A New Look to Three-Factor Fama-French Regression Model using Sample Innovations
AA New Look to Three-Factor Fama-French Regression Model usingSample Innovations
Javad Shaabani ∗ Ali Akbar Jafari † Abstract
The Fama-French model is widely used in assessing the portfolio’s performance compared tomarket returns. In Fama-French models, all factors are time-series data. The cross-sectional dataare slightly different from the time series data. A distinct problem with time-series regressions isthat R-squared in time series regressions are usually very high, especially compared with typicalR-squared for cross-sectional data. The high value of R-squared may cause misinterpretationthat the regression model fits the observed data well, and the variance in the dependent variableis explained well by the independent variables. Thus, to do regression analysis, and overcomewith the serial dependence and volatility clustering, we use standard econometrics time seriesmodels to derive sample innovations. In this study, we revisit and validate the Fama-Frenchmodels in two different ways: using the factors and asset returns in the Fama-French model andconsidering the sample innovations in the Fama-French model instead of studying the factors.Comparing the two methods considered in this study, we suggest the Fama-French model shouldbe consider with heavy tail distributions as the tail behavior is relevant in Fama-French models,including financial data, and the QQ plot does not validate that the choice of normal distributionas the theoretical distribution for the noise in the model.
Keywords:
Fama-French model; asset return modeling; sample innovations; econometrics. ∗ Department of Statistics, Yazd University, Yazd, Iran, [email protected]. † Department of Statistics, Yazd University, Yazd, Iran, [email protected]. a r X i v : . [ q -f i n . S T ] J un Introduction
Sharpe (1964) (Nobel laureate in economics) and Lintner (1965) introduced the capital asset pricingmodel (CAPM) to measure the investment risk and the return of an individual stock. The CAPMgives sturdy and intuitively pleasing forecasts about how to asses risk and the relation betweenrisk and expected return. Although the expected return of the market and market risk are thetwo factors taking into account in the CAPM model, the CAPM model employs only one variable,market return, to explain the stock returns. The standard formula of the CAPM is r t − r f,t = α + β ( r m,t − r f,t ) + (cid:15) t (1)where r is return on an asset; r f is risk-free rate; r m is the return of the market; β is the sensitivityof the asset returns to the excess market returns ( r m − r f ); α is the unexplained expected returnby asset; and (cid:15) is the market noise (error term associated with the excess return of the asset).The CAPM was widely used in assessing the portfolio’s performance compared to market returnsand measuring risk, beta, and alpha among practitioners. Many academic studies have determineduncertainty on reliability of the CAPM. FAMA and FRENCH (1992) by studying the share returnson the American Stock Exchange, the New York Stock Exchange, and Nasdaq observed thatdifferences in β over a lengthy period did not describe the performance of different stocks. Also,the linear relation between stock returns and β breaks down over shorter periods. These findingsimply the weaknesses of the CAPM model in applications.Eugene and Kenneth (1993) extended the CAPM model to create a better tool to evaluate port-folio performance and to predict the stock returns by adding two more factors to the CAPM. Thefactors in Fama-French model are (1) market risk, (2) the outperformance of small-cap companiesover large-cap companies, and (3) the excess returns of high book-to-price ratio companies versussmall low book-to-price ratio companies. The reason behind the Fama-French three-factor model isthat high value and small-cap companies tend to beat the overall market usually. The Fama-Frenchthree-factor’s representation model is r t − r f,t = α + β ( r m,t − r f,t ) + β SM B t + β HM L t + (cid:15) t (2)where SM B stands for “Small market capitalization Minus Big market capitalization” and
HM L for “High book-to-market ratio Minus Low book-to-market ratio;” they measure the historic excessreturns of small caps over big caps and of value stocks over growth stocks.In Fama-French models, all factors are time-series data. The cross-sectional data are slightlydifferent from the time series data. A distinct problem with time-series regressions is that R-squaredin time series regressions are usually very high, especially compared with typical R-squared for cross-sectional data. However, it does not imply that we learn more about factors impacting responsevariable from time-series data. One reason for artificially high values of R-squares and adjustedR-squares for time series regressions is because the response variable and dependent variables havetrends over time. A high R-squared can occasionally indicate that there is a problem with yourmodel. It may cause misinterpretation that the regression model fits the observed data well, andthe variance in the dependent variable is explained well by the independent variables. See . The book-to-bill ratio is the ratio of orders received to the amount billed for a specific period, usually one monthor one quarter. See https://en.wikipedia.org/wiki/Book-to-bill_ratio . (cid:15) t is not correlated with the explanatory variablesin any periods and (2) (cid:15) t is also uncorrelated with its past and future values (no autocorrelation).Thus we should text these assumptions before performing regression analysis.The hypotheses of independent and identically distributed for r t , r f , SM B , and
HM L arerejected. The Ljung-Box Q-test (Ljung and Box, 1978) indicates the presence of autocorrelationand heteroskedasticity in data set. The Engle-Granger co-integration (Fuller, 2009) test rejects thenull hypothesis of no co-integration among the time series. The explanatory variables drop therandom sampling, endogeneity, and homoscedasticity assumptions in the Gauss-Markov Theorem.Therefore, to do regression analysis, and overcome with the serial dependence and volatilityclustering, we use standard econometrics time series models to derive sample innovations. Insteadof studying the factors and asset returns, we consider their sample innovations derived from thetime series models. Consequently, we have iid standardized residuals for each factor as explanatoryvariables in the Fama-French models.Therefore, in this study, we revisit the Fama-French models in two different ways: • Using the factors and asset returns in the Fama-French model and validating the model byruing time series regression; • Filtering linear and nonlinear temporal dependencies in factors by applying time series model.Instead of studying the factors and asset returns in the model, considering the sample inno-vations for validating the Fama-French model; andWe use Microsoft’s stock data to test the robustness of the Fama-French (FF) models in explain-ing the variation in stock returns. The reason behind of choosing Microsoft stock is that Microsoftis the largest publicly traded US company and the top constituent of the S&P 500 index.The results from the QQ plot show the frequency of extreme events is higher than that implied bythe normal distribution. These result indicate that the distribution of the OLS regression estimatordoes not acquire the tail behavior. Tail behavior is relevant for regressions, including financial data.A potential reason for the significant variation in the estimated regression coefficients across time isthe heavy-tailed nature of the distribution of the innovations. It is an approved empirical fact thatmany financial variables are modeled better by distributions with tails heavier than the normaldistribution. The result of the two methods considered in this study, suggest the Fama-Frenchmodel should be consider with heavy tail distributions.We note that there is a comprehensive research that has attempted to model the tail behaviorof asset returns and to deal with non-normality of asset return. Modeling the tail behaviors ofasset returns is essential for risk managers. The the method of subordination (see Sato, 2002)widely proposed in literature to introduce additional parameters to the return model to reflect theheavy tail phenomena present in most asset returns and to generalize the classical asset pricingmodel (see Clark, 1973; Barndorff-Nielsen, 1977; Carr and Wu, 2004; Klingler et al., 2013; Shirvaniet al., 2020b). To incorporate the views of investors into asset return models and to deal withnon-normality of asset return, Shirvani et al. (2020a) a new process for asset returns in the form ofa mixed geometric Brownian motion and subordinated Levy process. In this study, to model the3sset return, we propose the Fama-French model with heavy tail distributions as the theoreticaldistribution for the noise in the model.There are three sections that follow in this paper. In the next section, Section 2, we describethe data source and data validation. Our methodology for modeling time series data is describedin Section 3. Section 4 compares the two methods used in this study. The results of each methodare reported in two subsections. Section 5 concludes the paper.
To examine the robustness of the Fama-French (FF) model in explaining the variation in stockreturns, we use the Microsoft stock price data as risky asset. The 10-year Treasury yield is usedas returns of the riskless asset. The historical data for factors in Fama-french model (SMB, HML,RMW, and CMA) are collected from the Kenneth French-Data Library. The data set are monthlyin the period from om March 1986 to February 2020. The factors in the Fama-French model withthe abbreviated names are follows. Market risk premium . Market risk premium (MRP) is the difference between the expected re-turn of the market and the risk-free rate. It provides an investor with an excess return ascompensation for the additional volatility of returns over and above the risk-free rate.
Small Minus Big . Small Minus Big (SMB) is a size effect based on the market capitalization of acompany. SMB measures the historic excess of small-cap companies over big-cap companies.
High Minus Low . High Minus Low (HML) is a value premium. It represents the spread in returnsbetween companies with a high book-to-market value ratio (value companies) and companieswith a low book-to-market value ratio.
The most important of the functional forms that we use in time series regressions is the naturallogarithm. As the time series data are not statistically significant, we account this change to removea time trend. In our analysis, the returns of the risky asset are obtained by the logarithmic returnas follows r ( t ) = ln S ( t ) S ( t − , (3)where S ( t ) is the price of Microsoft stock at day t .The Autoregressive Moving Average (ARMA) (Fuller, 2009) and the
Autoregressive ConditionalHeteroscedasticity (GARCH) model (Hamilton, 1994) are the standard tools for modeling the mean . https://ycharts.com/indicators/10_year_treasury_rate . https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html . See https://corporatefinanceinstitute.com/resources/knowledge/finance/fama-french-three-factor-model/ . , ,
1) with normal distribution assumption for innovations as follows r t = µ + ϕ ( r t − − µ ) + θa t − + a t ,a t = ε t σ t , ε t ∼ iid ,σ t = γ + αa t − + βσ t − , (4) µ t and σ t are the conditional mean and volatility of returns for stock and bonds, ε t is standardized iid –normal random variable, a t is referred to the market shocks, α ≥ , β ≥ , γ ≥ , δ, ϕ, θ arethe model parameters. We fit ARMA(1 , ,
1) with normal distribution assumption forinnovations on each factor. The model parameters are estimated with the use of the R-Package “rugarch” (Ghalanos et al., 2019). Instead of using the factors and asset returns in the Fama-Frenchmodel, we consider the sample innovations obtained from each time series model.
The following time-series regression (Fama-French 3 factor model) is used in this section: r t − r f,t = α + β ( r m,t − r f,t ) + β SM B t + β HM L t + (cid:15) t As to conduct OLS stationary test is required, we used the Augmented Dickey-Fuller. Thep-values( < .
01) are about zero indicating stationary time series data set. As the data was trans-formed into the log-return, we expected to have a stationary time series. Brooks (2019) explainedthat non-stationary data is led to spurious regressions. Table 1 shows the average monthly rateof return for Microsoft stock and the standard deviation for factors. As we observed from Table1, the average monthly rate of return for market portfolio is higher than the Microsoft stock. Itindicates the market outperforms the big market capitalization stocks. The reason behind it isthat the companies with small market capitalization perform better than the significant marketcapitalization portfolios. EXR MRP SMB HML RMW CMAMean 0.0055 0.0065 0.0003 0.0013 0.0034 0.0024Standard deviation 0.0409 0.0439 0.0309 0.0292 0.0245 0.0201Table 1: The average monthly rate of return for Microsoft stock and the standard deviation for thefactorsTable 2 reported the correlation coefficients between the explanatory variables and excess re-turn in the Fama-French 5 factor model. As the rule of thumb, the independent variables shouldnot be correlated or at least the correlation between independent variables should be low. In threefactor model (MRP, SMB and HML) we donot observe a high correlation value between the inde-pendent variables. In the three-factor model, all correlation coefficients are positive, and the lowestcorrelation observed between EXR and SMB. 5XR MRP SMB HMLEXR 1.000 0.569 0.080 -0.341MRP 0.569 1.000 0.216 -0.190SMB 0.080 0.216 1.000 -0.250HML -0.341 -0.190 -0.250 1.000Table 2: Correlation coefficients between the explanatory variablesTo visualize the pairwise correlations between the variables, We plotted the scatter plot andcorrelation heatmap in Figure 1. From Figure 1 we can discover more about the nature of theserelationships and the distribution’s shape of the variables. Although there are outliers in the databut there is not clustering by groups in the data. From Figure 1 we can quickly identify the relationbetween each factors in the model.
EXRN − . − . . . −0.2 −0.1 0.0 0.1 − . − . . . . −0.20 −0.10 0.00 0.10 Mkt.RF SMB −0.1 0.0 0.1 0.2 −0.10 0.00 0.05 0.10 − . − . . . − . . . . HML
EXRN Mkt.RF SMB HML −1.0−0.50.00.51.0
Figure 1: The scatterplot matrix and the correlation heatmap for Fama-French three factor model.
In this section we regress EXR on three factors with constant values added. The reason we addintercept to the regression model is that we need to make sure there is an intercept to confirm that6oefficients Standard Error t Stat P-valueIntercept 0.003 0.002 1.721 0.086MRP 0.504 0.037 13.467 0.000SMB -0.137 0.054 -2.527 0.012HML -0.370 0.057 -6.519 0.000Table 3: Regressions for the Fama-French three factor models explanatory variablesRegression StatisticsMultiple R 0.629R Square 0.395Adjusted R Square 0.389Standard Error 0.032Table 4: Regressions Statistics for the Fama-French three factor modelthe three factor model is correctly assess portfolios. If the intercept is omitted from the model, thismeans the model might not correctly evaluate excess return.Table 3 shows regressions statistics for the Fama-French three factor model’s explanatory vari-ables, where three factors explain the excess returns on Microsoft stock. In the regressions model,the intercept is not statistically significant (
P value = 0 . .
01 significant level. It is worth to note thatthe coefficients in regression model are close to the their correlation coefficients with the EXR. Theregression model suggests removing intercept and the SMB would not hurt the explanatory powerof the model if the EXR is regressed on the remaining two factors. The low value of R (0 . R values in stock market are less than 50%.Statistically significant coefficients continue to represent the mean change in the dependent variablegiven a one-unit shift in the independent variable. We note that in the Backward regression model,the R of two factor model with MRP and HML as explanatory variables is 0 .
70 179178 −0.10−0.050.000.050.10 −0.10 −0.05 0.00 0.05
Fitted values R e s i dua l s Residuals vs Fitted
170 179178 −4−2024 −2 0 2
Theoretical Quantiles S t anda r d i z ed r e s i dua l s Normal Q−Q
170 179178
Fitted values S t anda r d i z ed r e s i dua l s Scale−Location −4−2024 0.00 0.05 0.10
Leverage S t anda r d i z ed R e s i dua l s Residuals vs Leverage
Figure 2: Diagnostic plots for three factor regression model.The Scale-Location plot shows the residuals are spread equally along with the ranges of predic-tors. The plot shows the residuals appear randomly spread. From the Residuals vs. Leverage, wecan see some influential cases. Although we try the new model by excluding them, the regressionresults did not be altered when we exclude those cases. If we exclude the outliers case from theanalysis, the R changes from 0 .
395 to 0396 with small changes in coefficients. To see the cases out-side of a dashed line, we plot Cook’s distance in Figure 3. We note that the spread of standardizedresiduals changes as a function of leverage in all cases.8 −4−2024 0.00 0.05 0.10
Leverage S t anda r d i z ed R e s i dua l s Residuals vs Leverage
Leverage C oo k ' s d i s t an c e Cook's dist vs Leverage
Figure 3: Diagnostic plots for three factor regression model.We can test for the assumption that the error terms are are correlated with each other by usingDurbin-Watson, d, test statistic (see Durbin and Watson, 1992). Generally, a Durbin-Watson resultbetween 1.5 and 2.5 indicates, that any autocorrelation in the data will not have a discernible effecton your estimates. The test statistic values for our model is 2 . (cid:39)
0) we strongly reject the null hypothesis of normally distributed errors. Our residualsare not, according to our visual examination and this test, normally distributed.To check for multicollinearity, We use the Variance Inflation Factor. A general rule of thumb isthat
V IF > .
07, 1 .
11, and 1 . In time-series regression analysis, the dependent behavior of explanatory variables drops the randomsampling, Endogeneity, and homoscedasticity assumptions in the Gauss-Markov Theorem. Here toregress the EXR on three factors, we use standard econometrics models ARMA(1,1)-GARCH(1,1)with normal distribution to filter out the serial dependence and volatility clustering in each factor.Instead of studying factors and asset returns, we consider their sample innovations obtained fromthe time series model. The innovations are iid standardized residuals. Consequently, we have iidstandardized residuals for each factor as explanatory variables in the Fama-French models. Usingsample innovations instead the time series data set widely has been used in finance and economicto study the behavior of stock market and asset return process (see J¨aschke et al., 2011; Zhu et al.,2016; Shirvani, 2020). In electrical engineering, Ghaedi et al. (2016) used this method to studyhybrid electric vehicles. 9XRN MRP SMB HML RMW CMAEXRN 1.000 0.531 0.032 -0.281 -0.062 -0.353MRP 0.531 1.000 0.249 -0.208 -0.285 -0.307SMB 0.032 0.249 1.000 -0.120 -0.377 -0.092HML -0.281 -0.208 -0.120 1.000 0.035 0.622RMW -0.062 -0.285 -0.377 0.035 1.000 -0.002CMA -0.353 -0.307 -0.092 0.622 -0.002 1.000Table 5: Correlation coefficients between the innovations of explanatory variablesCoefficients Standard Error t Stat P-valueIntercept 0.1318 0.0676 1.950 0.0526MRP 0.3442 0.0729 4.719 0.0000SMB -0.1222 0.0659 -1.851 0.0657HML -0.4174 0.0782 -5.338 0.0000Table 6: Regressions for the innovation of the Fama-French three factor models explanatory vari-ables.Table 5 reports the correlation coefficients between the innovations of explanatory variables andexcess return in the Fama-French 5 factor model. As the rule of thumb, the independent variablesshould not be correlated, or at least the correlation between independent variables should below. Inthe three-factor model (MRP, SMB, and HML), we do not observe a high correlation value betweenthe independent variables. The lowest correlation coefficient is found between SMB and EXRN.By comparing the Tables 2 and 5, we observe the correlation coefficient between innovations areweaker than real data set.The scatter plots and correlation heatmap are plotted in Figure 4 to visualize the pairwisecorrelations between the variables. From Figure 4, we can discover more about the nature of theserelationships and the distribution’s shape of the variables. Again we observe outliers in the data,but there is not clustering by groups in the data. From Figure 4, we can quickly identify therelationship between each factor in the model.We regress innovations of EXR on innovations of three factors with constant values added.Table 6 shows regressions statistics for the Fama-French three-factor model’s explanatory variables,where the innovations of three factors explain the returns on Microsoft stock. Again the intercept isnot statistically significant ( P − value = 0 . R of innovation model (0 . . XRN − − −3 −2 −1 0 1 2 3 − − − −4 −2 0 2 MRP SMB −4 −2 0 2 4 −3 −2 −1 0 1 2 3 − − − − − HML
EXRN MRP SMB HML −1.0−0.50.00.51.0
Figure 4: The scatterplot matrix and the correlation heatmap for innovations of Fama-French threefactor model.As generally, the R values in the stock market are less than 50%, the variation of the EXR can beexplained well by the explanatory variables.Now, we go through the diagnostic plots for our model to check if the linear regression assump-tions are met. The diagnostic plots are shown in Figure 5. The QQ plot does not show a good fitand indicates skew distributions. Still, there is a problem in fitting on tails. It seems an asymmetrywith heavy tail distribution should be assumed here. The plot comes close to a straight line, exceptpossibly for the rear events, where we find a couple of residuals somewhat larger than expected.Thus, the normal assumption is not a proper assumption for both models.The Residuals vs. Fitted plot shows the residuals have non-linear patterns but not a distinctpattern. It seems that the spread of residuals around the zero line is equal. Although the spreadof residuals around the horizontal line has a negative trend for the negative fitted values, again, wecan not indicate non-linear relationships and rejecting the assumption of equal error variances forresiduals.Although the Scale-Location plot shows the residuals are not spread equally along with theranges of predictors, it shows the residuals appear randomly spread. From the Residuals vs. Lever-age plot, we can see some influential cases. We examined a new model by excluding influential cases.The regression results did not be altered when we exclude those cases. If we exclude the outliers11egression StatisticsMultiple R 0.5933R Square 0.3521Adjusted R Square 0.3410Standard Error 0.9051Table 7: Regression statistics for the innovaitons of three factor model.from the analysis, the R increases from 0 . . (cid:39)
0) we strongly reject the null hypothesis of normally distributed errors. Our residuals arenot, according to our visual examination and this test, normally distributed.To check for multicollinearity, We use the Variance Inflation Factor. A general rule of thumbis that
V IF > .
11, 1 . .
051 respectively. Thus, we are well within acceptable limits on VIF.
In this study, we revisited and validated the Fama-French models in two different ways: Using thefactors and asset returns in the Fama-French model and considering the sample innovations in theFama-French model instead of studying the factors. R of the model with the factors is (0 . R (= 0 . R of two-factor models with MRP and HML as explanatory variables are 0 . .
347 for thefactors and innovation models, respectively. The QQ plots do not show a good fit and indicate skewdistributions for both models. These plots demonstrated that the normal assumption is not a properassumption for both models. The Scale-Location plot shows the residuals are spread equally alongwith the ranges of predictors. The assumption that the error terms are correlated with each othertested by Durbin-Watson for both models. The test statistic values for both models demonstratedthat three is no autocorrelation problem with these models. The Jarque-Bera normality test withsmall p-values about zero strongly rejected the null hypothesis of normally distributed errors inboth models. We used the Variance Inflation Factor to check for multicollinearity, and results werewell within acceptable limits on VIF.Comparing the two methods considered in this study, we suggest the Fama-French model shouldbe consider with heavy tail distributions because • Tail behavior is relevant for regressions, including financial data,12 −3−2−10123 −3 −2 −1 0 1 2
Fitted values R e s i dua l s Residuals vs Fitted −2024 −2 0 2
Theoretical Quantiles S t anda r d i z ed r e s i dua l s Normal Q−Q
Fitted values S t anda r d i z ed r e s i dua l s Scale−Location −2024 0.000 0.025 0.050 0.075 0.100
Leverage S t anda r d i z ed R e s i dua l s Residuals vs Leverage
Figure 5: Diagnostic plots for three factor regression model. • QQ plot does not validate that the choice of normal distribution as the theoretical distributionfor the noise in our model, • In finance, the investors always try to estimate the tail, and hedge against tail risk aims toimprove returns over the long-term.Finally, we note that our study was based on the Microsoft stock returns process, and we letthe data speak for itself. Future research first can be conducted for the market indices, especially13 −2024 0.000 0.025 0.050 0.075 0.100
Leverage S t anda r d i z ed R e s i dua l s Residuals vs Leverage
Leverage C oo k ' s d i s t an c e Cook's dist vs Leverage
Figure 6: Diagnostic plots for three factor regression model.for S&P 500, Dow Jones Industrial Average index, and other stocks. It also can be generalizedfor other indices, especially for Real Estate indices. Secondly, we recommend reviewing the Fama-French five-factor model by considering the sample innovations instead of studying the factors.
References
Barndorff-Nielsen, O. (1977). Normal inverse Gaussian distributions and stochastic volatility mod-eling.
Scandinavian Journal of Statistics , 24:1–13.Brooks, C. (2019).
Introductory econometrics for finance . Cambridge university press.Carr, P. and Wu, L. (2004). Time-changed L´evy processes and option pricing.
Financial Economics ,17:113–141.Clark, P. (1973). A subordinated stochastic process model with fixed variance for speculative prices.
Econometrica , 41:135–156.Durbin, J. and Watson, G. S. (1992). Testing for serial correlation in least squares regression. i. In
Breakthroughs in Statistics , pages 237–259. Springer.Eugene, F. and Kenneth, F. (1993). Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics .FAMA, E. F. and FRENCH, K. R. (1992). The cross-section of expected stock returns.
The Journalof Finance , 47(2):427–465. 14uller, W. A. (2009).
Introduction to statistical time series , volume 428. John Wiley & Sons.Ghaedi, A., Dehnavi, S. D., and Fotoohabadi, H. (2016). Probabilistic scheduling of smart elec-tric grids considering plug-in hybrid electric vehicles.
Journal of Intelligent & Fuzzy Systems ,31(3):1329–1340.Ghalanos, A., Ghalanos, M. A., and Rcpp, L. (2019). Package “rugarch”.Hamilton, W. A. (1994).
Time series analysis . Princeton University Press, Princeton, N.J.Jarque, C. M. and Bera, A. K. (1980). Efficient tests for normality, homoscedasticity and serialindependence of regression residuals.
Economics letters , 6(3):255–259.J¨aschke, S., Siburg, K. F., and Stoimenov, P. A. (2011).
Modelling dependence of extreme eventsin energy markets using tail copulas . Universit¨atsbibliothek Dortmund.Klingler, S., Kim, Y., Rachev, S. T., and Fabozzi, F. J. (2013). Option pricing with time-changedL´evy processes.
Applied Financial Economics , 23(15):12–31.Lintner, J. (1965). The valuation of risk assets and the selection of risky investments in stockportfolios and capital budgets.
The Review of Economics and Statistics , 47(1):13–37.Ljung, G. M. and Box, G. E. (1978). On a measure of lack of fit in time series models.
Biometrika ,65(2):297–303.Sato, K. (2002).
Levy Processes and Infinitely Divisible Distributions . Cambridge University Press,New York.Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions ofrisk.
The Journal of Finance , 19(3):425–442.Shirvani, A. (2020). Stock returns and roughness extreme variations: A new model for monitoring2008 market crash and 2015 flash crash.
Applied Economics and Finance , 7(3):78–95.Shirvani, A., Hu, Y., Rachev, S. T., and Fabozzi, F. J. (2020a). Option pricing with mixedL´evy subordinated price process and implied probability weighting function.
The Journal ofDerivatives .Shirvani, A., Rachev, S. T., and Fabozzi, F. J. (2020b). Multiple subordinated modeling of assetreturns: Implications for option pricing.