induction hypothesis and the one-step Markov property we have
Note that for every realisation \(x_{t}\) of \(X_{t}(\omega )\) we have that almost surely under \({\mathbb {P}}^{x_t}\),
Therefore, by (11)–(13) and time-consistency we have for almost every \(\omega \in \Omega \):
and, by induction, the Markov property (4) holds for all functions f of the form (9).
Next we appeal to the monotone class theorem. Let \({\mathscr {H}}_0\) be the set of random variables having the form \(Z=f(X_0,\ldots ,X_{k})\) for some \(k \in {\mathbb {N}}_0\) and some f of the form (9). Clearly \({\mathscr {H}}_0\) is closed under the operation of taking the pointwise minimum. Let
We show that \({\mathscr {H}}_0 \subset {\mathscr {H}}\). Suppose that \(Z=f(X_0,\ldots , X_k) \in {\mathscr {H}}_0\). Then by conditional locality and the fact that \(Z\in {\mathscr {H}}_0\) we have that for each \(\omega \in \Omega \),
i.e. \(Z \in {\mathscr {H}}\).
The space \({\mathscr {H}}\) is closed under monotone limits and Theorem A.1 implies that \({\mathscr {H}}\) contains all bounded \(\sigma ({\mathscr {H}}_0)\)-measurable functions. Since \(\sigma ({\mathscr {H}}_0) = {\mathcal {F}}\) we conclude that the Markov property (4) holds on \({\mathscr {H}} = b{\mathcal {F}}\), completing the proof of claim (ii).
To prove claim (i), note that (7) applies directly to all functions f of the form (9), in which case time consistency does not need to be assumed. We then appeal to the monotone class theorem as we did for claim (ii). \(\square \)
The following result shows that Markovian families of conditional risk mappings which are continuous from above and below can be represented using the canonical form given in Proposition 2.8.
Proposition 2.13
Let \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) be a family of dynamic conditional risk mappings continuous from above and below. Then \(\varrho \) satisfies the Markov property if and only if for all \(x \in E\), \(t \in {\mathbb {N}}_0\), \(Z = f(X_{0},X_{1},\ldots ) \in b{\mathcal {F}}\) and \({\mathbb {P}}^{x}\)-almost every \(\omega \in \Omega \) we have
where \(Z_{t}(x_{0},\ldots ,x_{t}) :=f(x_{0},\ldots ,x_{t},X_{1},X_{2},\ldots )\).
The proof is omitted as it follows the path analogous to that of Proposition 2.11, namely showing the claimed property for simple random variables and appealing to the monotone class theorem.
Remark 2.14
-
(i)
If in addition to the hypotheses of Proposition 2.13 the family \(\varrho \) is time consistent, then we recover a version of the Markov property which is similar to that of Nendel (2021):
$$\begin{aligned} \rho ^{x}(Z) = \rho ^{x}\Big (\rho ^{X_{t}}(Z_{t}(X_{0},\ldots ,X_{t}))\Big ), \quad \forall x \in E, t \in {\mathbb {N}}_0, Z \in b{\mathcal {F}}, \end{aligned}$$where \(Z = f(X_{0},X_{1},\ldots )\) and \(Z_{t}(x_{0},\ldots ,x_{t}) :=f(x_{0},\ldots ,x_{t},X_{1},X_{2},\ldots )\). Note that in Definition 1.2 in Nendel (2021) random variables of the form \(Z = f(X_{0},X_{1},\ldots , X_t,X_{t+s})\) are taken into account.
-
(ii)
In Denk et al. (2018) a Kolmogorov-type theorem is established for conditional risk mappings which, like Proposition 2.13, leads to a risk mapping on path space, and Example 5.3 of the latter paper explores the case of discrete-time Markov chains.
2.4 Markov property in terms of acceptance sets
Particularly in the context of mathematical finance, conditional risk mappings can be characterised by their acceptance sets \({\mathcal {A}}_t^x\) (see Acciaio and Penner 2011, Sec. 1.4.1 or Föllmer and Schied 2016, Sec. 4.1), where
and so for completeness we also formulate the Markov property in these terms. First define another acceptance set, which will be useful in formulating the Markov property:
Note that for any \({\mathcal {F}}_{t,\infty }\)-measurable random variable \(Y={\hat{Y}} \circ \theta _t\) with the representation \(Y = f(X_t,X_{t+1},\ldots )\) one can define \(Y\circ \theta _{-t} := {\hat{Y}} = f(X_0,X_1,\ldots )\).
Lemma 2.15
The family \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) is Markov if and only if \(\rho _t^x :b{\mathcal {F}}_{t,\infty } \rightarrow b{\mathcal {F}}_{t,t}\) and for each \(x \in E\), \(t \in {\mathbb {N}}_0\) and \(Z \in b{\mathcal {F}}\) we have
Proof
Necessity is obvious. Conversely, suppose that \(\rho _t^x :b{\mathcal {F}}_{t,\infty } \rightarrow b{\mathcal {F}}_{t,t}\) and that the equivalence (14) holds. Fix \(Z \in b{\mathcal {F}}\), \(t \in {\mathbb {N}}_0\) and \(x \in E\).
Step 1. We show that
Proof of ‘\(\ge \)’. Let \(g(y) := \rho ^y(Z)\) for \(y\in E\). Then \({\mathbb {P}}^y\)-a.s. we have \(g(X_t)\circ \theta _{-t} = \rho ^{X_0}(Z) = \rho ^y(Z)\). Since \(\rho ^y(Z-g(X_t)\circ \theta _{-t}) = \rho ^y(Z-\rho ^y(Z)) = 0\) we have \(\rho ^{X_t}(Z-g(X_t)\circ \theta _{-t}) =0\), implying \(Z - g(X_t)\circ \theta _{-t} \in \tilde{{\mathcal {A}}}_t^{x}\).
Proof of ‘\(\le \)’. Let \(Y=g(X_t)\) belong to the set on the right-hand side of (15). Then for \(\Omega _1 = \{\omega \in \Omega : \rho ^{X_t}(Z - Y \circ \theta _{-t}) \le 0\}\) we have \({\mathbb {P}}^{x}(\Omega _1) = 1\). Let \(\Omega _0 = \{\omega \in \Omega : \rho ^{X_t}(Z)>Y\}\). To show that \(\Omega _0 \subset \Omega _1^c\), let \(\omega \in \Omega _0\) and \(x_t := X_t(\omega ) = \omega (t)\). Since \(\omega \in \Omega _0\), we have that \(\rho ^{X_t(\omega )}(Z) > Y(\omega )\), which is equivalent to \(\rho ^{x_t}(Z)>g(x_t)\). Then
i.e. \(\omega \in \Omega _1^c\). Since \({\mathbb {P}}^x(\Omega _1^c) = 0\), it follows that \({\mathbb {P}}^x(\Omega _0) = 0\), which finishes the proof of the claim of Step 1.
Step 2. To finish the proof, note from Acciaio and Penner (2011), Prop. 1.2 (modulo a minus sign which appears because Acciaio and Penner (2011) considers decreasing risk mappings) that
where the last inequality follows from the fact that \(Y = \rho _t^x(Z \circ \theta _t) \in b{\mathcal {F}}_{t,t}\) and \(Z \circ \theta _t - \rho _t^x(Z \circ \theta _t) \in {\mathcal {A}}_t^x\). Since for \(Y \in b{\mathcal {F}}_{t,t}\) we have from (14) that
Step 1 completes the proof. \(\square \)
Remark 2.16
The above lemma implies in particular that for a Markovian risk map, if \(Z \in b{\mathcal {F}}_{t,\infty }\), then \(\rho _t^x(Z)\) is \(\sigma (X_t)\)-measurable.
3 Examples
In this section we provide examples of Markovian families of dynamic conditional risk mappings. Note that the entropic and worst case risk mappings are time consistent (see Detlefsen and Scandolo 2005, Prop. 6 and Barron et al. 2003, Th. 2.8(b)(ii) respectively), while the mean semi-deviation risk mapping and average value at risk are not (Föllmer and Schied 2016, Ex. 11.13, Artzner et al. 2007, pp. 20–21). Below we take \(Z \in b{\mathcal {F}}, t \in {\mathbb {N}}_0, x \in E\).
3.1 Composite risk mappings
Let \(K \in {\mathbb {N}}_0\) and for \(k = 0,\ldots ,K\) let \(g_{k} :{\mathbb {R}}^{m_{k}} \times E \rightarrow {\mathbb {R}}\) be measurable functions bounded on compact sets such that \(m_{0} = 1\) and \(m_k = 2\) for \(k\ge 1\), with the map \(x \mapsto g_{k}(r_{k},x)\) bounded on E for every \(r_{k} \in {\mathbb {R}}^{m_{k}}\). Assume also that for each \(x \in E\), the sequence \((\rho _t^x)_{t \in {\mathbb {N}}_0}\) is a dynamic conditional risk mapping, where \(\rho _0^{x}(Z) = R_{K}^{x}(Z)\) and \(\rho ^{x}_{t}(Z) = R_{K}^{x}(Z \vert {\mathcal {F}}_{t})\) for \(t \ge 1\), with
This family clearly includes the linear expectation (\(K = 0\), \(g_{0}(z,x) = z\)) and its statistical estimation properties are studied in Dentcheva et al. (2017).
Lemma 3.1
The family of dynamic conditional risk mappings \(\varrho =((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) defined through (16) and (17) is Markovian.
Proof
The Markov property holds at \(k = 0\) since \({\mathbb {P}}^x\)-a.s.
Assuming that it holds at \(k-1\), the Markov property also holds at k:
\(\square \)
3.1.1 Entropic risk mapping
The entropic risk mapping (a special case of a certainty equivalent risk mapping, see Föllmer and Schied 2016, Def. 2.36. or Bäuerle and Rieder 2017) is Markovian since it is recovered from (17) by taking \(K = 1\), \(g_{1}(z,r,x) = \frac{1}{\gamma (x)}\ln (r)\) (restricting the domain of \(r \mapsto g_{1}(z,r,x)\) to \((0,\infty )\)) and \(g_{0}(z,x) = e^{\gamma (x)z}\) in (17), where \(\gamma :E \rightarrow (0,\infty )\) is measurable and bounded away from both 0 and \(\infty \), giving
3.1.2 Mean-semideviation risk mapping
Similarly, the mean–semideviation risk mapping satisfies the Markov property since it is recovered from (17) by taking \(K = 2\), \(g_{2}(z,r,x) = z + \kappa (x) \, r^{\frac{1}{p}}\), \(g_{1}(z,r,x) = ((z - r)^{+})^{p}\) and \(g_{0}(z,x) = z\) in (17), where \(\kappa :E \rightarrow [0,1]\) is measurable and \(p \ge 1\) is an integer, giving
3.2 Worst-case risk mapping
The worst-case risk mapping is given by the family
For \(t \ge 1\) this is the \({\mathcal {F}}_{t}\)-conditional \({\mathbb {P}}^{x}\)-essential supremum of Z, that is, the smallest \({\mathcal {F}}_{t}\)-measurable random variable dominating Z almost surely with respect to \({\mathbb {P}}^{x}\)Artzner et al. (2007), pp. 20-21.
Lemma 3.2
The family of dynamic conditional risk mappings given by (18) is Markovian.
Proof
Supposing first that Z is non-negative, then using Barron et al. (2003), Prop. 2.12 and the Markov property of the conditional expectation, we have \({\mathbb {P}}^{x}\)-a.s.:
while the case \(t=0\) establishes measurability in x. For general \(Z \in b{\mathcal {F}}\) we first set \(Z_{c} :=Z + c\) with \(c = \sup _{\omega }\vert Z(\omega ) \vert \), then use translation invariance with respect to constants (see Barron et al. 2003, Prop. 2.1),
completing the proof. \(\square \)
3.3 Value at risk
The value at risk may be defined by the family
where \(\lambda \in (0,1)\),
and
for \(t \ge 1\), see e.g. Föllmer and Schied (2016), Sec. 4.4 & Ex. 11.4.
Lemma 3.3
The family of dynamic conditional risk mappings given by (19) is Markovian.
Proof
We first show that \(x \mapsto \rho ^{x}(Z)\) is measurable. For \(y \in {\mathbb {R}}\) we have
where for each \(m\in {\mathbb {R}}\) the function \(f_m :E \rightarrow {\mathbb {R}}\) is defined by \(f_m(x) = {\mathbb {P}}^x(m < Z)\). Note that each \(f_m\) is measurable because for every \(A \in {\mathcal {F}}\) the mapping \(E \ni x \mapsto {\mathbb {P}}^x(A)\in {\mathbb {R}}\) is measurable by the measurability of \(E \ni x \mapsto {\mathbb {P}}^x\in {\mathscr {P}}({\mathcal {F}})\). Thus \(\{ x \in E :\rho ^{x}(Z) < y \}\) is a measurable set.
To show that \(x \mapsto \rho ^x(Z)\) is bounded note that, since Z is bounded, there exists \(M\in {\mathbb {R}}\) such that \(\{\vert Z\vert <M\}=\Omega \). Thus \(-M \le \rho ^x(Z) \le M\) for all \(x \in E\).
Next we show that \(\rho ^{x}_{t}(Z \circ \theta _{t}) = \rho ^{X_{t}}(Z)\) almost surely. Since Z is bounded, let \(m_{t}(X_{0},\ldots ,X_{t}) \in b{\mathcal {F}}_{t}\) satisfy \({\mathbb {P}}^x\)-a.s.
Then for \({\mathbb {P}}^x\)-almost all \(\omega \in \Omega \), by conditional locality and the Markov property we have
giving \(m_{t}(\omega ) \ge \rho ^{X_{t}(\omega )}(Z)\). We conclude that \(\rho ^{x}_{t}(Z \circ \theta _{t}) \ge \rho ^{X_{t}}(Z)\) almost surely under \({\mathbb {P}}^{x}\).
Conversely we have by the Markov property that \({\mathbb {P}}^{x}\text {-a.s.}\),
and, since \(\omega \mapsto \rho ^{X_{t}(\omega )}(Z)\) is bounded and \({\mathcal {F}}_{t}\)-measurable, we conclude that \(\rho ^{x}_{t}(Z \circ \theta _{t}) \le \rho ^{X_{t}}(Z)\). \(\square \)
3.4 Average value at risk
For \(\lambda \in (0,1)\) the average value at risk (see Acciaio and Penner 2011, Ex. 1.10) may be defined by the following family of dynamic conditional risk mappings:
where
and
for \(t \ge 1\).
Lemma 3.4
The family of dynamic conditional risk mappings given by (20) is Markovian.
Proof
This follows from the Markov property for \({\text {VaR}}^{x}_{\lambda }(\cdot \vert {\mathcal {F}}_{t})\) (Lemma 3.3), since \({\mathbb {P}}^x\)-a.s.
\(\square \)
4 Dual representation of convex Markovian risk mappings
In this section we characterise the dual representation of convex Markovian risk mappings. Recalling from Sect. 2.1 that \((q^X(B\vert x) :B \in {\mathcal {E}}, x \in E)\) is the kernel associated to the Markov process X under \({\mathbb {P}}\), we begin with the necessary definitions:
Definition 4.1
-
(i)
\({\mathcal {R}} :E \times b{\mathcal {E}} \rightarrow {\mathbb {R}}\) is a transition risk mapping (cf. Çavuş and Ruszczyński 2014; Fan and Ruszczyński 2018a; Ruszczyński 2010) if:
-
for all \(f \in b{\mathcal {E}}\), \(x \mapsto {\mathcal {R}}(x,f)\) is bounded and measurable,
-
for all \(x \in E\), \(f \mapsto {\mathcal {R}}(x,f)\) satisfies
-
normalisation: \({\mathcal {R}}(x,0) = 0\),
-
monotonicity: \({\mathcal {R}}(x,f) \le {\mathcal {R}}(x,g)\) for all \(f \le g\),
-
constant translation invariance: \({\mathcal {R}}(x,f+c) = {\mathcal {R}}(x,f) + c\) for all constants c.
-
-
-
(ii)
A transition risk mapping is convex if for all \(x \in E\), \(f,g \in b{\mathcal {E}}\) and \(\lambda \in [0,1]\) we have
$$\begin{aligned} {\mathcal {R}}(x,\lambda f + (1-\lambda ) g) \le \lambda {\mathcal {R}}(x,f) + (1-\lambda ) {\mathcal {R}}(x,g). \end{aligned}$$
Note that by Definitions 2.3 and 2.5, a transition risk mapping can be derived from a regular collection of risk mappings \((\rho ^x)_{x \in E}\) by writing
If the transition risk mapping \({\mathcal {R}}\) defined by (21) is convex and continuous from below it has the following dual representation (cf. Föllmer and Penner 2006, Th. 2.3):
where the penalty functions \(\alpha ^x : {\mathscr {P}}({\mathcal {E}}) \rightarrow {\mathbb {R}}\) are defined by
Letting \({\mathcal {K}}\) denote the set of kernels q such that \(q(\cdot \vert x)\) is absolutely continuous with respect to \(q^X(\cdot \vert x)\) for every \(x \in E\), we have the following proposition:
Proposition 4.2
Let \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) be a family of dynamic conditional risk mappings such that \((\rho ^x)_{x \in E}\) is regular and each risk mapping \(\rho ^x\) is convex and continuous from below. Then \(\varrho \) satisfies the one-step Markov property (8) if and only if for all \(x \in E\) and \(f \in b{\mathcal {E}}\) we have
\({\mathbb {P}}^x\)-almost surely.
Proof
Using kernels, for all \(x \in E\) and \(f \in b{\mathcal {E}}\) the representation (22) can be rewritten as (23). Indeed, it is clear that the right-hand side of (23) is less than or equal to the right-hand side of (22). For the reverse inequality, simply note that for every given \(x \in E\) and every \(Q \in {\mathscr {P}}({\mathcal {E}})\) such that \(Q \ll {\mathbb {P}}^x \circ X_1^{-1}\), we can associate a kernel \(q \in {\mathcal {K}}\) by setting \(q( \cdot \vert x') = Q\mathbbm {1}_{\{x\}}(x') + ({\mathbb {P}}^{x'} \circ X_1^{-1})(1-\mathbbm {1}_{\{x\}}(x'))\). Then Equation (23) implies
which shows that (24) is equivalent to the one step Markov property (8). \(\square \)
Note that in (23) we take the supremum (rather than essential supremum) over a potentially uncountable family of kernels. Therefore the regularity of the collection \((\rho ^x)_{x \in E}\) follows from the assumptions of Proposition 4.2 rather than from (23).
Corollary 4.3
Let \(\varrho :=((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) be a time-consistent family of dynamic conditional risk mappings such that each risk mapping \(\rho ^x\) is convex and continuous from above. Then \(\varrho \) satisfies (23)–(24) (for all \(x \in E\) and \(f \in b{\mathcal {E}}\)) iff the Markov property of Definition 2.6 holds.
Proof
Combining Lemma 4.21 and Theorem 4.22 in Föllmer and Schied (2016), we see that a convex conditional risk mapping that is continuous from above is also continuous from below (recall the sign difference in our work). The corollary is then an application of Proposition 2.11 to Proposition 4.2. \(\square \)
The following example identifies a maximising kernel in (24) in the case of the entropic risk mapping of Sect. 3.1.1.
Example 4.4
(Entropic risk) For \(q \in {\mathcal {K}}\) let
Fixing a bounded, measurable function \(f:E \rightarrow {\mathbb {R}}\), define the kernel \(q_{op}\) by
It is well known (Detlefsen and Scandolo 2005, Rem. 9) from the non-Markovian setting that for each \(x \in E\) the function (25) is the minimal penalty corresponding to the entropic risk mapping on \(b{\mathcal {E}}\) and that (26) defines a measure attaining the maximum in (24). In order to show that \(q_{op}\) defined in this way is indeed a kernel we show that for each \(A \in {\mathcal {E}}\) the function \(x \mapsto q_{op}(A\vert x)\) is measurable. Indeed we have \(q_{op}(A\vert x) = \int _E \mathbbm {1}_A(y) \frac{e^{\gamma (x)f(y)}}{\int _E e^{\gamma (x)f(z)} \, q^X({\textrm{d}}z\vert x)} \, q^X({\textrm{d}}y\vert x)\) and measurability follows since more generally, for any jointly measurable function \(g:E \times E \rightarrow {\mathbb {R}}\) and any kernel p, the function \(x \mapsto \int _E g(x,y) \, p({\textrm{d}}y \vert x)\) is measurable.
5 Applications
The probabilistic Markov property can provide a convenient tool to address, for example, optimal stopping problems with costs which are measurable only after the chosen stopping time. A first example is the case of exercise lag, where we seek
where functions \(c, g:E \mapsto {\mathbb {R}}\) and a potentially unbounded stopping time \(\sigma \in {\mathscr {T}}\) represent respectively an observation cost, exercise cost and exercise lag, and \(\varrho =((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) is a time-consistent Markovian family of dynamic risk mappings. The strong Markov property of Proposition 2.9 then allows dynamic programming to be applied indirectly by first transforming the objective function. Indeed it then follows by the recursive property (6), conditional locality, conditional translation invariance, the identity \(X_{\sigma \circ \theta _\tau + \tau } = X_\sigma \circ \theta _\tau \) and the strong Markov property that
where \(h(x) :=\rho ^x(g(X_\sigma ))\), and standard dynamic programming arguments can then be applied to obtain the Wald-Bellman equations
In the optimal prediction problem of the next section, use of the probabilistic Markov property enables dynamic programming to instead be applied directly.
5.1 Optimal prediction
Generalising (3), let
where \(x \in E = {\mathbb {R}}\), \(\Omega = {\mathbb {R}}^{{\mathbb {N}}_0}\), \(T \in {\mathbb {N}}_0\), \(\varrho =((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) is a Markovian family of dynamic conditional risk mappings, \(X_T^* = \max _{0 \le s \le T} X_s\) and \(g:{\mathbb {R}} \rightarrow {\mathbb {R}}\) is bounded and measurable.
We extend this probability space to include the process’ running maximum by letting \({\tilde{\Omega }} = ({\mathbb {R}} \times {\mathbb {R}})^{{\mathbb {N}}_0}\). On this space, we have the canonical process \((X_t({\tilde{\omega }}),M_{t}({\tilde{\omega }})) = ({\tilde{\omega }}^{1}(t),{\tilde{\omega }}^{2}(t)) = {\tilde{\omega }}(t)\). Setting \(\tilde{{\mathbb {F}}} = (\tilde{{\mathcal {F}}}_{t})_{t \in {\mathbb {N}}_0}\) with \(\tilde{{\mathcal {F}}}_{t} = \sigma (\{(X_{s},M_{s}) :s \le t\})\) and \(\tilde{{\mathcal {F}}} = \sigma \left( \cup _{t} \tilde{{\mathcal {F}}}_{t}\right) \), there exists (see, for example, Çinlar 2011, Th. 4.4.18) a unique probability measure \(\tilde{{\mathbb {P}}}^{x,m}\) on \(({\tilde{\Omega }},\tilde{{\mathcal {F}}})\) such that (X, M) is a time-homogeneous Markov chain on \(({\tilde{\Omega }},\tilde{{\mathcal {F}}},\tilde{{\mathbb {F}}},\tilde{{\mathbb {P}}}^{x,m})\) with \(\tilde{{\mathbb {P}}}^{x,m}(X_{0} = x, M_{0} = m) = 1\) and transition kernel \(q^{X,M}\) satisfying \(q^{X,M}({\textrm{d}}x',{\textrm{d}}m' \vert x, m) = \delta _{m\vee x'}({\textrm{d}}m') \, q^{X}({\textrm{d}}x' \vert x)\) for all \((x, m) \in {\mathbb {R}}^2\). Note that \(\tilde{{\mathbb {P}}}^{x,m}(M_{n} = X_n^*\vee m) = 1\) and, in particular, for \(m=x\) we have \(\tilde{{\mathbb {P}}}^{x,x}(M_{n} = X_n^*) = 1\). Recalling Proposition 2.8, define a regular collection of risk mappings by
and let \(((\rho _t^{x,m})_{t \in {\mathbb {N}}_0})_{x,m \in {\mathbb {R}}}\) be the associated Markovian family of dynamic conditional risk mappings.
Theorem 5.1
If \(((\rho _t^x)_{t \in {\mathbb {N}}_0})_{x \in E}\) is time consistent then, for each bounded measurable function \(g:{\mathbb {R}} \rightarrow {\mathbb {R}}\), the extended value function
satisfies the following modified Wald-Bellman equations:
The optimal prediction problem (3) satisfies \(V_{\text {pred}}^n(x) = {\tilde{V}}^n(x,x)\).
Proof
Set
Following the outline of Peskir and Shiryaev (2006), Sec. 1.2, we may now proceed in five steps:
Step 1. We show that for all \(n = 0,1,\ldots ,T\) and \(k = T-n, T-n-1, \ldots , 0\), we have
One can easily see that the claim is true for \(k=T-n\). Further, by the Markov property and backward induction, for all n the random variable \(S_n^T\) is \(\sigma (M_n,X_n)\)-measurable (cf. Remark 2.16). All subsequent equalities hold \({\tilde{{\mathbb {P}}}}^{x,m}\)-almost surely. For any \(Z = {{\hat{Z}}} \circ {\tilde{\theta }}_k \in b\tilde{{\mathcal {F}}}_{k,\infty }\) we have
This and the induction hypothesis imply
Step 2. Let \(\tau _n^T := \inf \{k=n,\ldots , T : S_k^T = \rho _k^{x,m}(g(M_T-X_k)) \}\). We show that \(\tau _n^T = n + \tau _0^{T-n} \circ {\tilde{\theta }}_n\).
Indeed,
Step 3. We show that for \(n=T, \ldots , 0\) we have
Note that on \(\{\tau _{n-1}^T \ge n\}\) we have \(\tau _{n-1}^T = \tau _n^T\) (by definition of these stopping times). From this and time consistency we have
By the induction hypothesis
Note that
Thus, (28) implies that \(\rho _{n-1}^{x,m}(g(M_T-X_{\tau _{k-1}^T})) = S_{n-1}^T\).
Step 4. We prove that
We have
On the other hand, one can show by induction that for each \(k=T, \ldots , 0\) and every \(\tau \in {\mathscr {T}}_{[k,T]}\) we have \(\rho _k^{x,m}(g(M_T-X_\tau )) \ge S_k^T\). The claim is true for \(k=T\), and we may write
By the induction hypothesis, since \(\tau \vee k \in {\mathscr {T}}_{[k,T]}\) we have
In particular for \(k=0\) we conclude by Step 3 that for every \(T \in {\mathbb {N}}_0\), the stopping time \(\tau _0^T\) is optimal and \({\tilde{V}}^T(x,m)=\rho ^{x,m}(g(M_T-X_{\tau _0^T}))\). Combining this with (30) gives (29).
Step 5. We have by the previous step and the Markov property that
Taking \(n=0\) we get \({\tilde{V}}^{T}(x,m) = \rho ^{x,m}(g(M_T-x)) \wedge \rho ^{x,m}({\tilde{V}}^{T-1}(X_1,M_1))\), and the result follows by construction. \(\square \)
References
Acciaio B, Penner I (2011) Dynamic risk measures. In: Advanced mathematical methods for finance. Springer, Heidelberg, pp 1–34. https://doi.org/10.1007/978-3-642-18412-3_1
Allaart P (2010) A general ‘bang-bang’ principle for predicting the maximum of a random walk. J Appl Probab 47(4):1072–1083. https://doi.org/10.1017/s0021900200007373
Artzner P, Delbaen F, Eber JM et al (2007) Coherent multiperiod risk adjusted values and Bellman’s principle. Ann Oper Res 152:5–22. https://doi.org/10.1007/s10479-006-0132-6
Artzner P, Delbaen F, Eber JM, et al (1999) Coherent measures of risk. Math Finance 9(3). https://doi.org/10.1111/1467-9965.00068
Barron EN, Cardaliaguet P, Jensen R (2003) Conditional essential suprema with applications. Appl Math Optim 48(3):229–253. https://doi.org/10.1007/s00245-003-0776-4
Bartl D (2020) Conditional nonlinear expectations. Stochastic Process Appl 130(2):785–805. https://doi.org/10.1016/j.spa.2019.03.014
Bäuerle N, Rieder U (2014) More risk-sensitive Markov decision processes. Math Oper Res 39(1):105–120. https://doi.org/10.1287/moor.2013.0601
Bäuerle N, Rieder U (2017) Partially observable risk-sensitive Markov decision processes. Math Oper Res 42(4):1180–1196. https://doi.org/10.1287/moor.2016.0844
Blumenthal RM, Getoor RK (1968) Markov processes and potential theory, Pure and Applied Mathematics, vol 29. Academic Press, New York-London
Bogachev VI (2007) Measure Theory, vol I. Springer, Berlin
Çavuş O, Ruszczyński A (2014) Risk-averse control of undiscounted transient Markov models. SIAM J Control Optim 52(6):3935–3966. https://doi.org/10.1137/13093902X
Cheridito P, Delbaen F, Kupper M (2006) Dynamic monetary risk measures for bounded discrete-time processes. Electron J Probab 11(3):57–106. https://doi.org/10.1214/EJP.v11-302
Çinlar E (2011) Probability and Stochastics, Graduate Texts in Mathematics, vol 261. Springer, New York. https://doi.org/10.1007/978-0-387-87859-1
Cohen SN, Elliott RJ (2008) Solutions of backward stochastic differential equations on Markov chains. Commun Stoch Anal 2(2):251–262. https://doi.org/10.31390/cosa.2.2.05
Cohen SN, Elliott RJ (2010) A general theory of finite state backward stochastic difference equations. Stoch Process Appl 120(4):442–466. https://doi.org/10.1016/j.spa.2010.01.004
de Cooman G, Hermans F, Quaeghebeur E (2009) Imprecise Markov chains and their limit behavior. Probab Eng Inf Sci 23(4):597–635. https://doi.org/10.1017/S0269964809990039
Delbaen F (2002) Coherent risk measures on general probability spaces. Advances in finance and stochastics. Springer, Berlin, pp 1–37
Dellacherie C, Meyer PA (1978) Probabilities and Potential, North-Holland Mathematics Studies, vol 29. Hermann; North-Holland Publishing Co., Paris
Denk R, Kupper M, Nendel M (2018) Kolmogorov-type and general extension results for nonlinear expectations. Banach J Math Anal 12(3):515–540. https://doi.org/10.1215/17358787-2017-0024
Dentcheva D, Ruszczyński A (2020) Risk forms: representation, disintegration, and application to partially observable two-stage systems. Math Program 181(2, Ser. B):297–317. https://doi.org/10.1007/s10107-019-01376-1
Dentcheva D, Penev S, Ruszczyński A (2017) Statistical estimation of composite risk functionals and risk optimization problems. Ann Inst Stat Math 69(4):737–760. https://doi.org/10.1007/s10463-016-0559-8
Detlefsen K, Scandolo G (2005) Conditional and dynamic convex risk measures. Finance Stoch 9(4):539–561. https://doi.org/10.1007/s00780-005-0159-6
du Toit J, Peskir G (2007) The trap of complacency in predicting the maximum. Ann Probab 35(1):340–365. https://doi.org/10.1214/009117906000000638
Fan J, Ruszczyński A (2018) Process-based risk measures and risk-averse control of discrete-time systems. Math Program. https://doi.org/10.1007/s10107-018-1349-2
Fan J, Ruszczyński A (2018) Risk measurement and risk-averse control of partially observable discrete-time Markov systems. Math Methods Oper Res 88(2):161–184. https://doi.org/10.1007/s00186-018-0633-5
Föllmer H, Penner I (2006) Convex risk measures and the dynamics of their penalty functions. Stat Decisions 24(1):61–96. https://doi.org/10.1524/stnd.2006.24.1.61
Föllmer H, Schied A (2016) Stochastic Finance. An Introduction in Discrete Time. De Gruyter Graduate, De Gruyter, Berlin, fourth revised and extended edition
Frittelli F, Rosazza Gianin E (2002) Putting order in risk measures. J Bank Finance 26(7):1473–1486
Hartfiel DJ (1998) Markov Set-chains, Lecture Notes in Mathematics, vol 1695. Springer, Berlin
Krak T, De Bock J, Siebes A (2017) Imprecise continuous-time Markov chains. Int J Approx Reason 88:452–528. https://doi.org/10.1016/j.ijar.2017.06.012
Lebedev AA (1992) On monotone dominated sublinear functionals on the space of measurable functions. Sib Math J 33(6):1028–1038. https://doi.org/10.1007/BF00971026
Lebedev AA (1993) Disintegration of dominated monotone sublinear functionals on the space of measurable functions. Sib Math J 34(6):1117–1134. https://doi.org/10.1007/BF00973475
Martyr R, Moriarty J, Perninge M (2022) Discrete-time risk-aware optimal switching with non-adapted costs. Adv Appl Probab 54(2):625–655
Nendel M (2021) Markov chains under nonlinear expectation. Math Finance 31(1):474–507. https://doi.org/10.1111/mafi.12289
Pedersen JL (2003) Optimal prediction of the ultimate maximum of Brownian motion. Stoch Stoch Rep 75(4):205–219. https://doi.org/10.1080/1045112031000118994
Peng S (2005) Nonlinear expectations and nonlinear Markov chains. Chin Ann Math 26(02):159–184. https://doi.org/10.1142/S0252959905000154
Peskir G, Shiryaev AN (2006) Optimal Stopping and Free-Boundary Problems. Lectures in Mathematics. ETH Zürich, Birkhäuser Basel, Basel, https://doi.org/10.1007/978-3-7643-7390-0
Pichler A, Schlotter R (2020) Martingale characterizations of risk-averse stochastic optimization problems. Math Program 181(2, Ser. B):377–403. https://doi.org/10.1007/s10107-019-01391-2
Ruszczyński A (2010) Risk-averse dynamic programming for Markov decision processes. Math Program 125(2, Ser. B):235–261. https://doi.org/10.1007/s10107-010-0393-3
Shen Y, Stannat W, Obermayer K (2013) Risk-sensitive Markov control processes. SIAM J Control Optim 51(5):3652–3672. https://doi.org/10.1137/120899005
Yam SCP, Yung SP, Zhou W (2009) Two rationales behind the ‘buy-and-hold or sell-at-once’ strategy. J Appl Probab 46(3):651–668. https://doi.org/10.1239/jap/1253279844
Acknowledgements
The authors would like to thank the Isaac Newton Institute for Mathematical Sciences for support and hospitality during the programme ‘The Mathematics of Energy Systems’ when work on this paper was undertaken. This work was partially supported by EPSRC grant numbers EP/R014604/1, EP/N013492/1 and EP/P002625/1. This work was supported by the Lloyd’s Register Foundation-Alan Turing Institute programme on Data-Centric Engineering under the LRF grant G0095.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix A: Monotone class theorem
Appendix A: Monotone class theorem
For the reader’s convenience we state the monotone class theorem in the form given in Th. 2.12.9 Bogachev (2007).
Theorem A.1
Let \({\mathscr {H}}\) be a class of real functions on a set \(\Omega \) such that \(1 \in {\mathscr {H}}\) and let \({\mathscr {H}}_0\) be a subset in \({\mathscr {H}}\). Then, any of the following conditions yields that \({\mathscr {H}}\) contains all bounded functions measurable with respect to the \(\sigma \)-algebra generated by \({\mathscr {H}}_0\):
-
(i)
\({\mathscr {H}}\) is a closed linear subspace in the space of all bounded functions on \(\Omega \) with the norm \(\vert f \vert := \sup _\Omega \vert f(\omega ) \vert \) such that \(\lim _{n\rightarrow \infty } f_n \in {\mathscr {H}}\) for every increasing uniformly bounded sequence of nonnegative functions \(f_n \in {\mathscr {H}}\), and, in addition, \({\mathscr {H}}_0\) is closed with respect to multiplication (i.e., \(f g \in {\mathscr {H}}_0\) for all functions \(f, g \in {\mathscr {H}}_0\)).
-
(ii)
\({\mathscr {H}}\) is closed with respect to the formation of uniform limits and monotone limits and \({\mathscr {H}}_0\) is an algebra of functions (i.e., \(f + g\), cf, \(f g \in {\mathscr {H}}_0\) for all \(f, g \in {\mathscr {H}}_0\), \(c \in {\mathbb {R}}\)) and \(1 \in {\mathscr {H}}_0\).
-
(iii)
\({\mathscr {H}}\) is closed with respect to monotone limits and \({\mathscr {H}}_0\) is a linear space containing 1 such that \(\min (f, g) \in {\mathscr {H}}_0\) for all \(f, g \in {\mathscr {H}}_0\).
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Kosmala, T., Martyr, R. & Moriarty, J. Markov risk mappings and risk-sensitive optimal prediction. Math Meth Oper Res 97, 91–116 (2023). https://doi.org/10.1007/s00186-022-00802-z
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00186-022-00802-z