Time-consistent Source: en.wikipedia.org/wiki/Time-consistent

Time consistency in the context of finance is the property of not having mutually contradictory evaluations of risk at different points in time. This property implies that if investment A is considered riskier than B at some future time, then A will also be considered riskier than B at every prior time.

Time consistency is a property in financial risk related to dynamic risk measures. The purpose of the time-consistent property is to categorize the risk measures which satisfy the condition that if portfolio (A) is riskier than portfolio (B) at some time in the future, then it is guaranteed to be riskier at any time prior to that point. This is an important property since if it were not to hold then there is an event (with probability of occurring greater than 0) such that B is riskier than A at time ${\displaystyle t}$ although it is certain that A is riskier than B at time ${\displaystyle t+1}$. As the name suggests a time inconsistent risk measure can lead to inconsistent behavior in financial risk management.

This article may be too technical for most readers to understand. Please help improve it to make it understandable to non-experts, without removing the technical details. (February 2018) (Learn how and when to remove this message)

A dynamic risk measure ${\displaystyle \left(\rho _{t}\right)_{t=0}^{T}}$ on ${\displaystyle L^{0}({\mathcal {F}}_{T})}$ is time consistent if ${\displaystyle \forall X,Y\in L^{0}({\mathcal {F}}_{T})}$ and ${\displaystyle t\in \{0,1,...,T-1\}:\rho _{t+1}(X)\geq \rho _{t+1}(Y)}$ implies ${\displaystyle \rho _{t}(X)\geq \rho _{t}(Y)}$.^[1]

Equality

For all ${\displaystyle t\in \{0,1,...,T-1\}:\rho _{t+1}(X)=\rho _{t+1}(Y)\Rightarrow \rho _{t}(X)=\rho _{t}(Y)}$

Recursive

For all ${\displaystyle t\in \{0,1,...,T-1\}:\rho _{t}(X)=\rho _{t}(-\rho _{t+1}(X))}$

Acceptance Set

For all ${\displaystyle t\in \{0,1,...,T-1\}:A_{t}=A_{t,t+1}+A_{t+1}}$ where ${\displaystyle A_{t}}$ is the time ${\displaystyle t}$ acceptance set and ${\displaystyle A_{t,t+1}=A_{t}\cap L^{p}({\mathcal {F}}_{t+1})}$^[2]

Cocycle condition (for convex risk measures)

For all ${\displaystyle t\in \{0,1,...,T-1\}:\alpha _{t}(Q)=\alpha _{t,t+1}(Q)+\mathbb {E} ^{Q}[\alpha _{t+1}(Q)\mid {\mathcal {F}}_{t}]}$ where ${\displaystyle \alpha _{t}(Q)=\operatorname {*} {esssup}_{X\in A_{t}}\mathbb {E} ^{Q}[-X\mid {\mathcal {F}}_{t}]}$ is the minimal penalty function (where ${\displaystyle A_{t}}$ is an acceptance set and ${\displaystyle \operatorname {*} {esssup}}$ denotes the essential supremum) at time ${\displaystyle t}$ and ${\displaystyle \alpha _{t,t+1}(Q)=\operatorname {*} {esssup}_{X\in A_{t,t+1}}\mathbb {E} ^{Q}[-X\mid {\mathcal {F}}_{t}]}$.^[3]

Due to the recursive property it is simple to construct a time consistent risk measure. This is done by composing one-period measures over time. This would mean that:

• ${\displaystyle \rho _{T-1}^{com}:=\rho _{T-1}}$
• ${\displaystyle \forall t<T-1:\rho _{t}^{com}:=\rho _{t}(-\rho _{t+1}^{com})}$^[1]

Both dynamic value at risk and dynamic average value at risk are not a time consistent risk measures.

The time consistent alternative to the dynamic average value at risk with parameter ${\displaystyle \alpha _{t}}$ at time t is defined by

${\displaystyle \rho _{t}(X)={\text{ess}}\sup _{Q\in {\mathcal {Q}}}E^{Q}[-X|{\mathcal {F}}_{t}]}$

such that ${\displaystyle {\mathcal {Q}}=\left\{Q\in {\mathcal {M}}_{1}:E\left[{\frac {dQ}{dP}}|{\mathcal {F}}_{j}\right]\leq \alpha _{j-1}E\left[{\frac {dQ}{dP}}|{\mathcal {F}}_{j-1}\right]\forall j=1,...,T\right\}}$.^[4]

The dynamic superhedging price is a time consistent risk measure.^[5]

The dynamic entropic risk measure is a time consistent risk measure if the risk aversion parameter is constant.^[5]

In continuous time, a time consistent coherent risk measure can be given by:

${\displaystyle \rho _{g}(X):=\mathbb {E} ^{g}[-X]}$

for a sublinear choice of function ${\displaystyle g}$ where ${\displaystyle \mathbb {E} ^{g}}$ denotes a g-expectation. If the function ${\displaystyle g}$ is convex, then the corresponding risk measure is convex.^[6]