2 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema 1 Readings
• Read BKM Ch 13
3 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema 2 What are we doing today
• Looking carefully at the evidence in support of a few models of
expected returns – CAPM – Multi-factor models – Other attempts
4 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema 3 CAPM evidence
• Ok, the CAPM states that:
• E[ri] = rf + βi(E[rm − rf ]) • How should we test it? • The problem here is that both E[ri] and E[rm −rf ] are not directly observable. – We cannot observe the expected return for Microsoft stock as of time t – Nor can we observe the expected return of the market premium (rm − rf )
5 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• So, we have to substitute actual returns for expected returns. This
is far from a perfect solution. Realised returns are VERY noisy compared to expected returns – To make this concrete: The expected excess return to a invest- ment with positive exposure to systematic risk should always be positive – However, out of the 3,575,388 monthly excess returns in the CRSP monthly data, 1,836,849 are in fact negative.
. count i f ! m i ss in g ( r ) 3 ,575 ,388 . count i f ! m i ss in g ( r ) & e r e t < 0 1 ,836 ,849
• (I am showing you a bit of Stata code along the way, just so you can get used to the idea that you can write code to do empirical work)
6 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• But with that caveat / limitation in mind, let’s proceed
• first estimate CAPM beta’s on 60 months of historical data → Bi,t = βdi,t – For each stock i at each time t over historical window t−1, t−60 (past 60 months) • Then regress excess returns (ri,t) on CAPM beta’s (Bi,t) – ri,t = α + γBi + ε
7 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• CAPM theory suggests that we should find α = 0 and γ = E[rm −
rf ] • This is what we actually get . reg r CAPMbeta , robust
Linear regression Number of obs = 2 ,337 ,301
F (1 , 2337299) = 3.71 Prob > F = 0.0541 R - squared = 0.0000 Root MSE = .16123
• The coefficient on CAPMbeta (γ) is -3 basis points per month,
which is much less than the 57 basis points per month earned by the broad market. So γ ̸= E[rm − rf ]. The relationship is not even positive (when it should be).
8 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• The intercept (_cons in the regression output, α in the equation)
should be (close to) zero, but it is 98 basis points per month. So α ̸= 0 with a t-statistic of 44.4 (very high statistical significance) • The R2 is less than 0.01%, meaning that more than 99.9% of the variation in realised returns cannot be explained by the CAPM model. • The hypothesis that the CAPM describes realised returns is rejected by the data. (Data can never confirm a hypothesis – however it can fail to reject it, but in this case it does not fail to reject it, that is, it rejects it.) • There is more. According to the CAPM, nothing except the CAPM beta should explain expected returns. The the coefficient on other explanatory variables should be zero. • Just for fun, let’s include a few other explanatory variables
9 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• So, variables other than CAPM beta does to seem help. But of course, they are not supposed to.
10 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• The R2 is now almost 0.1%. (Yay! We can explain one tenth of a
percent of the variation in monthly stock returns.) • The above analysis is not conclusive, BTW. – The CAPM beta is measured with error, which technically in- validates the OLS regression – This can be mitigated by testing on portfolios of stocks (so the measurement error is reduced), but that does not change the story • Roll’s critique – https://en.wikipedia.org/wiki/Roll%27s_critique – The true market portfolio is unobservable – Problematic, because that means technically testing the exact CAPM is impossible – As per Karl Popper, a theory that is incapable of being proven wrong – no matter what data is provided – no longer counts as science (the falsification requirement) 11 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema CAPM evidence (cont.)
• So where does this leave the CAPM
– It is a beautiful, elegant theory that describes how expected returns should behave (should → normative theory) – It does not describe how actual returns do behave (do → de- scriptive theory)
12 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema 4 Variations inspired by CAPM
• There has been a lot of work trying to resuscitate the CAPM
• It has led to more beautiful theory (that still does not really work) and some ugly ad-hoc factor models (that work somewhat better) • The theories – Consumption CAPM – Conditional CAPM – Intertemporal CAPM – Extensions to CAPM that include ◦ Liquidity ◦ Labour income ◦ Private business ◦ Macro factors (innovations in interest rates, credit spreads, industrial production, inflation)
13 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema Variations inspired by CAPM (cont.)
• The factor models
Market model (MKT): ri,τ = αtM KT + βtM KT M KTτ + εM i,τ KT for τ in window [t − w + 1, t]
CAPM model (CAPM):
ri,τ = αtCAP M + βtCAP M M KT RFτ + εCAP
i,τ M for τ in window [t − w + 1, t]
Fama-French 3-factor model (FF3):
ri,τ = αtF F 3 + βtF F 3M KT RF f f 3 M KT RF f f 3τ + βtF F 3SM Bf f 3 SM Bf f 3τ
+ βtF F 3HM Lf f 3 HM Lf f 3τ + εF F3 i,τ for τ in window [t − w + 1, t]
Note: the factors used in the FF3 model is suffixed by “ff3” to distinguish them from the factors used in the Fama-French 5-factor model. Fama-French-Cahart 4-factor model (FFC4):
ri,τ = αtF F C4 + βtF F C4M KT RF f f 3 M KT RF f f 3τ + βtF F C4SM Bf f 3 SM Bf f 3τ
+ βtF F C4HM Lf f 3 HM Lf f 3τ + βtF F C4U M D U M Dτ + εF F C4 i,τ for τ in window [t − w + 1, t]
Hou-Xue-Zhang q-factor model (HXZ4):
ri,τ = αtHXZ4 + βtHXZ4M KT M KTτ + βtHXZ4M E M Eτ
+ βtHXZ4I2A I2Aτ + βtHXZ4ROE ROEτ + εHXZ4 i,τ for τ in window [t − w + 1, t]
Fama-French 5-factor model (FF5):
ri,τ = αtF F 5 + βtF F 5M KT RF M KT RFτ + βtF F 5SM B SM Bτ
+ βtF F 5HM L HM Lτ + βtF F 5CM A CM Aτ + βtF F 5RM W RM Wτ + εF F5 i,τ for τ in window [t − w + 1, t]
14 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema Variations inspired by CAPM (cont.)
Hou-Xue-Zhang q5 5-factor model (HXZ5):
ri,τ = αtHXZ5 + βtHXZ5M KT M KTτ + βtHXZ5M E M Eτ
+ βtHXZ5I2A I2Aτ + βtHXZ5ROE ROEτ + βtHXZ5EG EGτ + εHXZ5 i,τ for τ in window [t − w + 1, t]
Fama-French 6-factor model (FF6):
ri,τ = αtF F 6 + βtF F 6M KT RF M KT RFτ + βtF F 6SM B SM Bτ
+ βtF F 6HM L HM Lτ + βtF F 6CM A CM Aτ + βtF F 6RM W RM Wτ + βtF F 6U M D U M Dτ + εF F6 i,τ for τ in window [t − w + 1, t]
The Geertsema-Lu (2020) Search Generated model (SG9):
+ βtSG9SM B SM Bτ + βtSG9IssuanceAndY ield IssuanceAndY ieldτ + βtSG9M KT RF M KT RFτ + βtSG9ST REV ST REVτ + βtSG9Seasonality Seasonalityτ + βtSG9CapexGrowth CapexGrowthτ βtSG9epsconsistency epsconsistency + εSG9 i,τ for τ in window [t − w + 1, t]
• The last 9 factor model explains away all of the 80 statistically
significant anomalies considered in that study (many more anom- alies are not statistically significant). This provides a sense of the “dimensionality” of the return generating process underlying stock returns.
15 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema 5 Final thoughts
• Why do we bother doing science?
• Scientists generally do science because they are curious. They want to know how things works. Finding the answer is the reward. – Data => Hypothesis => Empirical Testing => Model => Prediction – Scientists value the model, which encapsulates understanding, much more than the actual predictions • Society is willing to fund science because it works. To date, science has been the most successful approach to generating predictions that work. – Data => Hypothesis => Empirical Testing => Model => Prediction – Society values good predictions, but does not particularly care about the models 16 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema Final thoughts (cont.)
• We are now witnessing a large scale disruption in science, with
machine learning displacing theoretical models • With machine learning (ML), one can go from data directly to prediction – there is no need for hypotheses or models. – Data => ML => Prediction (note, no model or hypothesis) – In many fields, particularly those where simple theoretical mod- els cannot accommodate the true complexity of the data gen- erating process, we find that ML outperform theoretical models in terms of accurate prediction. • This is exactly what is happening in finance
17 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema Final thoughts (cont.)
• If you really want to predict actual returns, machine learning – not
finance theory – is your best bet* – *Augmented with human understanding of how the world works. Tesla is valuable because people understand that electric vehicles are likely to displace internal combustion vehicles in the future. But ML can only see what is in the past data, which does not include a world filled with electric vehicles. • This does not mean finance theory is not important; even if you use machine learning, an understanding of finance is essential if you want to avoid mistakes. (For instance, adjusting for de-listing returns or accounting for bid-ask bounce in illiquid securities) • This is the reason I’ll cover financial machine learning in a later topic.
18 FINANCE 361 Class Notes – University of Auckland – Copyright (C) Dr Paul Geertsema
