in medias res - 4.15.pptnymetro.chapter.informs.org/prac_cor_pubs/04-2015 thomas ball_in...
TRANSCRIPT
In Medias Res:
Fads, Trends, Bubbles and Massively Scaled Analyses in RTW Fashion
Thomas Ball
INFORMS
April 22, 2015
Highlights
• The importance and relevance of fashion
• Growth is king
• Watersheds in strategic thought
• Massively scaled analyses in the RTW fashion ecosystem
• Modeling of fads and trends: towards a protocol
• The bottom line
The Importance And Relevance Of Fashion
Fashion may be the world‘s largest and most important creative industry
• A global business with recent annual U.S. sales of more than $250 billion -- ~2% of GDP --larger than that of books, movies, and music combined
- Has blended into the wider arena of “Entertainment”
• Greenhouse for the analysis of fads and trends
• Provides economists, sociologists and other cultural thinkers and critics with canonical examples of consumption, conformity and consecration
• RTW fashion data possess all the computational challenges inherent in any massive analysis
Sources: Hemphill and Suk, The Law, Culture and Economics of Fashion, 2009; Aspers and Godart, Sociology of Fashion: Order and Change, 2013Cattani, Ferriani and Allison, Insiders, Outsiders, and the Struggle for Consecration in Cultural Fields, 2014; Teri Agins, The End of Fashion, 2000
Non Gustibus Disputandum Est
Defining Fads And Trends
Some tentative definitions – a few terms with many synonyms
• A fad is a short-term burst in behavior usually starting with explosive growth that rises to a peak followed by slower ebbing -- analogous to froth from waves breaking on a beach
– Bubbles are closely related to fads but are usually financial in nature and refer to unrealistic prices detached from intrinsic value -- can be both positive and negative
• A trend is a long-term or enduring influence on behavior -- analogous to open ocean waves
– The Ancients had no concept of a “trend” viewing existence as eternal and static
– Modern notions are typically credited as originating with Vico’s 1725 Nuova Scienza
– Population demographics are among the most important structural drivers of trends
Wave Motion Does Not Change With Water Depth
Sources: Daniel Bell, Personal communication, 2002; Didier Sornette, Financial Crisis Observatory, ETHZ; Robert Shiller, Irrational Exuberance, 1999; Eli Pariser, The Filter Bubble, 2011
Froth Forms
“Fads”“Trends”
Sources: http://www.theworldeconomy.org/MaddisonTables/MaddisontableB-10.pdf, http://kk.org/thetechnium/archives/2008/10/the_expansion_o.php , http://smartregion.org/2011/03/creative-classDaniel Bell, The Coming of Post-Industrial Society, 1974; Richard Florida, The Rise of the Creative Class, 2014; Deirdre McCloskey, Bourgeois Dignity, 2012; Google Ngrams Analysis
1500 1560 1600 1650 1690 1725 1775 1800 1850 1885 1925 1965 2010
In terms of the forces of history, we find ourselves in medias res…
Growth Is King
• Exponential growth of everything since the Industrial Revolution – GDP to double by 2050
• A “Knowledge Society” emerged when the Service Sector eclipsed Manufacturing in size and growth of occupations
• The production and flow of ideas is a primary source of growth
Wealth and Population1-2010 AD
$0
$10,000
$20,000
$30,000
1 1000 1500 1600 1700 1820 1870 1900 1950 1970 2010
0
2
4
6
Population
Year (Discontinuous)
U.S. Occupational Change1800-2010
Millions Billions
Wealth
Production of Ideas1500-2010
Year
# Books Printed
Agriculture
Manufacturing
Low Wage Service Sector
Knowledge Workers
Millions Employed70
60
50
40
20
30
10
0
Service Sector
Year
~Industrial Revolution ~Knowledge
Society~Industrial Revolution
~Knowledge Society
~Knowledge Society
Swing towards greater uncertainty and disruptions to equilibrium in normative business practices
Quantifiable Risk, Unknowable Uncertainty and The Business Landscape
Watersheds In Strategic Thought
A Clear Enough Future
What can be
known?
Linear forecasts drive strategyChange is gradual and incrementalBehavior is deterministic and predictableSustainabilitySix Sigma precisionConformance with commonsenseSandboxes
Alternate FuturesA few discrete outcomes
define the future
A Range of FuturesA range of possible outcomesNo natural scenarios
From Risk to Greater Complexity and UncertaintyLow High
1
2
3
• Anomalies regarding Porterian assumptions of sustainability of competitive advantage
• Hypercompetition in a widened arena of business operations versus myopic industry silos
• Shift from moment-based, linear models rooted in “normality” and simple iid relationships to models exploiting nonlinearity, complex dependence, power law distributions, infinite moments, heavy tails
True UncertaintyNonlinear systems dominateExtreme shifts can occur abruptly, without warningBehavior is deterministic but not predictableTransienceArbitrage in ignorance, diffidence, approximationExpectancy violationsSandpiles
?
Sources: Courtney, Kirkland and Viguerie, Strategy Under Uncertainty, HBR, 1999; Richard D’Aveni, Hypercompetition, 1991; Rita McGrath, The End of Competitive Advantage, 2012Frank Knight, Risk, Uncertainty and Profit, 1921; Nissam Taleb, Silent Risk, 2014; Embrechts, et al., Modeling Extremal Events, 1996; John Deighton, The Value of Data, 2013Arthur de Vany, Hollywood Economics, 2003; Anita Elberse, Blockbusters, 2013; David Hand, The Improbability Principle, 2014; Sarah Kaplan, Beyond Forecasting, 2014Cosima Shalizi, The Statistical Analysis of Complex Systems Models, 2010; Aswath Damodoran, Living with noise: Investing under uncertainty, 2013
FEEDBACK/PROPAGATE
3 Years of Sales
Decline
“Blooming, Buzzing Confusion”
Ze
itg
eis
t V
isio
na
rie
s –
“S
up
er
Ta
ste
rs”
Pla
ce
Be
ts
Sources and Tools
Long-Tailed
Product?Growth
MaturityDecline
Decelerating Sales
Pe
rform
an
ce
RDBMS Sales DataFast Fashion, e.g., ZaraMarketing SpendSocial Media OSIsCompetitive Info from Online
Comparison EnginesWeb ScrapingeBay
Ma
nu
fac
turi
ng
th
e P
ort
foli
o
Introduction
Accelerating Sales
0
Innovation-Development Pipeline Sales Curves and the Product Life CycleExecution
Paris’ Premiere Visione
Social Media OSIsPatents, Academic PapersFilm, Books, Mags, Art, etc.VC InvestmentsDemographics (Youth and
Agelessness)Tech ConferencesBlogs
Go – No Go
Visionaries Originate While Markets Imitate, Diffuse and Consecrate
Rapidly Cycling Fads And Trends, Massive Classification Of Fashion Styles Drive Need for Massively Scaled Analysis
Evidence-Based Decision-Making
R&D-Tacit KnowledgeText and Image MiningPrediction MarketsContinuous TrackingCompetitive Info from Online
Comparison EnginesWeb ScrapingHiring TrendseBay
La
un
ch
Time
Sc
an
, M
on
ito
r, O
rig
ina
te,
Imit
ate
From Uncertainty to Risk
Data MiningPredictive ModelingNetwork and Diffusion ModelsAgent-Based ModelsMassively Categorical ModelsMachine Learning AlgorithmsRecommender Systems
Sources and Tools
Fads, Trends And Explosive Self-Generating Demand
Unrelated phenomena such as the explosive flow of water out of a breached dam versus “fad” behaviors in Google search activity can be seen as structural homologues
Water Flow From A Breached Dam vs “Interest” in Justin Bieber
Sources: V. Seshadri, The Inverse Gaussian Distribution, 1994; Google Trends, March 2015
0
2 5
50
75
10 0
Jan-
09
Jul-
09
Jan-
10
Jul-
10
Jan-
11
Jul-11 Jan-
12
Jul-
12
Jan-
13
Jul-
13
Jan-
14
Jul-
14
Jan-
15
“J
us
tin
Bie
be
r” G
T S
ea
rch
In
tere
st
• There are at least two challenges inherent in this:
- Separating wheat from chaff or epiphenomenal flotsam and jetsam from emerging trends
- Finding scalable computational solutions for massive numbers of “Biebers”
Tracking Fads And Trends With Machine Learning Algorithms
*Word-of-mouth, open source indicatorsSources: http://www.jingdaily.com/from-social-status-to-self-expression-the-rapid-evolution-of-chinas-street-style/42059; Renee Dye, The Buzz on Buzz, 1999
Personal communications: Svante Jerling, P1.cn; Karen Moon, Trendalytics.co; Josh Clark, BoazandClark.com; David Wolfe, Doneger GroupIARPA.gov papers on OSIs, in Google search window enter “D12PC00337 OR D12PC00285 OR D12PC00347”
- Structuring unstructured information in the wilderness of data is difficult to do at scale
- “Words slip and slide and never stay in place” and so do images
• The evolution of Chinese fashion styles based on algorithmic image mining of tens of thousands of pictures taken on the streets of Shanghai and Beijing suggests:
- Tastes may have shifted from conspicuous status statements using big logo brands such as LVMH to niche brands
- Slowing economic growth as well as crackdowns on corruption and pirating also contributed
Text and image mining of WOM*, social media OSIs* is an analytic “Wild, Wild West”
Tracking Pictures of Handbags With Louis Vuitton Logo in China2008-2014
Modeling Fads And Trends
Fads and trends can be quantified using models of tech and new product innovation, diffusion, adoption and evolution rooted in the analysis of nonlinear logistic growth
Sources: Jesse Ausubel, DRAMs as Model Organisms for Study of Technological Evolution, 2001; Steven Johnson, Where Good Ideas Come From, 2010; Jonah Berger, Contagious, 2013Peres, Muller, Mahajan, Innovation diffusion and new product growth models: A critical review and research directions, 2009Alex Pentland, Social Physics : How good ideas spread, 2014; Julie Cohen, Configuring the Networked Self, 2012; Harrison White, Markets from Networks, 2002Lee Cooper, Market Share Analysis, 1989; Gelman and Hill, Data Analysis Using Regression and Multi-Level Models, 2007; Singer and Willet, Applied Longitudinal Data Analysis, 2003“Small World” social networks, e.g., “I shook Frank Sinatra’s hand,” Six Degrees of Kevin Bacon, Six Degrees of Francis Bacon
• Classic models focused on univariate time series and diffusion processes based on cumulants
of new adopters or sales, e.g., Gompertz and Bass-Anderson models, Fisher-Pry transforms
- Aggregates of individual decisions give a normative description of the adoption life cycle
- S-shaped curves identify inflection points and carrying capacities (ceilings or asymptotes)
• Recent research generalizes this framework to more complex, disaggregate, multi-level and multivariate growth processes leveraging, e.g., pooled time series, marketing mix or multi-level regression potentially with multiple DVs, network analysis, information theoretic frameworks, etc.
Classic Diffusion and S-Shaped CurvesEight Generations of DRAM Chips, 1970-2000
1970 1975 1980 1985 1990 1995 2000
Year
Social Networks
Differentiating Fads From Trends: Towards A Protocol
It is possible to distinguish fads from trends using a hybrid, generalized approach
- Based on pre-determined “burn-in” periods, partition products by phases of the life cycle
- Fads – “go, no go” phase for new products with time “zero” origin and data of short duration
- Use automated, cumulant diffusion models for insight into growth, ceilings or cancellation
- Early stages of an emergent fad have the least information and are the hardest to predict
-How much information (# data points) is needed for “go, no go” decisions?
- Trends – left-censored, pre-existing products with established sales curves
- Use actual sales, not cumulative, for insights into growth rates (slope, 1st derivative) and acceleration or momentum (Hessian, 2nd derivative, the rate of change in the slope)
0
1,000
2,000
3,000
0
100
200
300
400
500
Time
Product Sales for “Fads” and “Trends”Left Aligned, Past 12 months Cumulative
“Fad” ProductsCumulative Sales
“Go”
“Go?”
“No Go”
Cu
mu
lati
ve
U
nit
s
Ac
tua
l U
nits
Actual
“Zero” Time Origin
Left-Censoring
“Trend” Product Sales
0
20
40
60
80
-2.8 -1.3 -0.3 0.7 1.7 2.7 3.7 4.7 5.7 6.7 8.6
Classifying Trajectories Based On Growth And Acceleration
Growth
Product Growth vs Acceleration Rates
Ac
ce
lera
tio
n
A comparison of growth rates (slope or 1st derivative) with the rate of acceleration in those slopes (Hessian, 2nd derivative) suggests strong association but qualitatively distinct information
- Model accuracy may not be improved but are the insights greater from adding momentum?
-15
-10
-5
0
5
10
15
20
25
30
35
-4 -2 0 2 4 6 8 10
ILLUSTRATIVE
Distribution of Product Acceleration Rates
#
#
Distribution of Product Growth Rates
0
25
50
75
100
-21.
4
-4.6
-3.2
-2.2
-1.2
-0.2 0.8
1.8
2.8
3.8
4.8
5.8
6.8
7.9
8.9
9.9
11.0
12.2
14.1
15.8
18.8
31.2
“0” Points “0” Point
n~2,000
Challenges In Developing Generalized Growth Models
Is the trend really your friend?
*Heteroscedasticity, autocorrelation consistentSources: Ainslie and Steenburgh, Massively Categorical Variables: Revealing the information in zip codes, Marketing Science, 2002; Anita Elberse, Blockbuster, 2013
Van Den Bulte and Lilien, Bias and Systematic Change in the Parameter Estimates of Macro-Level Diffusion Models, 1997; Edward Thorpe, Beat the Dealer, 1964Ron Gallant, Nonlinear Models, 1988; Wang, Chen, Schifano, Wu and Yan, A Survey of Statistical Methods and Computing for Big DataEmmert-Streib and Delmer, Information Theory and Statistical Learning, 2009; Bikhchandani, Hirshleifer and Welch, A Theory of Fads, Fashion, Custom, and Cultural Change as InformationalCascades, 1992; Wang and Zhang, Reasons for Market Evolution and Budgeting Implications, 2008; Personal communications Mike Hanssens, Gary Lilien, Renana Peres, 2015Chattopadhyay and Lipson, Data smashing: uncovering lurking order in data, 2015; Andreas Brandmaier, pdc: An R Package for Complexity-Based Clustering of Time Series, 2014
- Univariate diffusion models are too simplistic
- Don’t capture important factors such as competition or the marketing mix (ex Extended Bass)
- Prima facie issues with HAC* and nonstationarity (cf. Box-Jenkins, “p’s and q’s”)
- Designed and built for successful new products -- but most new products fail(!)
- Hybrid, generalized models leverage multi-level, pooled, marketing mix, etc., robust regression
- Handle extreme value data in the original units of the dependent variable(s)
- Incorporate information related to competitive effects, marketing spend, market or consumer heterogeneity, social media, social networks, etc., as appropriate and available
- Find nonlinearities in momentum of growth rate (slope, 1st derivative) based on the rate of change or acceleration in that slope (Hessian, 2nd derivative)
- Exploit endogenous and combinatoric interactions inherent in massively categorical data to estimate, e.g., the impact of on sales of trends in a color or button size
- Random forests, Divide and Conquer, BLJs (bags of little jacknifes) compiled on massively parallel CPUs are among the approximating workarounds for scalable statistical modeling
- Another session could cover the advent of featureless, pattern matching, machine learning, complexity-based algorithms, e.g., permutation distribution clustering or “data smashing”
The Bottom Line
How can the analysis of fads, trends and bubbles be used to enhance business performance?
• Models are calibrated on a known world and projected into an unknown, uncertain future
– Model performance can be benchmarked several ways: 1) improvement over an incumbent method, 2) % correct prediction in a portfolio over and above random guessing and 3) prospective (not historic) predictive accuracy
• As in Vegas, beating the house 3%-4% of the time is pretty darn good
– That is, if one beats the house at all
– Relate to key corporate metrics such as YAG sales, stock price, financial ratios, etc.
• Stronger strategic planning, analysis and inventory control of new and existing products from the insights available in hybrid, generalized growth and diffusion analysis
– Widened set of strategic, evaluative and validatory metrics of prospective performance that go beyond purely data-based, historic predictive accuracy
– Evaluation of impact of marketing mix and other activities on product evolution
– Use in inventory control as an aid to answering questions related to potential market size, depth of purchase, duration and timing for when to get in or out
– Early warning for explosive product growth, negative revenue surprises
– Track cross-product elasticities and interdependence for cross-sell
• Extreme value, power law nature of information suggests changing emphasis from predicting averages to predicting quantiles or points on a heavy-tailed distribution
• Shifting views of growth as smooth and linear to recognition that it is inherently discontinuous, nonlinear, inefficient, lumpy, messy
Next Steps
- “Big whirls have little whirls that feed on their velocity, and little whirls have lesser whirls and so on to viscosity” – Lewis Frye Richardson
Exploring deeper mathematical metaphors for fads and trends in models of chaos and turbulence
Thank You!
Thomas [email protected]