betaFeedback

WorksEconomics, Econometrics and Finance

Neural Network Approach to Portfolio Optimization with Leverage Constraints:a Case Study on High Inflation Investment

Abstract

Motivated by the current global high inflation scenario, we aim to discover a dynamic multi-period allocation strategy to optimally outperform a passive benchmark while adhering to a bounded leverage limit. To this end, we formulate an optimal control problem to outperform a benchmark portfolio throughout the investment horizon. Assuming the asset prices follow the jump-diffusion model during high inflation periods, we first establish a closed-form solution for the optimal strategy that outperforms a passive strategy under the cumulative quadratic tracking difference (CD) objective, assuming continuous trading and no bankruptcy. To obtain strategies under the bounded leverage constraint among other realistic constraints, we then propose a novel leverage-feasible neural network (LFNN) to represent control, which converts the original constrained optimization problem into an unconstrained optimization problem that is computationally feasible with standard optimization methods. We establish mathematically that the LFNN approximation can yield a solution that is arbitrarily close to the solution of the original optimal control problem with bounded leverage. We further apply the LFNN approach to a four-asset investment scenario with bootstrap resampled asset returns from the filtered high inflation regime data. The LFNN strategy is shown to consistently outperform the passive benchmark strategy by about 200 bps (median annualized return), with a greater than 90% probability of outperforming the benchmark at the end of the investment horizon.

Introduction

Since the global outbreak of COVID-19 in March 2020, there has been a significant increase in worldwide inflation. Specifically, from May 2021 to February 2023, the 12-month change in the CPI index in the U.S. has not dropped below 5% [19]. Prior to the pandemic, the U.S. economy experienced nearly four decades of low inflation. The abrupt shift from a long-term low-inflation environment to a high-inflation environment has created substantial uncertainty and volatility in the financial markets. In 2022, the technology-heavy NASDAQ stock index recorded a yearly return of -33.10% [61].

Equally concerning is the uncertainty around the duration of this round of high inflation. Some believe that the geopolitical tensions and the COVID-19 pandemic will overturn the trend of globalization and lead to global supply chain restructuring (Javorcik, 2020) which may result in a higher cost of production in the foreseeable future. Moreover, Ball et al. [8] suggests that the future inflation rate may remain high if the unemployment rate remains low.

In this article, we aim to answer the following question: with the goal of outperforming a passive benchmark, how should an active investor optimize the portfolio during high inflation? It is important to note that we do not attempt to make predictions about future inflation conditions. Instead, we approach the problem by formulating a multi-period optimal control problem that considers bounded leverage constraints and specific investment criteria. This optimal control problem requires the specification of an appropriate objective function, realistic constraints, as well as stochastic models for returns of traded assets during high inflation regimes. Given the complexities and challenges associated, it is crucial to develop an efficient method capable of computing optimal solutions, accommodating flexible data sources, handling high-dimensional cases, and dealing with complex constraints. In this paper, we propose a framework to address these challenges.

In § 2, we assume that the real (inflation-adjusted) asset returns during a high-inflation regime follow stochastic processes and treat allocation decisions as the control of a dynamic system. Specifically, we formulate an optimal control problem to outperform a fixed-mix benchmark portfolio consistently throughout the investment horizon by minimizing a cumulative quadratic tracking difference (CD) objective. There is a large amount of extant literature on closed-form solutions for beating a stochastic benchmark under synthetic market assumptions [16, 17, 82, 10, 29, 51, 65, 3, 1]. In these articles, the common objective function involves a log-utility function, e.g. the log wealth ratio. Under the log wealth ratio formulation, it is often hard to accommodate a fixed stream of cash injections, which is a common characteristic of open-ended funds. Forsyth et al. (2022) consider a scenario where a fixed amount of cash injections is allowed and provides a closed-form solution under a cumulative quadratic tracking difference (CD) objective given the assumption that the stock price follows a double exponential jump-diffusion model and the bond price is deterministic. Since the assumption that the bond index price is stochastic and has jumps is more reasonable under a high-inflation scenario, we develop a closed-form solution under the case that both the stock index and the bond index follow jump-diffusion models.

The closed-form solution is derived, unfortunately, under unrealistic assumptions such as continuous rebalancing, infinite leverage, and continued trading when insolvent. A discrete-time multi-period asset allocation problem is generally solved using a dynamic programming (DP) based approach, which converts a multi-step optimization problem into multiple single-step optimization problems. However, van Staden et al. [87] point out that dynamic programming-based approaches require the evaluation of a high-dimensional performance criterion to obtain the optimal control which is comparatively low-dimensional. This means that solving the discrete-time problem numerically using dynamic programming-based techniques (for example numerical solutions to the corresponding PIDE [89], or reinforcement learning (RL) techniques [32, 67, 54, 36]) are inefficient and are computationally prone to known issues such as error amplification over recursions [91].

Acknowledging these limitations, in § 2.7, we propose to use a single neural network model to approximate the optimal control and solve the original optimal control problem directly via a single standard finite-dimensional optimization. This direct approximation of the control exploits the lower dimensionality of the optimal control and bypasses the problem of solving high-dimensional conditional expectations associated with DP methods. We note that the idea of using a neural network to directly approximate the control process is also used in [38, 18, 84, 75], in which they propose a stacked neural network approach that includes individual sub-networks for each rebalancing step. In contrast, we propose a single shallow neural network that includes time as an input feature, and thus avoids the need to have multiple sub-networks for each rebalancing step and greatly reduces the computational and modeling complexity. Furthermore, using time as a feature in the neural network approximation function is consistent with the observation that (under assumptions) the optimal control is naturally a continuous function of time, which we discuss in detail in § 2.4.

The idea of using a single neural network to approximate controls has also been explored in previous studies such as Li and Forsyth (2019) and Ni et al. (2022). These studies focus on portfolio optimization problems with long-only constraints. The neural network architecture proposed in these studies transforms the constrained portfolio optimization problems into unconstrained optimization problems, making them computationally easier to solve. However, these existing neural network architectures do not address the bounded leverage constraint, which limits the total long exposure in the portfolio. The limited literature on portfolio optimization with bounded leverage is likely due to the added complexity arising from the combination of long and short positions in the portfolio.

A significant contribution of this article is the introduction of a novel leverage-feasible neural network (LFNN) model, which converts the leverage-constrained optimization problem into an unconstrained optimization problem. This model enables the incorporation of the bounded leverage constraint into the portfolio optimization framework. Additionally, in § 2.8, we provide a mathematical proof that, under reasonable assumptions, the solution of the unconstrained optimization problem obtained using the LFNN model can approximate the optimal control of the original problem arbitrarily well. This mathematical justification validates the effectiveness and validity of the LFNN approach.

In § 3, we present a case study on active portfolio optimization in a high-inflation regime. To identify historical high-inflation periods, we employ a simple filtering method. Subsequently, we use bootstrap resampling to generate training and testing data sets, which consist of price paths for four assets: the equal-weighted and cap-weighted stock indexes, as well as the 30-day and 10-year U.S. treasury indexes.

Using the leverage-feasible neural network (LFNN) model and the cumulative quadratic tracking shortfall (CS) objective, we derive a leverage-constrained strategy for portfolio optimization. Our results demonstrate that the LFNN model produces a strategy that consistently outperforms the fixed-mix benchmark. Specifically, the strategy achieves a median (annualized) internal rate of return (IRR) that is more than 2% higher than the benchmark. Moreover, there is a probability of over 90% that the strategy will yield a higher terminal wealth compared to the benchmark.

These findings highlight the efficacy of the LFNN model in optimizing portfolios under high-inflation conditions. By incorporating the bounded leverage constraint and utilizing the CS objective, our approach enables investors to achieve superior performance and mitigate risks in a high-inflation environment.

Our contributions are summarized below:

  1. To gain intuition about the behavior of the optimal controls, we derive the closed-form solution under a jump-diffusion asset price model and other typical assumptions (such as continuous rebalancing) for a two-asset case. The closed-form solution provides important insights into the properties of the optimal control as well as meaningful interpretations of the neural network models that approximate the controls.

  1. We propose to represent the control directly by a neural network representation so that the stochastic optimal control problem can be solved numerically under realistic constraints such as discrete rebalancing and limited leverage. Particularly, we propose the novel leverage-feasible neural network (LFNN) model to convert the original complex leverage-constrained optimization problem into an unconstrained optimization problem that can be solved easily by standard optimization methods.

  1. We prove that, with a suitable choice of the hyperparameter of the LFNN model, the solution of the parameterized unconstrained optimization problem can approximate the optimal control arbitrarily well. This provides a mathematical justification for the validity of the LFNN approach. This is further supported by the numerical results that the performance of the LFNN model matches the clipped form of the closed-form solution on simulated data.

  1. In the case study on active portfolio optimization in high-inflation, we apply the neural network method to bootstrap resampled asset returns with four underlying assets, including the equal-weighted/cap-weighted stock indexes, and the 30-day/10-year treasury bond indexes. The dynamic strategy from

the learned LFNN model outperforms the fixed-mix benchmark strategy consistently throughout the investment horizon, with a 2% higher median (annualized) internal rate of return (IRR), and more than 90% probability of achieving a higher terminal wealth. Furthermore, the learned allocation strategy suggests that the equal-weighted stock index and short-term bonds are preferable investment assets during high-inflation regimes.

Outperform dynamic benchmark under bounded leverage

Sovereign wealth funds and benchmark targets

Instead of taking a passive approach, some of the largest sovereign wealth funds often adopt an active management philosophy and use passive portfolios as the benchmark to evaluate the efficiency of active management. For example, the Canadian Pension Plan (CPP) uses a base reference portfolio of 85% global equity and 15% Canadian government bonds [27]. Another example is the Government Pension Fund Global of Norway (also known as the oil fund) managed by Norges Bank Investment Management (NBIM), which uses a benchmark index consisting of 70% equity index and 30% bond index.1 The benchmark equity index is constructed based on the market capitalization for equities in the countries included in the benchmark. The benchmark index for bonds specifies a defined allocation between government bonds and corporate bonds, with a weight of 70 percent to government bonds and 30 percent to corporate bonds [64].

However, the excess return that these well-known sovereign wealth funds have achieved over their respective passive benchmark portfolios cannot be described as impressive. In the 2022 fiscal year report, CPP claims to have beaten the base reference portfolio by an annualized 80 bps after fees over the past 5 years [27]. On the other hand, NBIM reports a mere average of 27 bps of annual excess return over the benchmark over the last decade (see Table 2.1). It is worth noting that these behemoth funds achieve seemingly meager results by hiring thousands of highly paid investment professionals and spending billions of dollars on day-to-day operations. For example, the CPP 2021 annual report [26] lists personnel costs as CAD 938 million, for 1,936 employees, which translates to average costs of about CAD 500,000 per employee-year.

Year2012201320142015201620172018201920202021Average
Excess return (%)0.210.99-0.770.450.150.70-0.300.230.270.740.27

Table 2.1. Norges Bank Investment Management, relative return to benchmark portfolio

The stark contrast between the enormous spending of sovereign wealth funds and the meager outperformance of the funds relative to the passive benchmark portfolios is probably provocative to taxpayers and pensioners who invest their hard-earned money in the funds. Equally concerning is the potential of a long, persistent inflation regime and the funds’ ability to consistently beat the benchmark portfolio in such times. After all, both the CPP Investments and NBIM were established in the late 1990s, a decade after the last long inflation period ended in the mid-1980s.

These concerns prompt us to ask the following question: in a presumed persistent high-inflation environment, can a fund manager find a simple asset allocation strategy that consistently beats the benchmark passive portfolios by a reasonable margin (preferably without spending billions of dollars in personnel costs)?

Mathematical formulation

In this section, we mathematically formulate the problem of outperforming a benchmark. Let [t0(=0),T][t_0(=0),T] denote the investment horizon, and let W(t)W(t) denote the wealth (value) of the portfolio actively managed by the manager at time t∈[t0,T]t \in[t_0,T]. We refer to the actively managed portfolio as the “active portfolio”. Furthermore, let W^(t)\widehat{W}(t) denote the wealth of the benchmark portfolio at time t∈[t0,T]t \in[t_0,T]. To ensure a fair assessment of the relative performance of the two portfolios, we assume both portfolios start with an equal initial wealth amount w0>0w_0 > 0, i.e., W(t0)=W^(t0)=w0>0W(t_0) = \widehat{W}(t_0) = w_0 > 0.

Technically, the admissible sets of underlying assets for the active and passive portfolio need not be identical. However, for simplicity, we assume that both the active portfolio and the benchmark portfolio can allocate wealth to the same set of NaN_a assets.

Let vector S(t)=(Si(t):i=1,⋯ ,Na)⊤∈RNa\mathbf{S}(t) = (S_i(t) : i = 1,\cdots,N_a)^\top\in\mathbb{R}^{N_a} denote the asset prices of the NaN_a underlying assets at time t∈[t0,T]t \in[t_0,T]. In addition, let vectors p(t)=(pi(t):i=1,⋯ ,Na)⊤∈RNa\mathbf{p}(t) = (p_i(t) : i = 1,\cdots,N_a)^\top\in\mathbb{R}^{N_a} and p^(t)=(p^i(t):i=1,⋯ ,Na)⊤∈RNa\widehat{\mathbf{p}}(t) = (\widehat{p}_i(t) : i = 1,\cdots,N_a)^\top\in\mathbb{R}^{N_a} denote the fraction of wealth allocated to the NaN_a underlying assets at time t∈[t0,T]t \in[t_0,T], respectively, for the active portfolio and the benchmark portfolio. From a control theory perspective, the allocation vector p\mathbf{p} can be regarded as the control of the system, as it determines how the wealth of the active portfolio evolves over time. We will seek to find the optimal feedback control. In other words, the closed-loop controls (allocation decisions) are assumed to also depend on the value of the state variables (e.g. portfolio wealth). Therefore, we consider the control p\mathbf{p} to be a function of time as well as the relevant state variables. In addition, the benchmark portfolio allocation p^\widehat{\mathbf{p}} can be regarded as a known function of its state variables and time as well.

Mathematically, p(X(t))=(pi(X(t)):i=1,⋯ ,Na)⊤∈RNa\mathbf{p}(\mathbf{X}(t)) = (p_i(\mathbf{X}(t)) : i = 1,\cdots,N_a)^\top\in\mathbb{R}^{N_a} and p^(X^(t))=(p^i(X^(t)):i=1,⋯ ,Na)⊤∈RNa\widehat{\mathbf{p}}(\widehat{\mathbf{X}}(t)) = (\widehat{p}_i(\widehat{\mathbf{X}}(t)) : i = 1,\cdots,N_a)^\top\in\mathbb{R}^{N_a}, where X(t)∈X⊆RNx\mathbf{X}(t) \in\mathcal{X} \subseteq\mathbb{R}^{N_x} and X^(t)∈X^⊆RN^x\widehat{\mathbf{X}}(t) \in\widehat{\mathcal{X}} \subseteq\mathbb{R}^{\widehat{N}_x} are the state variables taken into account by the active portfolio and the benchmark portfolio respectively. Here we include tt in X(t)\mathbf{X}(t) and X^(t)\widehat{\mathbf{X}}(t) for notational simplicity. In this article, we consider the particular problem of outperforming a passive portfolio, in which X(t)=(t,W(t),W^(t))⊤\mathbf{X}(t) = (t,W(t),\widehat{W}(t))^\top.

We assume that the active portfolio and the benchmark portfolio follow the same rebalancing schedule denoted by T⊆[t0,T]\mathcal{T} \subseteq[t_0,T]. In the case of discrete rebalancing, T⊂[t0,T]\mathcal{T} \subset[t_0,T] is a discrete set. In the case of continuous rebalancing, T=[t0,T]\mathcal{T} = [t_0,T], i.e., rebalancing happens continuously throughout the entire investment horizon.

Additionally, we assume both portfolios follow the same deterministic sequence of cash injections, defined by the set Tc={c(t), t∈Tc}\mathcal{T}_c = \{c(t),\ t \in\mathcal{T}_c\}, where Tc⊆[t0,T]\mathcal{T}_c \subseteq[t_0,T] is the schedule of the cash injections. When Tc\mathcal{T}_c is a discrete injection schedule, c(t)c(t) is the amount of cash injection at tt. In the case of continuous cash injections, i.e., Tc=[t0,T]\mathcal{T}_c = [t_0,T], c(t)c(t) is the rate of cash injection at tt, i.e., the total cash injection amount during [t,t+dt][t,t+dt] is c(t)dtc(t)dt, where dtdt is an infinitesimal time interval. For simplicity, we assume that Tc=T\mathcal{T}_c = \mathcal{T}, so that the cash injections schedule is the same as the rebalancing schedule. At t∈Tt \in\mathcal{T}, W(t)W(t) and W^(t)\widehat{W}(t) always denote the wealth after the cash injection (assuming there is a cash injection event happening at tt).

The active and benchmark strategies, respectively, are defined as the sequence of the allocation fractions following the rebalancing schedule. Mathematically, the active and benchmark strategies are defined by sets

P={p(X(t)), t∈T},andP^={p^(X^(t)), t∈T}.(1)\mathcal{P} = \{\mathbf{p}(\mathbf{X}(t)),\ t \in\mathcal{T}\},\quad\text{and}\quad\widehat{\mathcal{P}} = \{\widehat{\mathbf{p}}(\widehat{\mathbf{X}}(t)),\ t \in\mathcal{T}\}. \tag*{(1)}

Denote A\mathcal{A} as the set of admissible strategies, which reflects the investment constraints on the controls. We assume that admissibility can vary with state and let {Xi:i=1,⋯ ,k}\{\mathcal{X}_i: i = 1,\cdots,k\} be a partition of X\mathcal{X} (the state variable space), i.e.

{⋃i=1kXi=X,Xi∩Xj=∅, ∀ 1≤i<j≤k,(2)\begin{cases} \bigcup_{i=1}^{k}\mathcal{X}_i = \mathcal{X},\\ \mathcal{X}_i \cap\mathcal{X}_j = \varnothing,\ \forall\,1 \leq i < j \leq k, \tag*{(2)} \end{cases}

and {Zi⊆RNa:i=1,⋯ ,k}\{\mathcal{Z}_i \subseteq\mathbb{R}^{N_a}: i = 1,\cdots,k\} be the corresponding value sets of feasible controls such that any feasible control p\mathbf{p} satisfies

p(x)∈Zi,∀x∈Xi, ∀i∈{1,⋯ ,k}.(3)\mathbf{p}(\mathbf{x}) \in\mathcal{Z}_i,\forall\mathbf{x} \in\mathcal{X}_i,\ \forall i \in\{1,\cdots,k\}. \tag*{(3)}

We say that strategy P\mathcal{P} is an admissible strategy, i.e., P∈A\mathcal{P} \in\mathcal{A}, if and only if

P={p(X(t)), t∈T | p(X(t))∈Zi, if X(t)∈Xi}(4)\mathcal{P} = \left\{\mathbf{p}(\mathbf{X}(t)),\ t \in\mathcal{T}\ \middle|\ \mathbf{p}(\mathbf{X}(t)) \in\mathcal{Z}_i,\ \text{if }\mathbf{X}(t) \in\mathcal{X}_i\right\} \tag*{(4)}

Consider a discrete rebalancing schedule T={tj,j=0,⋯ ,N}\mathcal{T} = \{t_j, j = 0, \cdots, N\} with NN rebalancing events, where t0<t1<⋯<tN=Tt_0 < t_1 < \cdots< t_N = T.2 Then, the wealth evolution of the active portfolio and the benchmark portfolio can be described by the equations

References

  1. [1]Al-Aradi, A. and S. Jaimungal (2018a). Outperformance and tracking: dynamic asset allocation for active and passive portfolio management. Applied Mathematical Finance 25(3), 268–294.arxiv.org/abs/1803.05819
  2. [2]Al-Aradi, A. and S. Jaimungal (2018b). Outperformance and tracking: Dynamic asset allocation for active and passive portfolio management. Applied Mathematical Finance 25(3), 268–294.arxiv.org/abs/1803.05819
  3. [3]Alekseev, A. G. and M. V. Sokolov (2016). Benchmark-based evaluation of portfolio performance: a characterization. Annals of Finance 12, 409–440.DOI
  4. [4]Alizadeh, A. H. and N. K. Nomikos (2007). Investment timing and trading strategies in the sale and purchase market for ships. Transportation Research Part B: Methodological 41(1), 126–143.DOI
  5. [5]Anarkulova, A., S. Cederburg, and M. S. O’Doherty (2022). Stocks for the long run? evidence from a broad sample of developed markets. Journal of Financial Economics 143(1), 409–433.DOI
  6. [6]Atkinson, A. B. (1987). On the measurement of poverty. Econometrica 55(4), 749–764.DOI
  7. [7]Baillif, M. (2022). Collared and non-collared manifold boundaries. L’Enseignement Mathématique 68(1), 161–180.DOI
  8. [8]Ball, L. M., D. Leigh, and P. Mishra (2022). Understanding us inflation during the covid era. Technical report, National Bureau of Economic Research.
  9. [9]Banz, R. (1981). The relationship between return and market value of common stocks. Journal of Financial Economics 9:1, 3–18.DOI
  10. [10]Basak, S., A. Shapiro, and L. Tepla (2006). Risk management with benchmarking. Management Science 52(4), 542–557.
  11. [11]Black, F. (1993). Estimating expected return. Financial Analysts Journal 49(5), 36–38.
  12. [13]Boyde, E. (2021). ETFs: why tracking difference usually matters more than tracking error. https://www.ft.com/content/44b08a6c-55d7-4841-8e8e-aad6451a4cc3. [Online; accessed 6-Dec-2022].
  13. [14]Brigo, D., A. Dalessandro, M. Neugebauer, and F. Triki (2008). A stochastic processes toolkit for risk management. arXiv preprint arXiv:0812.4210.arxiv.org/abs/0812.4210
  14. [15]Brown, M. (1962). Locally flat imbeddings of topological manifolds. Annals of Mathematics, 331–341.DOI
  15. [16]Browne, S. (1999). Beating a moving target: optimal portfolio strategies for outperforming a stochastic benchmark. Finance and Stochastics 3, 275–294.
  16. [17]Browne, S. (2000). Risk-constrained dynamic active portfolio management. Management Science 46(9), 1188–1199.DOI
  17. [18]Buehler, H., L. Gonon, J. Teichmann, and B. Wood (2019). Deep hedging. Quantitative Finance 19:8, 1271–1291.
  18. [19]Bureau of Labor Statistics (2023). Consumer price index summary.
  19. [20]Cavaglia, S., L. Scott, K. Blay, and S. Hixon (2022). Multi-asset class factor premia: A strategic asset allocation perspective. The Journal of Portfolio Management 48(4), 14–32.DOI
  20. [21]Charteris, A. and K. McCullough (2020). Tracking error vs tracking difference: Does it matter? Investment Analysts Journal 49(3), 269–287.DOI
  21. [22]Coache, A. and S. Jaimungal (2021). Reinforcement learning with dynamic convex risk measures. arXiv preprint arXiv:2112.13414.arxiv.org/abs/2112.13414
  22. [23]Cogneau, P. and V. Zakamouline (2013). Block bootstrap methods and the choice of stocks for the long run. Quantitative Finance 13(9), 1443–1457.DOI
  23. [24]Connelly, R. (1971). A new proof of brown’s collaring theorem. Proceedings of the American Mathematical Society 27(1), 180–182.DOI
  24. [25]Cont, R., C. Mancini, et al. (2011). Nonparametric tests for pathwise properties of semimartingales. Bernoulli 17(2), 781–813.DOI
  25. [26]CPP Investments (2021). Annual Report 2021. https://www.cppinvestments.com/the-fund/our-performance/financial-results/. [Online; accessed 5-Dec-2022].
  26. [27]CPP Investments (2022). Annual Report 2022. https://www.cppinvestments.com/wp-content/uploads/2022/06/CPP-Investments_F2022-Annual-Report-EN.pdf. [Online; accessed 5-Dec-2022].
  27. [28]Dang, D.-M. and P. A. Forsyth (2016). Better than pre-commitment mean-variance portfolio allocation strategies: A semi-self-financing Hamilton–Jacobi–Bellman equation approach. European Journal of Operational Research 250(3), 827–841.DOI
  28. [29]Davis, M. and S. Lleo (2008). Risk-sensitive benchmarked asset management. Quantitative Finance 8(4), 415–426.DOI
  29. [32]Dixon, M. F., I. Halperin, and P. Bilokon (2020). Machine learning in Finance, Volume 1406. Springer.
  30. [33]Efron, B. (1992). Bootstrap methods: another look at the jackknife. In Breakthroughs in statistics, pp. 569–593. Springer.
  31. [34]Forsyth, P. A. (2020). Optimal dynamic asset allocation for dc plan accumulation/decumulation: Ambition-cvar. Insurance: Mathematics and Economics 93, 230–245.DOI
  32. [36]Gao, Z., Y. Gao, Y. Hu, Z. Jiang, and J. Su (2020). Application of deep q-network in portfolio management. In 2020 5th IEEE International Conference on Big Data Analytics (ICBDA), pp. 268–275. IEEE.
  33. [37]Goodfellow, I., Y. Bengio, and A. Courville (2016). Deep learning. MIT press.
  34. [38]Han, J. et al. (2016). Deep learning approximation for stochastic control problems. arXiv preprint arXiv:1611.07422.arxiv.org/abs/1611.07422
  35. [39]Homer, S. and R. E. Sylla (1996). A history of interest rates. Rutgers University Press.
  36. [40]Hornik, K. (1991). Approximation capabilities of multilayer feedforward networks. Neural networks 4(2), 251–257.DOI
  37. [41]Hougan, M. (2015). Tracking difference, the perfect etf metric. Retrieved 8(1), 2020.
  38. [43]Johnson, B., H. Bioy, A. Kellett, and L. Davidson (2013). On the right track: Measuring tracking efficiency in etfs. The Journal of Index Investing 4(3), 35–41.
  39. [44]Kingma, D. P. and J. Ba (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.arxiv.org/abs/1412.6980
  40. [45]Kou, S. G. (2002). A jump-diffusion model for option pricing. Management science 48(8), 1086–1101.DOI
  41. [46]Kou, S. G. and H. Wang (2004). Option pricing under a double exponential jump diffusion model. Management science 50(9), 1178–1192.DOI
  42. [47]Kratsios, A. and I. Bilokopytov (2020). Non-euclidean universal approximation. Advances in Neural Information Processing Systems 33, 10635–10646.arxiv.org/abs/2006.02341
  43. [48]Lahaye, J., S. Laurent, and C. J. Neely (2011). Jumps, cojumps and macro announcements. Journal of Applied Econometrics 26(6), 893–921.
  44. [49]Li, D. and W.-L. Ng (2000). Optimal dynamic portfolio selection: Multiperiod mean-variance formulation. Mathematical finance 10(3), 387–406.
  45. [51]Lim, A. and B. Wong (2010a). A benchmark approach to optimal asset allocation for insurers and pension funds. Insurance: Mathematics and Economics 46(2), 317–327.DOI
  46. [52]Lim, A. E. and B. Wong (2010b). A benchmarking approach to optimal asset allocation for insurers and pension funds. Insurance: Mathematics and Economics 46(2), 317–327.DOI
  47. [53]Lin, Y., R. D. MacMinn, and R. Tian (2015). De-risking defined benefit plans. Insurance: Mathematics and Economics 63, 52–65.DOI
  48. [54]Lucarelli, G. and M. Borrotti (2020). A deep q-learning portfolio management framework for the cryptocurrency market. Neural Computing and Applications 32(23), 17229–17244.
  49. [55]L’Her, J.-F., R. Stoyanova, K. Shaw, W. Scott, and C. Lai (2016). A bottom-up approach to the risk-adjusted performance of the buyout fund market. Financial Analysts Journal 72(4), 36–48.DOI
  50. [56]MacMinn, R., P. Brockett, J. Wang, Y. Lin, and R. Tian (2014). The securitization of longevity risk and its implications for retirement security. Recreating sustainable retirement, 134–160.DOI
  51. [57]Mancini, C. (2009). Non-parametric threshold estimation for models with stochastic diffusion coefficient and jumps. Scandinavian Journal of Statistics 36(2), 270–296.DOI
  52. [58]Merton, R. C. (1976). Option pricing when underlying stock returns are discontinuous. Journal of financial economics 3(1-2), 125–144.
  53. [59]Moody’s (2023a). Moody’s Seasoned Aaa Corporate Bond Yield [AAA], retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/AAA. [Online; accessed 18-Jan-2023].
  54. [60]Moody’s (2023b). Moody’s Seasoned Baa Corporate Bond Yield [BAA], retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/BAA. [Online; accessed 18-Jan-2023].
  55. [61]NASDAQ (2023). NASDAQ Composite Index. https://www.nasdaq.com/market-activity/index/comp/historical. [Online; accessed 15-Feb-2023].
  56. [63]Nicolosi, M., F. Angelini, and S. Herzel (2018). Portfolio management with benchmark related incentives under mean reverting processes. Annals of Operations Research 266, 373–394.DOI
  57. [64]Norges Bank (2022). Investment Strategy. https://www.nbim.no/en/the-fund/how-we-invest/investment-strategy/. [Online; accessed 5-Dec-2022].
  58. [65]Oderda, G. (2015). Stochastic portfolio theory optimization and the origin of rule based investing. Quantitative Finance 15(8), 1259–1266.DOI
  59. [66]Øksendal, B. K. and A. Sulem (2007). Applied stochastic control of jump diffusions, Volume 498. Springer.
  60. [67]Park, H., M. K. Sim, and D. G. Choi (2020). An intelligent financial portfolio trading strategy using deep q-learning. Expert Systems with Applications 158, 113573.arxiv.org/abs/1907.03665
  61. [68]Patton, A., D. N. Politis, and H. White (2009). Correction to “automatic block-length selection for the dependent bootstrap” by d. politis and h. white. Econometric Reviews 28(4), 372–375.DOI
  62. [69]Phalippou, L. (2014). Performance of buyout funds revisited? Review of Finance 18(1), 189–218.DOI
  63. [70]Plyakha, Y., R. Uppal, and G. Vilkov (2021). Equal or value weighting? implications for asset-pricing tests. In C. Zopounidis, R. Benkraiem, and I. Kalaïtzoglou (Eds.), Financial Risk Management and Modeling, pp. 295–347. Springer International Publishing.DOI
  64. [71]Politis, D. N. and J. P. Romano (1991). A circular block-resampling procedure for stationary data. Purdue University. Department of Statistics.
  65. [72]Politis, D. N. and J. P. Romano (1994). The stationary bootstrap. Journal of the American Statistical Association 89(428), 1303–1313.DOI
  66. [73]Politis, D. N. and H. White (2004). Automatic block-length selection for the dependent bootstrap. Econometric Reviews 23(1), 53–70.DOI
  67. [74]Powell, W. (2023). A universal framework for sequential decision problems. OR/MS Today February. https://tinyurl.com/PowellORMSfeature/.DOI
  68. [75]Reppen, A. M., H. M. Soner, and V. Tissot-Daguette (2022). Deep stochastic optimization in finance. arXiv preprint arXiv:2205.04604.arxiv.org/abs/2205.04604
  69. [76]Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206–215.DOI
  70. [77]Scott, L. and S. Cavaglia (2017). A wealth management perspective on factor premia and the value of downside protection. The Journal of Portfolio Management 43(3), 33–41.DOI
  71. [78]Shahzad, S. J. H., E. Bouri, D. Roubaud, L. Kristoufek, and B. Lucey (2019). Is bitcoin a better safe-haven investment than gold and commodities? International Review of Financial Analysis 63, 322–330.DOI
  72. [79]Silver, D., G. Lever, N. Heess, T. Degris, D. Wierstra, and M. Riedmiller (2014). Deterministic policy gradient algorithms. In International conference on machine learning, pp. 387–395. PmLR.
  73. [80]Simonian, J. and A. Martirosyan (2022). Sharpe parity redux. The Journal of Portfolio Management 48(4), 183–193.DOI
  74. [81]Taljaard, B. H. and E. Mare (2021). Why has the equal weight portfolio underperformed and what can we do about it? Quantitative Finance 21(11), 1855–1868.DOI
  75. [82]Tepla, L. (2001). Optimal investment with minimum performance constraints. Journal of Economic Dynamics and Control 25, 1629–1645.DOI
  76. [83]The Federal Reserve (2011). What is an acceptable level of inflation? https://www.federalreserve.gov/faqs/5D58E72F066A4DBDA80BBA659C55F774.htm. [Online; accessed 5-Dec-2022].
  77. [84]Tsang, K. H. and H. Y. Wong (2020). Deep-learning solution to portfolio selection with serially dependent returns. SIAM Journal on Financial Mathematics 11 (2), 593–619.
  78. [85]van Staden, P. M., D.-M. Dang, and P. A. Forsyth (2021). On the distribution of terminal wealth under dynamic mean-variance optimal investment strategies. SIAM Journal on Financial Mathematics 12 (2), 566–603.DOI
  79. [86]van Staden, P. M., P. A. Forsyth, and Y. Li (2022). Across-time risk-aware strategies for outperforming a benchmark. Working paper.DOI
  80. [87]van Staden, P. M., P. A. Forsyth, and Y. Li (2023). Beating a benchmark: dynamic programming may not be the right numerical approach. SIAM Journal on Financial Mathematics 14:2, 407–451.DOI
  81. [88]Vigna, E. (2014). On efficiency of mean–variance based portfolio selection in defined contribution pension schemes. Quantitative finance 14 (2), 237–258.
  82. [89]Wang, J. and P. A. Forsyth (2010). Numerical solution of the hamilton–jacobi–bellman formulation for continuous time mean variance asset allocation. Journal of Economic Dynamics and control 34 (2), 207–230.DOI
  83. [90]Wang, J. and P. A. Forsyth (2012). Comparison of mean variance like strategies for optimal asset allocation problems. International journal of theoretical and applied finance 15 (02), 1250014.DOI
  84. [91]Wang, R., D. P. Foster, and S. M. Kakade (2020). What are the statistical limits of offline rl with linear function approximation? arXiv preprint arXiv:2010.11895.arxiv.org/abs/2010.11895
  85. [92]Yao, D. D., S. Zhang, and X. Y. Zhou (2006). Tracking a financial benchmark using a few assets. Operations Research 54 (2), 232–246.DOI
  86. [93]Zhang, Q. and Y. Gao (2017). Portfolio selection based on a benchmark process with dynamic value-at-risk constraints. Journal of Computational and Applied Mathematics 313, 440–447.DOI
  87. [94]Zhao, Y. (2007). A dynamic model of active portfolio management with benchmark orientation. Journal of Banking & Finance 31 (11), 3336–3356.DOI
  88. [95]Zhou, X. Y. and D. Li (2000). Continuous-time mean-variance portfolio selection: A stochastic lq framework. Applied Mathematics and Optimization 42, 19–33.DOI
  89. [96]Ziemba, W. T. (2005). The symmetric downside-risk sharpe ratio. The Journal of Portfolio Management 32 (1), 108–122.

Paper details

vlnc:W_bszd2v5q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents