Balancing quantitative models with common sense 2008

Balancing Technological Change with Intuition

Nick Wade

Director, Asia Marketing

Northfield Information Services Asia Ltd.

[email protected]+81 (0)3 5403 4655+61 (0)2 9238 4284

04/10/23 2

“30,000 foot view” or “999 things to find later on Google”

Northfield

What is changing in our world?

What are the implications?

Degustation Menu of Recent Quant Tools

A few applications of technology

04/10/23 3

Northfield Overview

Northfield – established 1985, open philosophy, over 300 clients

Northfield products and services fit each part of the investment process

Innovative risk modelling principles & broad coverage

Appropriate Risk models – and hence our motivation

04/10/23 4

What is Changing

Data Availability, timeliness, accuracy(?)

Market Dynamics

Models & Techniques

Business Processes - Applications

Client, Regulatory and Competitive Environment

04/10/23 5

DATA: Distinguishing Signal from Noise

Old Problem: no dataCurrent Problem: no filter

Still need to know:how is it calculated?is it reliable?is it comparable?

What is information, and what is noise (subjective)?

- Data is a commodity, not a competitive edge- Whoever has the best filter wins?

04/10/23 6

TECHNOLOGY: Just because we can doesn’t mean we should

Northfield MARS: Fully customized tax-aware optimization of 200,000 HNW client accounts in a couple of hours (good idea)

Intra-day performance attribution(bad idea)

Statistical inference problemsCross-term problems

Technology allows us to do great things……and a lot of things we shouldn’t…

04/10/23 7

MARKET: market evolution example 1

effect of hedge funds’ increasing share of the volume – highly correlated “bad days” across “uncorrelated” asset classes as hedge funds have “fire sale” and close out liquid positions to meet margin calls. Business as usual on up days.

Implications:Asymmetric betaDifferent up/down market correlationsdifferent small/large move correlations

Time for Non-linear asymmetric models?

04/10/23 8

MARKET: market evolution example 2

Speculative Trading – example China. See Derman paper on “temperature”. This plays hell with many of our usual assumptions, but can be intelligently modelled fairly easily.

Implications:Include trend (bubble) riskAdjust for skew, kurtosis

Northfield currently do this for all our equity risk models…

04/10/23 9

MARKET: market evolution example 3

Illiquid assets / funds / appraisal pricing / mark-to-market issues – adjusting for lack of real trade information

Political risk of pricing being controlled by portfolio management who bought it / broker they bought it from.Expect to see a mini-industry in third-party “objective” pricing?

04/10/23 10

MARKET: market evolution example 4

HF strategies: various ways to disguise short-vol strategies, but if it quacks like a duck… selling insurance e.g. OTM puts to ramp up return adds value little by little every day most days, very occasional catastrophic lossesIncreasing bets after lossesCan happen by accident – shorts go bad

Implications: more sophisticated risk measures needed to capture short-vol strategies, bets moving opposite to rules

04/10/23 11

Pause for thought:

As markets evolve we must question our assumptions:

Arbitrage Pricing Model:perfect competition

linear relationship between factors & return

Modern Portfolio TheoryQuadratic utility function – not true for levered HF or trading desk

04/10/23 12

MODELS: it’s not just linear regression

TechniqueDecision Systems: AHP, ANP, MCDMTime-Series Models: GMM, Markov Chains, HME, etcMachine Learning:

support vector machines, radial basis functions, “boosting”Complex Event Processing, Event Stream ProcessingTemporal Difference learning

ApplicationDynamic Alpha Models: Sorensen, MacquarieHybrid Risk Models (Northfield)Simultaneous Estimation of risk models (UBS, MacQueen, Heston & Rouwenhorst)

04/10/23 13

Models: decision systems

Objective: robust result based on expert opinion and auditable process adds consistency to typically subjective rules of thumb…

Applications:

Suitability in asset allocation

Credit risk scoring

Examples:

Analytic Hierarchy Process, Analytic Network Process

Multiple Criteria Decision Modelling

04/10/23 14

Models: Time-Series Techniques

Objective: Predicting the next observation from a series of observations where the distribution varies over time

Arima, Garch, Kalman

Multiple generators e.g. Hidden Markov Experts (GMO implemented this)

Issues: - over-fitting- reliance on history / sample- tells us nothing about why the distributions change…- linear regression analogy: 1000 factors, great R2, lousy out-of-sample performance

04/10/23 15

Models: Machine-Learning

Objective: Pattern recognition – humans are brilliant at it, and most of machine learning attempts to replicate how we learn.

Applications:Recognizing patterns in financial time series

Predicting next observation based on learning

04/10/23 16

Machine Learning Example: Complex Event ProcessingObjective: Pattern Recognition – imply the “complex” event from an event

“cloud”

White gown, tuxedo, bells, rice in air => weddingWhite gown, tuxedo, band playing, rice in air => rowdy dinner party

Obvious applications to algorithmic tradingAvoid specifying distribution of underlying factors

Problems:Assumption that history will persist – but 1998 Russian debt crisis, whilst very similar to debt aspects of sub-prime crisis, did not spill over into equities…

04/10/23 17

Machine Learning Example: Temporal Difference Learning

Objective: adapt/revise consecutive forecasts based on the accuracy of recent forecasts

Example: 5 day weather forecast around a hurricane event

Used as a training technique in some machine-learning situations

04/10/23 18

Machine Learning Example: “boosting”

Objective: make a good prediction out of a set of poor predictions

Gambler 1: apportion money across a team of friends based on how they are doing, aim to get close to what we would have won if we’d bet everything with luckiest friendGambler 2: take a number of “rules of thumb” that work sometimes, or better than random, and use those together to make a good prediction

Application: obvious application to dynamic weighting of alpha signalsAsset allocation?

04/10/23 19

Machine learning Example: SVM, RVM, and sparse Bayesian modelsObjective: pattern recognition in data, for example handwriting

Support Vector MachinesRelevance Vector MachinesInformative Vector MachinesLinear / Non-linear Kernel MethodsSparse Bayesian ModelsOnline Gaming? (Microsoft)

RVM advantage – avoid a lot of cross-validation with SVM, but EM based so can get stuck in local minima…

Problem: all predicated on the idea that the future will be the same as the past, and subjective in the sense that they are sample dependent

04/10/23 20

Risk Model Design: Hybrid Risk Model

Objective: improve risk forecast accuracy by better inclusion of real-world dynamics

Hybrid Model: temporary factors

x2 variance not (x-µ)2

Parkinson volatility

Exponential weighting

Our existing linear framework can accommodate a variety of real-world violations of Arbitrage Pricing Theory

04/10/23 21

Pause for thought

Advances in machine learning and time-series techniques allow us to (over-) fit a much wider variety of data series

New techniques like boosting and temporal difference learning may help us improve e.g. switching between alpha strategies or asset allocation

Innovations in decision systems can help us make subjective processes more robust

Caveat: As these techniques become more accessible and more widely known, the goalposts move…

04/10/23 22

BUSINESS PROCESSES:

AHP: suitable asset allocationDefine a questionnaire for suitability issues

Expert Involvement: hierarchy of asset classes

Expert Involvement: hierarchy of questions

MARS: portfolio manufacturingLink accounting, custodian, client, and trading systems

Allow full customization of investment choices

Automated to demand minimal intervention

04/10/23 23

REGULATORY/COMPETITIVE PRESSURES:

Increasing regulatory pressure on risk

Client pressure – risk is value added?

Time-to-market, barriers to entry, IP

Soft-dollar disappearing

Best execution pressure

“Centralized Portfolio Management” - Vanguard

04/10/23 24

ConclusionsMarket Dynamics changing as participants change

Technology offers compelling advantagesResist temptation to overfit!High Frequency Data – more trouble than it’s worth?

Major Modelling Advances in the last few years

Stay current on modelling, but balance technological advances with your own intuition