Don’t Lose Your Shirt
Trading
Mean-Reversion
Quantopian Conference
April 2017
A Bit Of History
Edith Mandel | Greenwich Street Advisors, LLC
2
Guiding Principles
—  Bayesian Approach:
—  When the market not agreeing with Model,
posterior confidence in Model has to decline
—  Et
*[Model_Gains]*pt(Model✓)+λ*pt(Model✗),
— pt: posterior distribution at time=t
— Et
*[Gains|pt]: risk-adjusted forecast
—  Target a portfolio Πt with Max{Et
*[Gains|pt]}
—  Πt is updated continuously
—  Diversified set of trades
—  Don’t be fooled by lack of randomness (low
dimensionality)
—  Systemic & idiosyncratic risk factors
Edith Mandel | Greenwich Street Advisors, LLC
3
When A Trend Develops
Et
*[Gains|pt] is Reduced
Existing Trade is Sub-
Optimal Source of Risk
No Longer Part of Optimal
Portfolio
Edith Mandel | Greenwich Street Advisors, LLC
4
Cointegration via CCA
—  Decomposition of the panel, Yt, with covariance
matrix,Σ,into M non-stationary and N-M
stationary (i.e. cointegrated) directions
—  M will change throughout economic cycles
—  Methodologies to generate cointegrated vectors
—  Multivariate MLE, Johansen (1988)
—  CCA, Box-Tiao (1977)
—  Multivariate level regressions & CCA, Chou & Ng
(1994)
—  Regularized CCA
Edith Mandel | Greenwich Street Advisors, LLC
5
Cointegration via Box-Tiao
Edith Mandel | Greenwich Street Advisors, LLC
6
—  CCA with {Yt, Yt-1}
—  Yt=Yt-1A+εt
—  Q=Σ-1AΣAT
—  Q is a measure of relative
predictability
—  E=Eigenvectors of Q are
the factor loadings of the
Cointegration Model
—  Ft=ET Yt are the Factors
(Canonical Variates)
Cointegration Challenges
—  Larger portfolios are more cointegrated, but
difficult to execute
—  D’Aspermont (2008)
—  Maximizing mean-reversion under cardinality
constraints
—  Changes in cointegrated relationships
—  Ft=μt+ktFt-1+εt
—  H0: [μt,kt]=[μ0,k0]
—  SPC methods (CUSUM-like) applied to εt
— Ploberger, Kramer (1992)
— CUSUM(εt)->Brownian bridge
— Pr(max|CUSUM(εt)|>a) ->Posterior pt
Edith Mandel | Greenwich Street Advisors, LLC
7
Regimes: Mean-Reversion &
Trend (I)
Edith Mandel | Greenwich Street Advisors, LLC
8
Regimes: Mean-Reversion &
Trend (II)
Edith Mandel | Greenwich Street Advisors, LLC
9
Rapid Change
(Permanent Break)
Edith Mandel | Greenwich Street Advisors, LLC
10
Regime Changes
—  ‘Permanent’ breaks are rare
—  Regime-Switching Models (RSM) are more
suitable
—  Ang, Timmermann (2011)
—  2-State Auto-Regressive Model:
—  Ft=μs(t)+ks(t)Ft-1+σs(t) εt
—  s(t) is a Markov chain with time-invariant or time-
varying transitions:
— Pr{s(t)=0|s(t-1)=0}=p00,t Pr{s(t)=1|s(t-1)=1}=p11,t
—  Assumption of recurrent regimes is supported
by data
Edith Mandel | Greenwich Street Advisors, LLC
11
Factor-Based Approach
—  Factors Ft=ET Yt determine cointegrated vectors
—  M non-stationary factors (systemic risk factors), FR
—  K strictly stationary factors, FS
— Stable (but small) gains
— Suitable for limit order placement
—  N-(M+K) locally stationary factors, FLS
— Recurrent regimes
— 2-State Auto-Regressive RSM
—  F=[FR,FS,FLS]
—  Factor-replicating portfolios
—  Et
*[Gains|pt]=f{Et [Ft (Gains)], ΣF,pt}
Edith Mandel | Greenwich Street Advisors, LLC
12
Factors of US Trsy Market
Locally Stationary Factor Strictly Stationary Factor
Edith Mandel | Greenwich Street Advisors, LLC
13
Risk-Management
— Optimal portfolio Πt=Πt-1+wt
—  wt: new positions
— σ2(Πt)=EΣFET
—  F=[FR,FS,FLS]
—  E: factor exposures of Πt-1+wt
Edith Mandel | Greenwich Street Advisors, LLC
14
‘Bigger Data’ Approach
— Dynamics of different types of assets
(Stocks, Interest Rates, FX) are different
— Dynamics of cointegrated vectors are
more uniform
—  Crossing asset class boundaries
—  Far bigger training set to learn from
Edith Mandel | Greenwich Street Advisors, LLC
15

"Don't Lose Your Shirt Trading Mean-Reversion" by Edith Mandel, Principal at Greenwich Street Advisors, LLC & Adjunct Professor at NYU Tandon School of Engineering

  • 1.
    Don’t Lose YourShirt Trading Mean-Reversion Quantopian Conference April 2017
  • 2.
    A Bit OfHistory Edith Mandel | Greenwich Street Advisors, LLC 2
  • 3.
    Guiding Principles —  BayesianApproach: —  When the market not agreeing with Model, posterior confidence in Model has to decline —  Et *[Model_Gains]*pt(Model✓)+λ*pt(Model✗), — pt: posterior distribution at time=t — Et *[Gains|pt]: risk-adjusted forecast —  Target a portfolio Πt with Max{Et *[Gains|pt]} —  Πt is updated continuously —  Diversified set of trades —  Don’t be fooled by lack of randomness (low dimensionality) —  Systemic & idiosyncratic risk factors Edith Mandel | Greenwich Street Advisors, LLC 3
  • 4.
    When A TrendDevelops Et *[Gains|pt] is Reduced Existing Trade is Sub- Optimal Source of Risk No Longer Part of Optimal Portfolio Edith Mandel | Greenwich Street Advisors, LLC 4
  • 5.
    Cointegration via CCA — Decomposition of the panel, Yt, with covariance matrix,Σ,into M non-stationary and N-M stationary (i.e. cointegrated) directions —  M will change throughout economic cycles —  Methodologies to generate cointegrated vectors —  Multivariate MLE, Johansen (1988) —  CCA, Box-Tiao (1977) —  Multivariate level regressions & CCA, Chou & Ng (1994) —  Regularized CCA Edith Mandel | Greenwich Street Advisors, LLC 5
  • 6.
    Cointegration via Box-Tiao EdithMandel | Greenwich Street Advisors, LLC 6 —  CCA with {Yt, Yt-1} —  Yt=Yt-1A+εt —  Q=Σ-1AΣAT —  Q is a measure of relative predictability —  E=Eigenvectors of Q are the factor loadings of the Cointegration Model —  Ft=ET Yt are the Factors (Canonical Variates)
  • 7.
    Cointegration Challenges —  Largerportfolios are more cointegrated, but difficult to execute —  D’Aspermont (2008) —  Maximizing mean-reversion under cardinality constraints —  Changes in cointegrated relationships —  Ft=μt+ktFt-1+εt —  H0: [μt,kt]=[μ0,k0] —  SPC methods (CUSUM-like) applied to εt — Ploberger, Kramer (1992) — CUSUM(εt)->Brownian bridge — Pr(max|CUSUM(εt)|>a) ->Posterior pt Edith Mandel | Greenwich Street Advisors, LLC 7
  • 8.
    Regimes: Mean-Reversion & Trend(I) Edith Mandel | Greenwich Street Advisors, LLC 8
  • 9.
    Regimes: Mean-Reversion & Trend(II) Edith Mandel | Greenwich Street Advisors, LLC 9
  • 10.
    Rapid Change (Permanent Break) EdithMandel | Greenwich Street Advisors, LLC 10
  • 11.
    Regime Changes —  ‘Permanent’breaks are rare —  Regime-Switching Models (RSM) are more suitable —  Ang, Timmermann (2011) —  2-State Auto-Regressive Model: —  Ft=μs(t)+ks(t)Ft-1+σs(t) εt —  s(t) is a Markov chain with time-invariant or time- varying transitions: — Pr{s(t)=0|s(t-1)=0}=p00,t Pr{s(t)=1|s(t-1)=1}=p11,t —  Assumption of recurrent regimes is supported by data Edith Mandel | Greenwich Street Advisors, LLC 11
  • 12.
    Factor-Based Approach —  FactorsFt=ET Yt determine cointegrated vectors —  M non-stationary factors (systemic risk factors), FR —  K strictly stationary factors, FS — Stable (but small) gains — Suitable for limit order placement —  N-(M+K) locally stationary factors, FLS — Recurrent regimes — 2-State Auto-Regressive RSM —  F=[FR,FS,FLS] —  Factor-replicating portfolios —  Et *[Gains|pt]=f{Et [Ft (Gains)], ΣF,pt} Edith Mandel | Greenwich Street Advisors, LLC 12
  • 13.
    Factors of USTrsy Market Locally Stationary Factor Strictly Stationary Factor Edith Mandel | Greenwich Street Advisors, LLC 13
  • 14.
    Risk-Management — Optimal portfolio Πt=Πt-1+wt — wt: new positions — σ2(Πt)=EΣFET —  F=[FR,FS,FLS] —  E: factor exposures of Πt-1+wt Edith Mandel | Greenwich Street Advisors, LLC 14
  • 15.
    ‘Bigger Data’ Approach — Dynamicsof different types of assets (Stocks, Interest Rates, FX) are different — Dynamics of cointegrated vectors are more uniform —  Crossing asset class boundaries —  Far bigger training set to learn from Edith Mandel | Greenwich Street Advisors, LLC 15