nonlinear manifold learning for financial markets...

25
Nonlinear Manifold Learning for Financial Markets Integration George Tzagkarakis 1 & Thomas Dionysopoulos 1,2 1 EONOS Investment Technologies, Paris (FR) 2 Dalton Strategic Partnership, London (UK) DLM International Workshop Nice, 4 – 6 Sept 2017

Upload: others

Post on 30-Jan-2021

3 views

Category:

Documents


0 download

TRANSCRIPT

  • Nonlinear Manifold Learning for Financial Markets Integration

    George Tzagkarakis1 & Thomas Dionysopoulos1,2

    1 EONOS Investment Technologies, Paris (FR)2 Dalton Strategic Partnership, London (UK)

    DLM International Workshop

    Nice, 4 – 6 Sept 2017

  • Motivation

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Extract meaningful information

    from financial data

    Build smart trading

    strategies

    Quants Traders

  • Investment process

    Risk premia vs Risk factors

    Manifold learning for financial data

    Detection of critical transitions in financial markets

    Conclusions

    Overview

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Investment process

    Risk premia vs Risk factors

    Manifold learning for financial data

    Detection of critical transitions in financial markets

    Conclusions

    Overview

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Investment process

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Views(How do I produce my market signals)

    Portfolio Construction(How do I aggregate my views together to

    positions)

    Dynamic Budgeting(How does my portfolio evolve in time)

    • MVP [MinVar Portfolio]: lowest possible risk (volatility); concentration in low volatility asset classes

    • MCP [MinCorr Portfolio]: lowest volatility weighted average corrcoeff between asset classes; asset classes with low corr and volatility relative to other asset classes within the portfolio receive higher weight

    • RPP [Risk Parity Portfolio]: asset classes contribute the same amount of risk (volatility) to the overall portfolio; assets with lower risk (e.g. bonds) get a larger part of the portfolio than risky ones

    • Driver-based models: identify mathematical relationship between business drivers and study their financial outcomes under certain operational decisions

    • Rolling forecasting: continuous look forward on N-month basis updated monthly/quarterly to adjust positions in order to reach the financial goals

    Repeated pattern

    Statistics vs ThresholdMovAvg, MovVol, …

    {-1, 1}

  • Investments for the long haul

    Buy-and-Hold mentality

    Resist the temptation to react or predict the stock market’s every next move

    Successful passive investors keep their eye on the prize (returns) and ignore short-term setbacks, even sharp downturns

    Passive vs Active investing

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Beat the stock market’s average returns and take full advantage of short-term price fluctuations

    Involves a much deeper analysis and expertise to decide when to pivot into or out of a particular asset

    Accurate determination of when and where prices change will be critical

    Passive Active

  • Understand the true drivers of returns (From assets to small number of common factors)

    Importance of true diversification/stability as opposed to decorrelation

    Time-varying risk budgeting

    Challenges of active investing

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Investment process

    Risk premia vs Risk factors

    Manifold learning for financial data

    Detection of critical transitions in financial markets

    Conclusions

    Overview

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Commodities

    Equities

    Forex

    Bonds

    Interest rates

    Assets Systematic Exposure

    From assets to risk premia

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Momentum

    Mean Reversion

    Value

    Volatility

    Low Risk

    Liquidity

    RISK PREMIA

    [A market segmentation]

    Investors treat markets as a purely economic system

    Investment decisions are made empirically based on economic interpretations of the markets

  • Premia vs Factors

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Segment market without relying on economic interpretations Exploit the power of mathematical and signal processing tools

    Concept: Focus on achieving orthogonality

    Total risk = Sum of marginal risks Risk Decomposition

    Segmentation of market lacks economic interpretation BUT we are able to extract hidden information

    Green World

    (e.g. equities)Investors

    Blue

    Yellow

    Quants

  • Commodities

    Equities

    Forex

    Bonds

    Interest rates

    Assets Dimensionality Reduction

    From assets to risk factors

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    PCA, PPCA

    Fourier Transform

    Wavelet Transform

    Machine Learning

    Pattern Analysis

    Manifold Learning

    RISK FACTORS

    [Market portfolios]

    We treat markets as a mathematical system

    Investment decisions are made based on market information extracted via DSP/ML/… methods

  • Chasing the decorrelation

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    [Invariance]

    Time Seriesto

    Returns

    From ReturnsTo

    Smooth Manifolds

    The desired portfolio is NOT the solution

    to some optimization problem but an

    invariant of a dynamical system

  • Investment process

    Risk premia vs Risk factors

    Manifold learning for financial data

    Detection of critical transitions in financial markets

    Conclusions

    Overview

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Optimal portfolio as an invariant of a dynamical system

    Optimal portfolio as an invariant

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Traditional view(optimization-based)

    Utility function: max𝐄 𝒓𝒆𝒕𝒖𝒓𝒏𝒔 +min 𝑹𝒊𝒔𝒌Solution: straight line

    Modern view(invariant-based)

    Invariance with respect to the 1st (returns) and 2nd

    (Var) derivatives(changes of these two derivatives affect portfolio performance)

    Allow more features of the portfolio to remain invariant (e.g. correlations (= angles) between assets)

  • Financial data as dynamical systems

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Time-Delay Embedding

    Financial Time Series

    Manifold Learning

    Early Warning

    • Financial markets are highly complex,nonlinear dynamical systems

    • Financial time series are comprehensivereflections of market condition/operationsand provide the ground for market analysis

    • Dynamic nature of original system is oftencorrupted by irrelevant components thatdisturb useful intrinsic features

    • Extract underlying manifold structure thatgoverns the dynamical system; embeddingin a more stable/smooth low-dimensionalspace

  • Taken’s theorem: complete information about the hidden state of dynamical systems can be preserved in observed time series

    Phase space reconstruction

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Critical Parameters

    τ Delay

    m Embedding dimension

    N ( = n - (m - 1)τ ) Number of states

    1st min of Average Mutual Information

    1st min of False Nearest Neighbors (%)

    AM

    I

    Time lag

    FN

    N (

    %)

    Dimension

  • Information-based manifold learning

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Obtain the attractor manifolds by preserving geodesic distances between points in the state space

    Classical Manifold Learning

    Data points in state space

    Data representation via probability distributions (e.g. risk quantification)

    Considering only the geometric structure of a data space hides essential characteristics of the data and destroys the proximity relations (topology) of the original data space

    Financial Practice

    PDFs of data points in state space

    Measure informationchange between data points

  • Information-based manifold learning

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Sta

    te (p

    has

    e) s

    pac

    e

    Kernel Density Estimator

    K: Gaussian kernelh: plug-in bandwidth selection

    Information SimilarityGlobal Relationship

    Matrix

    Extract an extended local linear structure (adjacency doesnot entirely depend on the states’ geometric relations)while retaining the global topological characteristics in theinherent low-dimensional manifold

  • Employ locally linear embedding (LLE)Maps its inputs into a single global coordinate system of lower dimensionality

    Optimizations do not involve local minima

    Recovers global nonlinear structure from locally linear fits; local geometry of locally linear patches characterized by linear coefficients that reconstruct each data point from its neighbors

    Information-based manifold learning

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Embedding Cost FunctionTranslation-free embedding

    Rotation/Scaling-free embedding

    Solution: d eigenvectorscorresponding to the smallest

    d eigenvalues of A

  • Investment process

    Risk premia vs Risk factors

    Manifold learning for financial data

    Detection of critical transitions in financial markets

    Conclusions

    Overview

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Hard to predict accurately the market shifting points

    Detect gradual increase of transition points’ likelihood

    HMM classifier on the learned manifold of probability distributions 3 states (classes) [High, Medium, Low risk]

    Early warning for market transitions

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Learned Manifold

    Initial State Probability Distribution (GMMs)

    Construct State Transition Matrix P

    Compute Posterior Probability

    Classification of the corresponding time series point (end of phase space

    vectors)

  • S&P500 Index: daily closing prices in the period 2005-2016

    Estimated phase space parameters: m = 8, τ = 23

    Manifold learning/Early warning over sliding windows: length = 250, step = 25

    3 warning states (posterior thresholds): 50%-70% Low, 70%-90% Medium, 90%-100% High

    Early warning for market transitions

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

    Manifold of S&P500

    Early warning signals

    Pre-crisis period Crisis period

    • High posteriors concentratedin 2007-2009

    • Medium posteriors (earlysigns) in 2005-2006 (US realestate bubble)

  • Investment process

    Risk premia vs Risk factors

    Manifold learning for financial data

    Detection of critical transitions in financial markets

    Conclusions

    Overview

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • Take home messages

    Manifold learning (ML) is an efficient framework for unveiling intrinsic dynamic structure of financial systems

    Traditional geometry-based ML methods are not proper to handle investors’ probabilistic (risk-based) view

    Information distance-based ML measures more complex relationships between financial data in a phase space

    ML coupled with a conventional HMM enabled accurate identification of critical market transitions, providing reliable early warnings for investors

    Future work: Study the effect of differential curvature of a financial system through its attractor manifold

    as an indicator of market resilience against external disturbances Examine alternative manifold learning techniques and distance measures adapted to

    financial data

    DLM International Workshop

    Nice, 4 – 6 Sep 2017

  • DLM International Workshop

    Nice, 4 – 6 Sep 2017