【论文精读】A New Predictive Measure Using Agent-Based Behavioral Finance

A New Predictive Measure Using Agent-Based Behavioral Finance

Todd Feldman · Shuming Liu

这篇论文2018年发表在Computational Economics,据摘要介绍其主要是对 Friedman, Abraham, 2009年发表 Journal of Economic Dynamics and Control上的论文Bubbles and crashes: Gradient dynamics in financial markets中所构建模型的calibration,所以我对两篇论文都进行了研读。第一天我读的是关于Agent-Based Model and Simulation的tutorial式的介绍,感觉还是有点虚,所以想结合实际应用看看ABS模型究竟是如何使用的。基于我之前对Quantitative Finance的研究,以及对行为金融学的兴趣,所以想通过这篇论文对ABMS一探究竟。

A New Predictive Measure Using Agent-Based Behavioral Finance这篇论文主要对Friedman, Abraham,的模型(一下简称FA模型)进行参数校验(calibration)以及研究异质性信念(heterogeneity in beliefs )如何影响股票市场的波动和收益。

simulation model的工作流程如下:

  1. every agent or portfolio manager is randomly assigned a portfolio size and an allocation to a risky asset at the start of the simulation 每位agent,即基金经理设置任意的组合头寸大小以及风险资产的配置比例
  2. the portfolio managers update their allocations to the risky asset based on the historical market return, their current idiosyncratic alpha, and a measure of risk from behavioral biases 基金经理根据历史市场收益,异质性的alpha以及行为偏差更新风险资产的配置比例。
  3. simulation program averages each manager’s allocation to the risky asset across all managers weighting by the portfolio size. This measure forms the risky asset’s average market demand. 求出所有基金经理风险资产头寸的平均值,用以衡量风险资产的平均市场需求。
  4. a change in the average market demand leads to a change in the risky asset’s price and a change in the price leads to a change in the average market demand. 风险资产的平均市场需求的变化导致风险资产的价格变动,风险资产的价格变动亦会导致风险资产的平均市场需求的变化。这个feedback loop会持续直到simulation结束。

论文将美国上千位权益性共同基金经理的数据放入FA模型中,结果显示,FA预测的股票价格序列与标普500序列的波峰和波谷类似,二阶矩相似,分布相同。

在异质性信念的研究中,作者使用FA模型calibrate的头寸大小以及头寸波动的结果作为他们异质性信念的代理变量。结果显示,投资者异质性信念的差异越大,未来股票市场decline的可能性越大。信念头寸大小的波动性与股票市场的波动性有着显著的正相关关系。

FA Simulation Model

t h e   p r i c e   o f   t h e   r i s k y   a s s e t   风 险 资 产 的 价 格    P t = 1 R s − g u ˉ t δ g : t h e   g r o w t h   r a t e , R s t h e   d i s c o u n t   r a t e 1 R s − g :   t h e   p r e s e n t   v a l u e   o f   f u t u r e   d i v i d e n t s u ˉ t : t h e   m e a n   a l l o c a t i o n   a c r o s s   a l l   f u n d   m a n a g e r s   w e i g h t e d   b y   p o r t f o l i o   s i z e δ : a   p a r a m e t e r   t h a t   c a p t u r e s   t h e   s e n s i t i v i t y   o f   p r i c e   t o   b u y i n g   p r e s s u r e . the~price~of~the~risky~asset~\\ 风险资产的价格~~P_t=\frac{1}{R_s-g}\bar u_t^\delta \\ g: the~growth~rate, R_sthe~discount~rate\\ \frac{1}{R_s-g}:~the~present~value~of~future~dividents\\ \bar u_t : the~ mean~ allocation~ across~ all~ fund~ managers~ weighted~by~portfolio~size\\ \delta :a~ parameter~ that~ captures~ the~ sensitivity~ of~ price~ to~ buying~pressure.\\ the price of the risky asset   Pt=Rsg1uˉtδg:the growth rate,Rsthe discount rateRsg1: the present value of future dividentsuˉt:the mean allocation across all fund managers weighted by portfolio sizeδ:a parameter that captures the sensitivity of price to buying pressure.

The realized yield on the risky asset is the log derivative of Eq.(1),
风 险 资 产 的 收 益 率 R 1 , t = ( R s − g ) u ˉ t − δ + g + δ ( u ˉ t − u ˉ t − 1 ) u ˉ t − 1 风险资产的收益率\\R_{1,t}=(R_s-g)\bar u_t^{-\delta}+g+\delta \frac{(\bar u_t-\bar u_{t-1})}{\bar u_{t-1}} R1,t=(Rsg)uˉtδ+g+δuˉt1(uˉtuˉt1)
where the first term represents the dividend yield, the second term represents capital gains due to underlying growth, and the third term represents short term capital gains (or losses) due to buying (or selling) pressure.
The payoff function of manager i is,
基 金 经 理 的 收 益 方 差 ϕ ( u i , t ) = u i , t ( R 1 , t + α i , t ~ − R 0 ) − 1 2 c 2 t ∗ u i , t 2 基金经理的收益方差\\ \phi (u_{i,t})=u_{i,t}(R_{1,t}+\widetilde{\alpha_{i,t}}-R_0)-\frac{1}{2}c2_t*u_{i,t}^2 ϕ(ui,t)=ui,t(R1,t+αi,t R0)21c2tui,t2
Managers adjust their exposure to risk by following the slope of the payoff function,
收 益 方 程 的 斜 率 ϕ ( u i , t ) ~ = ( R 1 , t + α i , t ~ − R 0 ) − c 2 t ∗ u i , t 收益方程的斜率\\ \widetilde{ \phi (u_{i,t})}=(R_{1,t}+\widetilde{\alpha_{i,t}}-R_0)-c2_t*u_{i,t} ϕ(ui,t) =(R1,t+αi,t R0)c2tui,t
每位基金经理都会调整他的风险头寸使其与收益方程的斜率成比例。如果$ \widetilde{ \phi (u_{i,t})}$ 为正,她会增加她的风险资产头寸。收益方程越陡峭这个操作的频率越高。
u i , t = ϕ u i , t ~ + u i , t − 1 u ˉ t = ∑ u i , t u_{i,t}=\widetilde{\phi u_{i,t}}+u_{i,t-1}\\ \bar u_t=\sum u_{i,t} ui,t=ϕui,t +ui,t1uˉt=ui,t

$ \widetilde{\alpha_{i,t}}$ is an idiosyncratic mean reverting alpha, 服从mean-reverting Ornstein-Uhlenbeck process 。
α i , t ~ = e − η α i , t − 1 ~ + 1 − e − 2 η 2 η σ ν \widetilde{\alpha_{i,t}}=e^{-\eta }\widetilde{\alpha_{i,t-1}}+\sqrt{ \frac{1-e^{-2\eta}}{2\eta}}\sigma \nu αi,t =eηαi,t1 +2η1e2η σν
c2 is the perceived risk factor. determined by using a loss function and exponential averaging
c 2 t = γ L N , t ~ L i , t = m a x 0 , − R G i , t R G i , t = ( R 1 , t + α i , t ~ ) u i , t L i , t ~ = e − η L i , t − 1 ~ + ( 1 − e − η ) L i , t c2_t=\gamma\widetilde{L_{N,t}}\\ L_{i,t}=max{0,-R_{G_i,t}}\\ R_{G_i,t}=(R_{1,t}+\widetilde{\alpha_{i,t}})u_{i,t}\\ \widetilde{L_{i,t}}=e^{-\eta}\widetilde{L_{i,t-1}}+(1-e^{-\eta})L_{i,t} c2t=γLN,t Li,t=max0,RGi,tRGi,t=(R1,t+αi,t )ui,tLi,t =eηLi,t1 +(1eη)Li,t
作者使用Morningstar 上面的美国1990年1月4日至2012年12月31日的至少有5%的资金投资于股票的股权基金的周频数据,一共17616支基金;还使用了3个月美国国债收益率数据、Baa级信用风险溢价、美国年度GDP增长率。

Calibration Model

作者用3个月的国债收益率作为 R 0 , t R_{0,t} R0,t, 用Baa级周频债券风险溢价作为 R s , t R_{s,t} Rs,t, 美国GDP增长率作为g,通过CAPM方程和跨期计算的方式来计算 α \alpha α, 通过对基金经理的头寸加权平均来计算 c 2 t c2_t c2t.
α i , t = R i , t − R 0 , t − β i , t ∗ ( R S & P , t − R 0 , t ) α i , t ~ = e − η α i , t − 1 ~ + ( 1 − e − η ) α i , t \alpha_{i,t}=R_{i,t}-R_{0,t}-\beta_{i,t}*(R_{S\&P,t}-R_{0,t})\\ \widetilde{\alpha_{i,t}}=e^{-\eta}\widetilde{\alpha_{i,t-1}}+(1-e^{-\eta})\alpha_{i,t} αi,t=Ri,tR0,tβi,t(RS&PtR0,t)αi,t =eηαi,t1 +(1eη)αi,t

本文通过回归来获得收益率序列
R i , t + k = β 0 + β 1 ∗ S p r e a d B a a − A a a + β 2 ∗ T b i l l 10 y r − 3 m t h + β 4 ∗ ( R m − R f ) + β 5 ∗ S M B + β 6 ∗ H M L + β t ∗ r e c e s s i o n + β 8 ∗ V o l + β 9 ∗ σ u + ε R_{i,t}+k=\beta_0+\beta_1*Spread_{Baa-Aaa}+\beta_2*Tbill_{10yr-3mth}\\ +\beta_4*(R_m-R_f)+\beta_5*SMB+\beta_6*HML+\beta_t*recession \\+\beta_8*Vol+\beta_9*\sigma_u+\varepsilon Ri,t+k=β0+β1SpreadBaaAaa+β2Tbill10yr3mth+β4(RmRf)+β5SMB+β6HML+βtrecession+β8Vol+β9σu+ε
Heterogeneity in Beliefs 的实证分析的部分,也是通过回归进行的。VIX回归主要是想检验头寸的标准差和VIX 的标准差是不是差异不大。
V I X % = β 0 + β 1 ∗ S p r e a d B a a − A a a + β 2 ∗ T b i l l 10 y r − 3 m t h + β 4 ∗ ( R m − R f ) + β 5 ∗ S M B + β 6 ∗ H M L + β t ∗ r e c e s s i o n + β 8 ∗ V o l + β 9 ∗ σ u + ε VIX_\%=\beta_0+\beta_1*Spread_{Baa-Aaa}+\beta_2*Tbill_{10yr-3mth}\\ +\beta_4*(R_m-R_f)+\beta_5*SMB+\beta_6*HML+\beta_t*recession \\+\beta_8*Vol+\beta_9*\sigma_u+\varepsilon VIX%=β0+β1SpreadBaaAaa+β2Tbill10yr3mth+β4(RmRf)+β5SMB+β6HML+βtrecession+β8Vol+β9σu+ε
作者还跑了Logit回归看是否当估计的头寸的标准差增加时市场会衰退。
R e c e s s i o n = β 0 + β 1 ∗ S p r e a d B a a − A a a + β 2 ∗ T b i l l 10 y r − 3 m t h + β 4 ∗ ( R m − R f ) + β 5 ∗ S M B + β 6 ∗ H M L + β t ∗ r e c e s s i o n + β 8 ∗ V o l + β 9 ∗ σ u + ε Recession=\beta_0+\beta_1*Spread_{Baa-Aaa}+\beta_2*Tbill_{10yr-3mth}\\ +\beta_4*(R_m-R_f)+\beta_5*SMB+\beta_6*HML+\beta_t*recession \\+\beta_8*Vol+\beta_9*\sigma_u+\varepsilon Recession=β0+β1SpreadBaaAaa+β2Tbill10yr3mth+β4(RmRf)+β5SMB+β6HML+βtrecession+β8Vol+β9σu+ε

Results

下图显示的标普500和模拟出来的收益率序列结果涨跌情况类似。

下图显示的是FA Model Validation。结果显示,eta越高(记忆率越低),经济泡沫期的估计价格越准确,eta越低,经济衰退期的估计价格越准确。这意味着在经济泡沫期,基金经理不怎么考虑历史情况,而基金经理对经济危机的记忆力是长久的。

下图显示的是VIX回归的结果以及标普500和预测的序列的分布情况

下面显示的是recession回归的结果:

看完这篇论文之后我发现其实大部分用还是金融经济学和计量经济学的基础知识,而且也没有想象中那么难。在中国市场实证要安排上日程!

欢迎关注二幺子的知识输出通道:

avatar

已标记关键词 清除标记
相关推荐
Model predictive control (MPC) has a long history in the field of control en- gineering. It is one of the few areas that has received on-going interest from researchers in both the industrial and academic communities. Four major as- pects of model predictive control make the design methodology attractive to both practitioners and academics. The first aspect is the design formulation, which uses a completely multivariable system framework where the perfor- mance parameters of the multivariable control system are related to the engi- neering aspects of the system; hence, they can be understood and ‘tuned’ by engineers. The second aspect is the ability of the method to handle both ‘soft’ constraints and hard constraints in a multivariable control framework. This is particularly attractive to industry where tight profit margins and limits on the process operation are inevitably present. The third aspect is the ability to perform on-line process optimization. The fourth aspect is the simplicity of the design framework in handling all these complex issues. This book gives an introduction to model predictive control, and recent developments in design and implementation. Beginning with an overview of the field, the book will systematically cover topics in receding horizon con- trol, MPC design formulations, constrained control, Laguerre-function-based predictive control, predictive control using exponential data weighting, refor- mulation of classical predictive control, tuning of predictive control, as well as simulation and implementation using MATLAB and SIMULINK as a platform. Both continuous-time and discrete-time model predictive control is presented in a similar framework.
©️2020 CSDN 皮肤主题: 数字20 设计师:CSDN官方博客 返回首页