Abstract
Compared with the extensive empirical literature on contrarian strategy, we develop a dynamic mean-variance model with geometric value-reversion asset prices, which implies a contrarian strategy. The model is solved (semi) explicitly under three asset price evaluations: constant valuation, exponential-varying valuation, and geometric average valuation. From a mathematical perspective, it is nontrivial to solve the extended HJB equations under stochastic opportunities. We demonstrate that our strategy exhibits the same monotonicity as that of the traditional constant relative risk-averse utility, and the welfare loss of using the dynamic mean-variance criterion is rather small, supporting that our model is a good approximation to the constant relative risk-averse utility. Empirical tests show that our strategies can help an investor achieve a less volatile wealth trajectory.
Similar content being viewed by others
Availability of Data and Material
All original datasets used come from public websites.
Notes
For a function F of \(\left( t,P,R\right) \), the infinitesimal operator is \(\mathcal {D}F\left( t,P,R\right) =\frac{\partial F}{\partial t}+( \mu _{t}+\lambda ( \alpha _{t}-\ln P)) P\frac{\partial F}{\partial P}+\left[ r_{t}+\left( \mu _{t}-r_{t}+\lambda \left( \alpha _{t}-\ln P\right) \right) \pi _{t}-\frac{1}{2}\sigma _{t}^{2}\pi _{t}^{2}\right] \frac{\partial F}{\partial R}+\frac{1}{2}\left[ \sigma _{t}^{2}P^{2}\frac{\partial ^{2} F}{\partial P^{2}}+2\sigma _{t}^{2}\pi _{t}P\frac{\partial ^{2}F}{\partial P\partial R}\right. +\left. \sigma _{t}^{2}\pi _{t}^{2}\frac{\partial ^{2}F}{\partial R^{2} }\right] \). \(\mathcal {D}^{*}\) is the infinitesimal operator when replacing \(\pi \) by the equilibrium \(\hat{\pi }\) in \(\mathcal {D}\).
By the structure of \(R_{T}\), it is easy to see that the expectation and variance, hence g and V, are separable with the starting return \(R_{t}\).
For a function F of \(\left( t,P,\bar{P},R\right) \), the infinitesimal operator is \(\mathcal {D}F\left( t,P,\bar{P},R\right) =\frac{\partial F}{\partial t}+\left( \mu +\lambda \right. \left. (\ln \bar{P}-\ln P) \right) P\frac{\partial F}{\partial P}+\bar{P}\left[ \frac{\ln P-\ln \bar{P}}{t}\right] \frac{\partial F}{\partial \bar{P}}+\left[ r+\left( \mu -r+\lambda \left( \ln \bar{P}-\ln P\right) \right) \pi -\frac{1}{2}\sigma ^{2}\pi ^{2}\right] \frac{\partial F}{\partial R}+\frac{1}{2}\left[ \sigma ^{2}P^{2}\right. \left. \frac{\partial ^{2}F}{\partial P^{2}}+2\sigma ^{2}\pi P\frac{\partial ^{2}F}{\partial P\partial R}+\sigma ^{2}\pi ^{2}\frac{\partial ^{2}F}{\partial R^{2}}\right] \). \(\mathcal {D}^{*}\) is the infinitesimal operator when replacing \(\pi \) by the optimal \(\hat{\pi }\) in \(\mathcal {D}\).
References
Basak S, Chabakauri G (2010) Dynamic mean-variance asset allocation. Rev Financ Stud 23:2970–3016
Basu P, Vinod HD (1994) Mean reversion in stock prices: implications from a production based asset pricing model. Scand J Econ 96(1):51–65
Benth FE, Karlsen KH (2005) A note on Merton’s portfolio selection problem for the Schwartz mean-reversion model. Stoch Anal Appl 23(4):687–704
Björk T, Murgoci A, Zhou XY (2014) Mean-variance portfolio optimization with state-dependent risk aversion. Math Financ 24(1):1–24
Björk T, Khapko M, Murgoci A (2017) On time-inconsistent stochastic control in continuous time. Finance Stochast 21(2):331–360
Chan KC (1988) On the contrarian investment strategy. J Bus 61(2):147–163
Chang RP, McLeavey DW, Rhee SG (1995) Short-term abnormal returns of the contrarian strategy in the Japanese stock market. J Bus Financ Acc 22(7):1035–1048
Chen Q, Hua X, Jiang Y (2018) Contrarian strategy and herding behaviour in the Chinese stock market. Eur J Financ 24(16):1552–1568
Chen J, Davison M, Escobar M, Zafari G (2021) Robust portfolios with commodities and stochastic interest rates. Quant Finance 21(6):991–1010
Cheng S, Hameed A, Subrahmanyam A, Titman S (2017) Short-term reversals: the effects of past returns and institutional exits. J Financ Quant Anal 52(1):143–173
Chiu M, Wong H (2014) Mean-variance asset-liability management with asset correlation risk and insurance liabilities. Insur Math Econ 59:300–310
Dai M, Jin H, Kou S, Xu YH (2021) A dynamic mean-variance analysis for log returns. Manage Sci 67(2):1093–1108
De Bondt WF, Thaler R (1985) Does the stock market overreact? J Financ 40(3):793–805
De Bondt WF, Thaler R (1987) Further evidence on investor overreaction and stock market seasonality. J Financ 42(3):557–581
Jegadeesh N (1990) Evidence of predictable behavior of security returns. J Financ 45(3):881–898
Kim H (2009) On the usefulness of the contrarian strategy across national stock markets: a grid bootstrap analysis. J Empir Financ 16(5):734–744
Lehmann BN (1990) Fads, martingales, and market efficiency. Q J Econ 105(1):1–28
Li D, Ng WL (2000) Optimal dynamic portfolio selection: multiperiod mean-variance formulation. Math Financ 10:387–406
Markowitz H (1952) Portfolio selection. J Financ 7(1):77–91
Otchere I, Chan J (2003) Short-term overreaction in the Hong Kong stock market: can a contrarian trading strategy beat the market? J Behav Financ 4(3):157–171
Poterba JM, Summers LH (1988) Mean reversion in stock prices: evidence and implications. J Financ Econ 22(1):27–59
Schwartz ES (1997) The stochastic behavior of commodity prices: implications for valuation and hedging. J Financ 52(3):923–973
Zaitsev VF, Polyanin AD (2002) Handbook of exact solutions for ordinary differential equations chapter 2, table 14 and table 15. CRC Press
Zhou XY, Li D (2000) Continuous-time mean-variance portfolio selection: a stochastic LQ framework. Appl Math Optim 42:19–53
Funding
Xu is supported by the Natural Science Foundation of China (No.12271391; No.11871050) and Tang Scholar Fund.
Author information
Authors and Affiliations
Contributions
The authors are listed in alphabetical order. L.Z. performed the numerical solutions for Sections 4.2 and 4.3, P.P. derived all the mathematical results, X.Y. contributed to writing, reviewing, and providing guidance throughout the paper, and Z.W. prepared the figures for Sections 4.1 and 4.4. All authors reviewed the manuscript.
Corresponding author
Ethics declarations
Competing Interests
The authors declare no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix 1 Proof of Theorem 2.1
(i) We assume that V and g are seperableFootnote 2 in R, i.e., they have the following structures:
Substituting (A.1) and (A.2) into PDE (2.6) and PDE (2.7), we obtain:
subject to \(\tilde{V}(T,P)=0\), \(f(T,P)=0\). By the first-order condition, we derive a representation of the equilibrium \(\hat{\pi }\) as (2.8). It follows putting \(\hat{\pi }\) into f that, the function f satisfies PDE (2.9), subject to terminal condition \(f\left( T,P\right) =0\). Then, following the line of Björk et al. (2017), it is easy to verify that the strategy (2.10) is really an equilibrium of the original problem.
(ii) First we make a transform on the variable \(x=\ln P\). Then PDE (2.9) becomes
Let \(f\left( t,x\right) =-\frac{\left( 1+\gamma \right) ^{2}}{\gamma ^{2} }\ln u\left( t,x\right) \). Then PDE (A.3) is rewritten as
Let \(\omega =\alpha +\frac{\mu -r}{\lambda }-x\), \(\tilde{u}(t,\omega )=u(t,x).\) Putting \(\frac{\partial u}{\partial x}=-\frac{\partial \tilde{u}}{\partial \omega }\), \(\frac{\partial ^{2}u}{\partial x^{2}}=\frac{\partial ^{2}\tilde{u}}{\partial \omega ^{2}}\) into PDE (A.4), we get
Set \(\tilde{u}=\exp \left\{ k(T-t)+\eta \omega +\theta \omega ^{2}\right\} V(t,\omega )\). Substituting
into the above PDE, we obtain
Let the coefficient of V be 0 by separating the variable \(\omega \), we get the following algebraic equations:
where
From the last and the second equations, we deduce that
where \(q=\sqrt{b^{2}+2\sigma ^{2}c}=\frac{\lambda \sqrt{1+2\gamma }}{(1+\gamma )}.\) Thus (A.5) satisfies
Then we denote \(\hat{a}:=a+\sigma ^{2}\eta =\mp \frac{ab}{q}\), \(\hat{b}:=2\sigma ^{2}\theta -b=\pm q\). So (A.9) is rewritten as
From the terminal structure of \(V\left( T,\omega \right) ,\) it is natural to suppose \(V\left( t,\omega \right) =\exp \{A\left( t\right) +B\left( t\right) \omega +C\left( t\right) \omega ^{2}\}\). Substituting \(V_{t}=\left( A_{t}+B_{t}\omega +C_{t}\omega ^{2}\right) \cdot V\), \(V_{\omega }=\left( B+2C\omega \right) \cdot V\), \(V_{\omega \omega }=\left( B^{2}+2C+4BC\omega +4C^{2}\omega ^{2}\right) \cdot V\) into (A.10) and by separating variables, we obtain the following system of ordinary differential equations (ODEs):
Solving the last equation, we have
Then
Denote \(\bar{B}=\eta +B\). Then it follows (A.11) that,
Note that the coefficient of \(\bar{B}\) and the nonhomogeneous term are functions of time t, it is not easy to get the solution of ODE (A.12). Since (A.12) is a linear equation, we can get the solution first in integral form, then by a quite daunting and tedious calculation, the explicit solution is obtained as follows:
From the deformations of the above series of formulas, it yields
Then the equilibrium strategy (2.8) can be expressed explicitly by
We then rewrite the above \(\hat{\pi }(t,P)\) as (2.10).
The proof is complete. \(\square \)
Appendix 2 Proof of Corollary 2.1
(i) From the expression of \(\hat{\pi }\), it’s not difficult to find that \(\hat{\pi }\) decreases as \(\ln P\) increases.
(ii) When \(\ln P<\alpha +\frac{\mu -r}{\lambda }\) and \(r<\frac{\sigma ^{2}}{2}\), we find that \(\frac{\lambda \gamma (1+2\gamma )\left( \mu -r+\lambda (\alpha -\ln P)\right) }{(1+\gamma )^{3}\sigma ^{2}}>0\) and \(\frac{\lambda ^{2} \gamma (1+2\gamma )(\sigma ^{2}-2r)}{2(1+\gamma )^{3}\sigma ^{2}}>0\). On the contrary, when \(\ln P>\alpha +\frac{\mu -r}{\lambda }\) and \(r>\frac{\sigma ^{2} }{2}\), it is easy to get that \(\frac{\lambda \gamma (1+2\gamma )\left( \mu -r+\lambda (\alpha -\ln P)\right) }{(1+\gamma )^{3}\sigma ^{2}}<0\) and \(\frac{\lambda ^{2}\gamma (1+2\gamma )(\sigma ^{2}-2r)}{2(1+\gamma )^{3}\sigma ^{2} }<0\). At the same time, \(\hat{\Theta }(\tau )\) is increasing in the time horizon \(\tau .\) Moreover,
is also increasing in the time horizon \(\tau \). Consequently, we obtain the desired results. \(\square \)
Appendix 3 Proof of Corollary 2.2
(i) In order to calculate the limit of \(\hat{\pi }(t,P)\) when \(\tau \) tends to infinity, we first calculate the limits of \(\hat{\Theta } (\tau )\) and \(\hat{B}\left( \tau \right) \).
Note that
Therefore,
(ii) Denote
Hence,
and
To compute the hedging ratio \(\frac{I}{I+1}\), we first calculate the limit of I
Thus,
The proof is complete. \(\square \)
Appendix 4 Proof of Proposition 4.1
Benth and Karlsen (2005) provide a semi-explicit solution for the CRRA optimization with geometric mean reversion and a risk aversion coefficient \(\tilde{\gamma }\in \left( 0,1\right) \). In our paper, \(\tilde{\gamma }=\gamma +1\) should be greater than 1 because \(\gamma \ge 0\) in the MV model. Here, we outline the proof and provide an explicit solution. Utilizing the classical dynamic programming principle, we derive the HJB equation for the value function:
subject to \(v(T,W,P)=\frac{W^{1-\tilde{\gamma }}-1}{1-\tilde{\gamma }}.\)
Assuming there exists a sufficiently smooth solution v to the HJB equation, the first-order condition for an optimal strategy gives
Then substituting \(\pi ^{*}\) into (D.1), we get the following PDE:
As in Benth and Karlsen (2005), we make the assumption that the solution of (D.2) takes the form
Then from (D.2) we get
which indeed is a one-dimensional linear parabolic equation. Similar to the baseline model in Section 2, we assume \(g(t,P)=\exp \left( f_{0}(t)+f_{1}(t)\ln P+f_{2}(t)(\ln P)^{2}\right) .\) Then the optimal control can be rewritten as:
where \(f_{1},\) \(f_{2}\) satisfy the following system of ODEs:
Through a series of calculations, we get
where \(\hat{\Theta }\left( t\right) ,\) \(\hat{B}\left( t\right) \) are defined by (2.12), (2.13) with \(b=\frac{\lambda }{\tilde{\gamma }} \), \(q=\frac{\lambda }{\sqrt{\tilde{\gamma }}}\). \(\square \)
Appendix 5 Proof of Theorem 3.1
Let \(\alpha _{t}:=\ln \bar{P}_{t}=\ln \bar{P}_{0}+\beta t\). Similar to the baseline model in Section 2, we can get an equilibrium strategy by
where the function f satisfies the following PDE:
First we make a transform on the variable \(x=\ln P-\ln \bar{P}_{0}\). Then PDE (E.2) adapts to
Let f(x) \(=-\frac{\left( 1+\gamma \right) ^{2}}{\gamma ^{2}}\ln u\left( t,x\right) \). Then PDE (E.3) satisfies
Let \(\omega =\frac{\mu -r}{\lambda }-x\), \(\tilde{u}(t,\omega )=u(t,x).\) Then it yields
Set \(\tilde{u}=\exp \left\{ k(T-t)+\eta (\omega +\beta t)+\theta (\omega +\beta t)^{2}\right\} V(t,\omega )\). We obtain
Let the coefficient of V be 0 by separating the variable \((\omega +\beta t)\). Then we get the following ODE systems:
Set
Then we get the ODE system the same as (A.6). Hence \(\theta \), \(\eta \) have the same forms with (A.8). Consequently, (E.5) is reduced to
Then we denote \(\hat{a}:=a+\sigma ^{2}\eta =\mp \frac{ab}{q}\), \(\hat{b}:=2\sigma ^{2}\theta -b=\pm q\). Hence the above PDE becomes
We assume V has the following structure:
Substituting it into (E.6), and by separating variables, we get the same ODE system as (A.11). By solving the ODE system, we get an equilibrium strategy. The proof is complete. \(\square \)
Appendix 6 Proof of Theorem 3.2
By the general extended HJB in Björk et al. (2017), we have,
subject to \(V(T,P,\bar{P},R)=R\), \(g(T,P,\bar{P},R)=R\), where \(\mathcal {D},\mathcal {D}^{*}\) are the infinitesimal operatorsFootnote 3. Now we assume that V and g are separable in R, i.e., they have the following structures:
Substituting (F.3) and (F.4) into PDE (F.1) and PDE (F.2), it yields that
subject to \(\tilde{V}(T,P,\bar{P})=0\), \(f(T,P,\bar{P})=0\).
Similar to the baseline model in Section 2, by the first-order condition, we get an equilibrium strategy by
where the function f satisfies the following PDE
subject to terminal condition \(f\left( T,\bar{P},P\right) =0\).
Let \(f\left( t,\bar{P},P\right) =-\frac{\left( 1+\gamma \right) ^{2}}{\gamma ^{2}}\ln u\left( t,\bar{P},P\right) \). Then PDE (F.5) becomes
with terminal condition \(u\left( T,\bar{P},P\right) =1\). Let \(x=\ln \bar{P}-\ln P\), \(\tilde{u}(t,x):=u(t,\bar{P},P)\). Then we deduce
As in the baseline model, we set \(\tilde{u}(t,x)=\exp (A(t)+B(t)x+C(t)x^{2}).\) Substituting it into PDE (F.6) and by separating variables, we obtain
where we denote \(\hat{a}=-(\mu -\frac{\gamma ^{2}(\mu -r)}{(1+\gamma )^{2}} -\frac{1}{2}\sigma ^{2})\), \(\hat{b}=-\frac{\lambda (1+2\gamma )}{(1+\gamma )^{2}}\).
From the above transformations, it’s easy to get that
where B(t) and C(t) satisfy (F.7). Therefore,
The proof is complete. \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Lu, Z., Pang, P., Xu, Y. et al. Portfolio Selection with Contrarian Strategy. Methodol Comput Appl Probab 26, 16 (2024). https://doi.org/10.1007/s11009-024-10085-y
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s11009-024-10085-y