assessing the asymmetric information associated with the equity market a cart based decision rule...
TRANSCRIPT
Owen P. Hall, Jr., P.E., Ph.D.Owen P. Hall, Jr., P.E., Ph.D.
Pepperdine UniversityPepperdine University
CART ConferenceCART Conference
May, 2012May, 2012
San Diego, CASan Diego, CA
Assessing the Asymmetric Information Associated with the Equity Market:
A CART Based Decision Rule Analysis
Presentation AgendaPresentation Agenda
OverviewOverview Problem StatementProblem Statement Results AnalysisResults Analysis ConclusionsConclusions
Problem StatementProblem Statement
Assess the effectiveness of analytics to Assess the effectiveness of analytics to detect asymmetric information associated detect asymmetric information associated with the equity marketwith the equity market ModelsModels
• Probabilistic Neural netsProbabilistic Neural nets• CARTCART
FactorsFactors• Classic (e.g., Price Momentum)Classic (e.g., Price Momentum)• Tobin’s QTobin’s Q• EntropyEntropy
ChallengeChallenge In an efficient market, the current In an efficient market, the current
prices of securities represent prices of securities represent unbiased estimates of their true or unbiased estimates of their true or fairfair market value at all timesmarket value at all times
This principle suggests that neither This principle suggests that neither technical analysis nor fundamental technical analysis nor fundamental analysis can assist investors in analysis can assist investors in identifying undervalued or identifying undervalued or overvalued stocksovervalued stocks
I'd be a bum in the street with a tin cup if the markets were efficient -- Warren Buffett
Classic FactorsClassic Factors
Price MomentumPrice Momentum Earnings MomentumEarnings Momentum ValuationValuation SystemSystem EconomicEconomic
EntropyEntropy
The basic idea is that more volatile The basic idea is that more volatile securities have a greater entropy state than securities have a greater entropy state than more stable securities more stable securities
Two fundamentally different phenomena Two fundamentally different phenomena exist in which time based securities data exist in which time based securities data deviate from constancy:deviate from constancy: Exhibit larger standard deviationsExhibit larger standard deviations Appear highly irregularAppear highly irregular
The standard deviation measures the extent The standard deviation measures the extent of deviation from centrality while entropy of deviation from centrality while entropy delineating the extent of irregularity or delineating the extent of irregularity or complexity of the data setcomplexity of the data set
EntropyEntropy Two entropy modelsTwo entropy models
Approximate entropy (ApEn)Approximate entropy (ApEn) Sample entropy (SaEn)Sample entropy (SaEn)
Model inputsModel inputs Time seriesTime series Matching template length (M)Matching template length (M) Matching tolerance level (r)Matching tolerance level (r)
Time series length (50 months)Time series length (50 months)
Tobin’s QTobin’s Q Q = Market value / Replacement valueQ = Market value / Replacement value
Reflects the expected current and future Reflects the expected current and future profitability of capital profitability of capital
Q values less than one identify under Q values less than one identify under valued equitiesvalued equities
Q values greater than one suggest than Q values greater than one suggest than capex will increase share holder wealthcapex will increase share holder wealth
Q values less than one suggest making Q values less than one suggest making acquisitions is cheaper than capexacquisitions is cheaper than capex
Valueline Timeliness RanksValueline Timeliness Ranks (1965 – 2009)(1965 – 2009)
Rank Weekly (%) Yearly (%)
1 15,575 30,778
2 10,727 4,174
3 4,924 252
4 2,846 - 60
5 5,266 -99
DatabaseDatabase
2008 (4) – 2010(1) – 6 Quarters2008 (4) – 2010(1) – 6 Quarters SourcesSources
Value Line Investment SurveyValue Line Investment Survey Ford Equity ResearchFord Equity Research Mergent OnlineMergent Online
Sample Size (100 ~ 400)Sample Size (100 ~ 400) Target Variable – PGQ (binary- lagged)Target Variable – PGQ (binary- lagged)
Two Step Analytic ProcessTwo Step Analytic Process
Screen variables with Screen variables with neutral netsneutral nets
Develop decision rules Develop decision rules using CARTusing CART
Holdout AssessmentHoldout Assessment
Probabilistic Neural NetworksProbabilistic Neural Networks
An extension to the classical backward An extension to the classical backward propagation neural netpropagation neural net
Non-parametricNon-parametric “ “Black Box”Black Box” Results often difficult to interpret and Results often difficult to interpret and
operationalizeoperationalize
CARTCART Non-parametricNon-parametric Interactive effectsInteractive effects Non-normally distributed variablesNon-normally distributed variables Decision tree logic makes it easier to Decision tree logic makes it easier to
apply model outcomesapply model outcomes Model is extremely robust to the effect of Model is extremely robust to the effect of
outliersoutliers Results easy to interpret and implementResults easy to interpret and implement
Neural Net ResultsNeural Net Results
Rank 8-4 9-1 9-2 9-3 9-4 10-1
1 PSS ROA PSS SMO CNE PRM
2 PRM SUE PVA PSS EMO Q
3 PVA PSS SEP ROA SMO ROA
4 ROA SMO PRM EMO VMO VMO
5 SEP EMO SMO SEP PEG EMO
6 VMO SEP Q Q Q SMO
7 EMO PRM VMO PRM SUE PSS
8 SUE PVA PEG EMO PRM PER
9 PEG PEG EMO PEG PSS SEP
10 SMO VMO ROA PVA SEP COM
Classification AnalysisClassification Analysis(9/4 -> 10/1)(9/4 -> 10/1)
Actual
Predicted 1 0
1 31 15 67% PPV1
0 16 33 67% NPV2
Total 47 48
66% 69%
Sensitivity Specificity
1PPV = ratio of the number of winners classified correctly divided by the total number of securities classified as winners.2NPV = ratio of the number of losers classified correctly divided by the total number of securities classified as losers.
ResultsResults(Modified Sharp Ratio)(Modified Sharp Ratio)
Case Qtrs./Sample
Size
Quarter Value Line Ones
Going Long
NSI Selling Short
NSI
1 1/89 9-2 0.289 0.392 38 0.210 53
2 1/91 9-3 0.775 0.853 51 -0.022 37
3 1/88 9-4 1.177 0.771 53 -0.043 40
4 1/93 10-1 0.513 0.553 38 0.485 56
5 1/94 10-2 -0.580 -0.328 46 -0.583 49
6 2/180 9-3 0.775 0.800 23 0.789 65
7 2/179 9-4 1.177 0.598 62 0.749 31
8 2/181 10-1 0.513 0.514 49 0.512 45
9 2/187 10-2 -0.580 -0.498 59 -0.728 36
10 4/361 10-1 0.513 0.613 49 0.418 45
11 4/366 10-2 -0.580 -0.493 70 -0.605 25
ConclusionsConclusions Modeling approach generally Modeling approach generally
performed as well or better than performed as well or better than Valueline 100Valueline 100
CART results provide an CART results provide an operational strategyoperational strategy
Adding transaction costs reduces Adding transaction costs reduces model effectivenessmodel effectiveness
Portfolio size based on binary Portfolio size based on binary target variable remains target variable remains problematicalproblematical
Future ResearchFuture Research
Expand data set from 6 to 12 Expand data set from 6 to 12 quartersquarters
Ternary classification targetTernary classification target Variable selection optimizationVariable selection optimization Add economic factorsAdd economic factors
CPICPI UEMUEM
Explore “super” factorsExplore “super” factors Q / ApEnQ / ApEn PRM / SpEnPRM / SpEn