ROE and Fama-French 7/26/2016 Setting the Stage The Analysis The Results Conclusions 2 Setting the Stage • Insurance companies require capital in order to function. • Raising and holding capital costs money. • The cost of capital is a necessary part of the insurance rate. Per ASOP 30: “Property/casualty insurance rates should provide for all expected costs, including an appropriate cost of capital associated with the specific risk transfer.” 3 Setting the Stage What is an appropriate cost of capital? Per the Supreme Court: “[T]he return to the equity owner should be commensurate with returns on investments in other enterprises having corresponding risks. That return, moreover, should be sufficient to assure confidence in the financial integrity of the enterprise, so as to maintain its credit and to attract capital.”* *Hope Natural Gas, 320 U.S. at 603 (citations omitted) 4 Setting the Stage In Actuarial Pricing, we are used to having to project future, unknown quantities, but the cost of raising and maintaining capital is more difficult to quantify than most other expenses. Q: How do you know if you are accurately estimating the company’s cost of capital? 5 Setting the Stage Cost of capital discussions often suffer from lack of context. • Why should we target the return we are targeting? • What capital should that return be based on? And why? 6 Setting the Stage Allstate Example: • Allstate is a publicly owned company that raises capital by issuing shares of stock (and bonds). • Investors expect a return on their investment. • The return is expected on their investment. • The return should be commensurate with the risk. The goal: determine the return to shareholders (i.e. on the market value of the company) that is commensurate with the risk of investing in Allstate. 7 Setting the Stage Income to Shareholders is provided through Underwriting Income and Investment Income. Underwriting Income Income to Shareholders Income to Company Investment Income 8 Setting the Stage Pricing calculations work in reverse, by determining the appropriate income to Shareholders, then determining the needed Underwriting Income. Underwriting Income Income to Shareholders Income to Company Investment Income 9 Setting the Stage We need to determine what return on the market value (MV) of Allstate is commensurate with the risk. • There are many methodologies that have been developed to measure returns on MV. • Markowitz portfolio theory from the 1950s developed into the Capital Asset Pricing Model (CAPM) in the 1960s: 𝑟𝑖 = (𝑟𝑚 −𝑟𝑓 ) ∗ 𝛽𝑖 + 𝑟𝑓 • In the 1990s, CAPM expanded into various multi-factor models, including the Fama-French Three-factor Model (FF3F): 𝑟𝑖 = (𝑟𝑚 −𝑟𝑓 ) ∗ 𝛽𝑖 + 𝛽𝑠 ∗ 𝜋𝑠 + 𝛽ℎ ∗ 𝜋ℎ + 𝑟𝑓 10 Setting the Stage A common approach* for calculating the betas is a two-step process: • Step one – using Simple Linear Regression (SLR) calculate for each company a beta relative to the market based on 60 months of company-specific data: Company A: 𝑟𝐴 = (𝑟𝑚 −𝑟𝑓 ) ∗ 𝛽𝐴 + 𝛽𝑠−𝐴 ∗ 𝜋𝑠 + 𝛽ℎ−𝐴 ∗ 𝜋ℎ + 𝑟𝑓 Company B: 𝑟𝐵 = (𝑟𝑚 −𝑟𝑓 ) ∗ 𝛽𝐵 + 𝛽𝑠−𝐵 ∗ 𝜋𝑠 + 𝛽ℎ−𝐵 ∗ 𝜋ℎ + 𝑟𝑓 Company C: 𝑟𝐶 = (𝑟𝑚 −𝑟𝑓 ) ∗ 𝛽𝐶 + 𝛽𝑠−𝐶 ∗ 𝜋𝑠 + 𝛽ℎ−𝐶 ∗ 𝜋ℎ + 𝑟𝑓 etc. Note: at this point, all of the betas are company-specific. *This is the approach outlined in “ESTIMATING THE COST OF EQUITY CAPITAL FOR PROPERTY-LIABILITY INSURERS” by J. David Cummins and Richard D. Phillips, The Journal of Risk and Insurance, 2005, Vol. 72, No. 3, 441-478 11 Setting the Stage • Step Two – Use the betas calculated in Step One as the dependent variable in a system of linear equations. Using the market betas as an example: 𝛽𝐴 = 𝜔𝐴.𝑋 𝛽𝑋 + 𝜔𝐴.𝑌 𝛽𝑌 + 𝜔𝐴.𝑍 𝛽𝑍 𝛽𝐵 = 𝜔𝐵.𝑋 𝛽𝑋 + 𝜔𝐵.𝑌 𝛽𝑌 + 𝜔𝐵.𝑍 𝛽𝑍 𝛽𝐶 = 𝜔𝐶.𝑋 𝛽𝑋 + 𝜔𝐶.𝑌 𝛽𝑌 + 𝜔𝐶.𝑍 𝛽𝑍 In this example, X, Y, and Z represent industries, and ω represents the percentage of the firm’s sales in a given industry. Note: after Step Two, the calculated betas are now industry betas. This is called the “Full-information Beta” (FIB) approach. 12 Setting the Stage Q: How do we know if a methodology is any good? • There is significant debate in academic literature regarding whether the CAPM framework holds. • Articles defending specific methodologies generally use regression output, such as pvalues, to show that certain parameters are statistically significant, and therefore the methodology holds. • As a test, I checked all of the p-values from the first-run regression (SLR on a company-specific basis) to see how many of them were below 5% or 10%: Market Size Value Betas 311,564 311,564 311,564 P < .05 29.4% 13.8% 12.5% P < .10 36.6% 21.0% 19.3% 13 Setting the Stage Q: How do we know if one methodology is better than another? • The approach used to defend individual methodologies in literature (using regression output) is akin to fitting a model on a Training data set. • Modern-day model building typically does not stop there – after building a model on a Training data set, the model is then run on a Test data set to see how it performs. • Using this approach, we can see how well different methodologies perform relative to each other. 14 The Analysis Testing • Available data included monthly stock returns from 1989 through 2013. • The Training/Testing of the methodologies was designed to mirror our actual Pricing process – calculate the parameters based on N consecutive years of data; use the calculated parameters to predict a company’s risk premium in the following chronological year. • Example: Calculation data Test data 2000 2001 2002 2003 2004 2005 • Note: this is not a test of predicting the stock market, but rather a test of predicting a company’s return given the market/size/value return in that future period. 15 The Analysis • • Cummins/Phillips Methodology (Base Scenario) • First regression run – Simple Linear Regression (SLR) on 60 months of returns for each individual company, using the Sum Beta approach • Second regression run – Full-Information Beta calculation solving a system of Linear Equations, weighted by market value, for industry betas using Seemingly Unrelated Regression (SUR) Additional Options • Regression Techniques • Industry factors • Hierarchical model • Bayesian model • Regression + Shrinkage • Quantile regression • Detail Tweaks • Weighting • Months of data • Capping • Sum Beta • Seemingly Unrelated Regression 16 The Analysis • Pooling • When data can be categorized at multiple levels – in this case, by company and by industry – decisions need to be made regarding what level of pooling will be done in the regression analysis. • No pooling – data is not pooled by industry at all, and regression is run only at the individual-company level. • Complete pooling – data is not kept at the individual-company level at all, and regression is run entirely at the industry level. No Pooling Partial Pooling Complete Pooling SLR Hierarchical Models Bayesian Models FIB (sort of) SLR - Industry Factors • Note that CAPM testing has always been done using portfolios (not necessarily by industry) in order to minimize estimation error. 17 The Analysis • Regression Detail Tweaks • • • Weighting – for regressions that involve some pooling of the data, weighting the data by some measure (such as market value) might provide greater accuracy. Months of data • From the 2013 Ibbotson SBBI Valuation Yearbook: “The amount of history included in beta calculations done by commercial beta services is fairly consistent at five years. Using five years of data is a rather arbitrary decision that attempts to use as much data as possible without including irrelevant historical data. Using five years of data would ideally cover a number of different economic scenarios such as expansion and contraction in the economy.” • In my experience, most people use five years, and some use fewer than five. I’ve never seen anyone use more than five. Capping - Betas could be capped at some quantile to help eliminate the crazies. 18 The Analysis • Measurement Criteria • Accuracy • • After calculating the betas on the Training data set, we can calculate predictions on the Test data set and compare to the actual value. Accuracy can be measured based on Mean Squared Error or Mean Absolute Error on the Test predictions. • Stability • • • In ratemaking, we not only care about accuracy, but also the stability of our projections. • Stability can be measured based on the average absolute change in estimated betas and/or returns from one year to the next. Unfortunately, stability and accuracy tend to be opposed to each other. The Actuary needs to decide if the desire is to be most accurate, most stable, or have some blend of the two. 19 The Results Given all of the options mentioned above, there are too many scenarios to look at in one shot. However, we can make decisions on some of the smaller adjustments first: • Sum Beta – the Sum Beta approach is commonly used today, but in a comparison across methodologies, the use of this approach does not appear to impact accuracy much, but it results is much less beta stability. Sum Beta Methodology Impact Change in Change in Avg ABS Error Beta Stability FIB 0.002 0.073 SLR (CAPM) 0.001 0.060 SLR (FF3F) 0.004 0.127 Hierarchical Model 0.000 0.034 Better Same Worse 20 The Results • Seemingly Unrelated Regression (SUR) – The SUR approach is only needed when solving systems of linear equations. This is only relevant in the Full Information Beta (FIB) approach. It appears to have minimal impact. Seemingly Unrelated Regression Change in Change in Avg ABS Error Beta Stability FIB (0.000) (0.001) • Weighted Regression – weighting is more meaningful in methodologies that involve some level of pooling, though in most cases it appears to have minimal, and mixed, impact: Weighted Regression Change in Change in Avg ABS Error Beta Stability FIB 0.001 0.025 SLR - Industry Parameters 0.000 0.022 Hierarchical Model 0.000 (0.030) Better Same Worse 21 The Results • Years of Data – Despite common practice, using more than 5 years of data was uniformly better, with 10 years slightly outperforming 8 years. Years of Data 5-yr vs. 8-yr 5-yr vs. 10-yr Change in Change in Change in Change in Avg ABS Error Beta Stability Avg ABS Error Beta Stability FIB 0.000 (0.032) 0.000 (0.038) SLR (FF3F) (0.001) (0.116) (0.001) (0.146) SLR - Industry Parameters (0.000) (0.034) (0.000) (0.042) Hierarchical Model 0.000 (0.042) 0.000 (0.058) Quantile (0.001) (0.116) (0.001) (0.146) SLR + Shrinkage (0.000) (0.079) (0.000) (0.100) Better Same Worse 22 The Results Beta Capping – Testing showed that capping betas at some quantile (to rein in the tail values) improved the results in all cases. Of the quantile values that I tried, 90%/10% appeared to perform best. Beta Estimates - SLR 10-yr 25000 Beta Capping FIB 10-yr SLR 10-yr Quantile 10-yr SLR IND 10-yr SLR + Shrinkage 10-yr Hierarchical 10-yr Change in Change in Avg ABS Error Beta Stability (0.000) (0.016) (0.001) (0.040) (0.001) (0.058) (0.000) (0.015) (0.001) (0.030) (0.000) (0.028) 20000 15000 Frequency • Frequency 10000 5000 0 -5 -4.5 -4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 More Market Beta Better Same Worse 23 The Results Stability vs. Accuracy 0.500 0.450 FF3F SLR (5-yr) - Step 1 of Base Scenario 0.400 Avg Beta Change It can be helpful to compare methodologies by charting the measures of accuracy and stability on two axes of a graph. In this graph, lower on the yaxis and left on the x-axis are better. FF3F SLR 0.350 FIB Quantile 0.300 SLR Industry 0.250 Ridge Bayesian 0.200 FIB - Base Scenario 0.150 Quantile - 10-yr Capped 0.100 Lasso - 10-yr Capped Hierarchical CAPM MKT CAPM - 10-yr Capped Ridge - 10yr Capped Lasso Vasicek 0.050 0.0900 0.0920 0.0940 0.0960 Market 0.0980 0.1000 0.1020 0.1040 0.1060 0.1080 0.1100 Avg ABS Diff 24 The Results Things of note on the previous chart: • • • • • On average, there is nothing less accurate than the Base Scenario. With the benefit of additional years of data and capping, methodologies that use companyspecific data can be both more accurate and more stable than the current approach. Both CAPM and simply using the Market Betas perform surprisingly well. The methodologies that incorporate industry information perform surprisingly poorly. • This could be related to how we currently identify industry information in the data. The previous two points, plus the fact that the best-performing methodologies are those that incorporate parameter shrinkage, long data horizons, and extreme-event capping, suggest that over time company returns regress towards the market means, and many methodologies tend to over-fit based on short-term data. However, the results on the whole portfolio are not identical to the results on a specific company. As a larger, more-stable company, Allstate is probably more likely to be fit and predicted accurately by models (see next slide). 25 The Results Stability vs. Accuracy - Allstate 0.2000 FIB - Base Scenario 0.1800 Avg Beta Change 0.1600 FF3F SLR 0.1400 FIB Quantile 0.1200 SLR Industry 0.1000 0.0800 Ridge Quantile 10-yr Bayesian Hierarchical FF3F SLR - 10-yr Capped CAPM 0.0600 Hierarchical - 10yr Capped MKT 0.0400 Lasso Ridge - 10-yr Capped Vasicek 0.0200 0.0450 0.0460 0.0470 0.0480 0.0490 0.0500 0.0510 0.0520 0.0530 Market 0.0540 0.0550 Avg ABS Diff 26 The Results Things of note on the previous chart: • Market Betas or CAPM perform the worst, by far. • The current method also leaves opportunity for improvement. • For Allstate-specific data, the gains in accuracy are minimal, but the stability can be improved greatly. • Allstate does appear to track better with industry information than the average company does. • The best-performing methodologies are still those that incorporate shrinkage. 27 The Results The results also appear to be very consistent across years in terms of relative performance: Average Absolute Error 0.1800 0.1600 0.1400 0.1200 Ridge 0.1000 Quantile Bayesian 0.0800 Hierarchical FIB (Base Scenario) 0.0600 0.0400 0.0200 2013 2012 2011 2010 2009 2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 28 The Results Still, the methodologies can provide a range of answers for Allstate: Calculated Allstate Cost of Equity 20% 18% 16% Calculated Cost of Equity 14% SLR (5-yr) Bayes (10-yr) Capped 12% Ridge (10-yr) Capped FIB - Base Scenario 10% Hierarchical (10-yr) Capped 8% SLR (10-yr) Capped Quantile (10-yr) Capped 6% CAPM (10-yr) Capped 4% 2% 0% 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 29 Conclusions Final Thoughts: • We can easily improve on the base scenario. • The best methods appear to use company-specific information, along with a 10-year time horizon and Beta capping. • Using the average of multiple methodologies can help safeguard against the particularities of any one approach being overly impacted by some data quirk. • If averaging methodologies, it makes sense to select ones that are conceptually different from each other. • This analysis is just a starting point. There are many other modeling techniques that could be tested. 30