РАЗВЕДОЧНЫЙ АНАЛИЗ ДАННЫХ
diplom_1802_imp = read.csv('/Users/olesyamba/Downloads/data/diplom_1802_imp.csv')
diplom_1802_w_imp = read.csv('/Users/olesyamba/Downloads/data/diplom_1802_w_imp.csv')diplom_1802_imp = filter(diplom_1802_imp, year != "2016")
desc = stat.desc(diplom_1802_imp[, c(6, 7,10:20)])
desc = t(round(desc, digits = 2))
desc = desc[, c('nbr.val','min', 'max', 'mean', 'median', 'std.dev')]
desc## nbr.val min max mean median std.dev
## price_to_book 612 0.04 11.17 3.02 2.67 1.43
## polarity_news 612 0.03 0.14 0.09 0.09 0.01
## polarity_twitter 612 0.01 0.30 0.12 0.12 0.02
## yoy_revenue_growth 612 -11.55 261.18 19.02 16.64 27.85
## google_trends 612 -1021.73 683.00 167.66 164.36 102.03
## number_of_news 612 0.00 32.15 2.15 1.28 3.26
## number_of_likes_twitter 612 0.02 220.08 15.42 14.91 12.68
## number_of_tweets 612 0.28 487.69 76.10 72.89 38.56
## sales 612 -189.12 103.60 25.30 26.19 12.00
## ROA 612 -22.57 38.05 14.15 13.96 6.65
## fin_leverage 612 -105.15 43.95 0.56 0.45 6.60
## buyback_yield 612 -24.97 37.22 2.38 2.29 3.92
## log_sales 607 1.82 4.65 3.25 3.30 0.30
На графике представлены описательные статистики до предварительной обработки, включающей обработку пропущенных значений, обработку выбросов, а также уменьшение дисперсии некоторых переменных с целью борьбы с шумом.
diplom_1802_w_imp = filter(diplom_1802_w_imp, year != "2016")
desc = stat.desc(diplom_1802_w_imp[, c(6, 7,10:20)])
desc = t(round(desc, digits = 2))
desc = desc[, c('nbr.val','min', 'max', 'mean', 'median', 'std.dev')]
desc## nbr.val min max mean median std.dev
## price_to_book 612 0.04 11.17 3.02 2.67 1.43
## polarity_news 612 0.03 0.14 0.09 0.09 0.01
## polarity_twitter 612 0.01 0.30 0.12 0.12 0.02
## yoy_revenue_growth 612 -11.55 47.14 16.73 16.64 12.82
## google_trends 612 10.00 683.00 169.48 164.36 89.90
## number_of_news 612 0.00 32.15 2.20 1.38 3.28
## number_of_likes_twitter 612 0.02 220.08 15.42 14.91 12.68
## number_of_tweets 612 0.28 487.69 76.10 72.89 38.56
## sales 612 7.80 103.60 25.76 26.22 8.02
## ROA 612 -1.22 38.05 14.24 13.96 6.33
## fin_leverage 612 -5.96 18.82 1.03 0.45 3.15
## buyback_yield 612 -4.15 17.90 2.51 2.35 3.50
## log_sales 612 2.17 4.65 3.24 3.30 0.31
КОРРЕЛЯЦИОННЫЙ АНАЛИЗ
library("Hmisc")## Loading required package: lattice
##
## Attaching package: 'lattice'
## The following object is masked from 'package:regclass':
##
## qq
## Loading required package: survival
## Loading required package: Formula
##
## Attaching package: 'Hmisc'
## The following object is masked from 'package:parsnip':
##
## translate
## The following object is masked from 'package:simputation':
##
## impute
## The following objects are masked from 'package:dplyr':
##
## src, summarize
## The following objects are masked from 'package:base':
##
## format.pval, units
res2 <- rcorr(as.matrix(diplom_1802_w_imp[, c(6:20)]))matrix_r = as.data.frame(res2$r)
matrix_p = as.data.frame(res2$P)
empty_as_na <- function(x){
ifelse(is.na(x), as.integer('1'), x)
}
matrix_r = round(matrix_r, 3)
matrix_p = matrix_p %>% mutate_each(funs(empty_as_na)) ## Warning: `funs()` was deprecated in dplyr 0.8.0.
## ℹ Please use a list of either functions or lambdas:
##
## # Simple named list: list(mean = mean, median = median)
##
## # Auto named with `tibble::lst()`: tibble::lst(mean, median)
##
## # Using lambdas list(~ mean(., trim = .2), ~ median(., na.rm = TRUE))
## Warning: `mutate_each_()` was deprecated in dplyr 0.7.0.
## ℹ Please use `across()` instead.
## ℹ The deprecated feature was likely used in the dplyr package.
## Please report the issue at <�]8;;https://github.com/tidyverse/dplyr/issues�https://github.com/tidyverse/dplyr/issues�]8;;�>.
for (i in seq.int(1,15,1)){
for (j in seq.int(1,15,1)){
if (matrix_p[i,j] <= 0.01){
matrix_r[i,j] = str_c(as.character(matrix_r[i,j]), as.character('*'))
}
if (matrix_p[i,j] <= 0.05){
matrix_r[i,j] = str_c(as.character(matrix_r[i,j]), as.character('*'))
}
if (matrix_p[i,j] <= 0.1){
matrix_r[i,j] = str_c(as.character(matrix_r[i,j]), as.character('*'))
}
}
}matrix_r## price_to_book polarity_news subjectivity_news
## price_to_book 1 0.446*** 0.485***
## polarity_news 0.446*** 1 0.784***
## subjectivity_news 0.485*** 0.784*** 1
## pol_sub_news 0.361*** 0.728*** 0.726***
## polarity_twitter 0.193*** 0.353*** 0.432***
## yoy_revenue_growth 0.076* -0.315*** -0.38***
## google_trends -0.04 -0.054 -0.122***
## number_of_news -0.16*** 0.066 0.069*
## number_of_likes_twitter -0.158*** -0.02 0.003
## number_of_tweets -0.222*** -0.127*** -0.122***
## sales -0.028 0.132*** 0.141***
## ROA 0.685*** 0.25*** 0.382***
## fin_leverage -0.109*** -0.033 0.009
## buyback_yield -0.196*** -0.158*** -0.057
## log_sales -0.015 0.143*** 0.19***
## pol_sub_news polarity_twitter yoy_revenue_growth
## price_to_book 0.361*** 0.193*** 0.076*
## polarity_news 0.728*** 0.353*** -0.315***
## subjectivity_news 0.726*** 0.432*** -0.38***
## pol_sub_news 1 0.351*** -0.368***
## polarity_twitter 0.351*** 1 -0.336***
## yoy_revenue_growth -0.368*** -0.336*** 1
## google_trends -0.107*** -0.19*** 0.129***
## number_of_news 0.074* 0.091** -0.119***
## number_of_likes_twitter 0.01 0.035 -0.037
## number_of_tweets -0.116*** -0.01 0.094**
## sales 0.116*** 0.052 -0.195***
## ROA 0.269*** 0.261*** -0.279***
## fin_leverage 0.03 -0.041 0.045
## buyback_yield -0.06 -0.024 0.006
## log_sales 0.14*** 0.096** -0.23***
## google_trends number_of_news number_of_likes_twitter
## price_to_book -0.04 -0.16*** -0.158***
## polarity_news -0.054 0.066 -0.02
## subjectivity_news -0.122*** 0.069* 0.003
## pol_sub_news -0.107*** 0.074* 0.01
## polarity_twitter -0.19*** 0.091** 0.035
## yoy_revenue_growth 0.129*** -0.119*** -0.037
## google_trends 1 0.258*** 0.085**
## number_of_news 0.258*** 1 0.478***
## number_of_likes_twitter 0.085** 0.478*** 1
## number_of_tweets 0.153*** 0.549*** 0.739***
## sales 0.304*** 0.336*** 0.041
## ROA -0.145*** -0.039 -0.1**
## fin_leverage -0.084** -0.029 0.015
## buyback_yield -0.075* 0.141*** 0.028
## log_sales 0.282*** 0.266*** 0.038
## number_of_tweets sales ROA fin_leverage
## price_to_book -0.222*** -0.028 0.685*** -0.109***
## polarity_news -0.127*** 0.132*** 0.25*** -0.033
## subjectivity_news -0.122*** 0.141*** 0.382*** 0.009
## pol_sub_news -0.116*** 0.116*** 0.269*** 0.03
## polarity_twitter -0.01 0.052 0.261*** -0.041
## yoy_revenue_growth 0.094** -0.195*** -0.279*** 0.045
## google_trends 0.153*** 0.304*** -0.145*** -0.084**
## number_of_news 0.549*** 0.336*** -0.039 -0.029
## number_of_likes_twitter 0.739*** 0.041 -0.1** 0.015
## number_of_tweets 1 0.021 -0.154*** -0.013
## sales 0.021 1 0.052 -0.128***
## ROA -0.154*** 0.052 1 -0.07*
## fin_leverage -0.013 -0.128*** -0.07* 1
## buyback_yield 0.047 0.369*** -0.032 0.388***
## log_sales -0.003 0.923*** 0.087** -0.127***
## buyback_yield log_sales
## price_to_book -0.196*** -0.015
## polarity_news -0.158*** 0.143***
## subjectivity_news -0.057 0.19***
## pol_sub_news -0.06 0.14***
## polarity_twitter -0.024 0.096**
## yoy_revenue_growth 0.006 -0.23***
## google_trends -0.075* 0.282***
## number_of_news 0.141*** 0.266***
## number_of_likes_twitter 0.028 0.038
## number_of_tweets 0.047 -0.003
## sales 0.369*** 0.923***
## ROA -0.032 0.087**
## fin_leverage 0.388*** -0.127***
## buyback_yield 1 0.349***
## log_sales 0.349*** 1
Цикл позволяет вывести корреляционную матрицу с указанием уровней значимости * = 10%, ** = 5%,*** = 1% соответственно. Борьба с мультиколлинеарностью проводится далее. На данном этапе лишь подтверждается ее необходимость.
МОДЕЛИРОВАНИЕ
dataPanel <- pdata.frame(diplom_1802_w_imp, index=c("ticker","year"))# трансформация в панельные данные
filter(dataPanel, year == "2017") %>%
group_by(sector)%>%
count(sector) # выводим представленный в выборке набор секторов## # A tibble: 7 × 2
## # Groups: sector [7]
## sector n
## <chr> <int>
## 1 Communication Services 13
## 2 Consumer Discretionary 15
## 3 Consumer Staples 7
## 4 Health Care 13
## 5 Industrials 8
## 6 Information Technology 42
## 7 Utilities 4
МОДЕЛЬ МНОЖЕСТВЕННОЙ ЛИНЕЙНОЙ РЕГРЕССИИ
dataPanel_std <- pdata.frame(data.frame(diplom_1802_w_imp[,c(1:5)], scale(diplom_1802_w_imp[,-c(1:5)])), index=c("ticker","year")) #дополнительно создаем дф с стандартнизированными переменнымиols1 <-plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield , data = filter(dataPanel_std, year != "2016"), model="within")summary(ols1) # оцениваем первую спецификацию как обычную множественную линейную регрессию со стандартизированными коэффициентами ## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel_std,
## year != "2016"), model = "within")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -1.903461 -0.172692 -0.021048 0.172409 1.617357
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 0.1975949 0.0280822 7.0363 6.535e-12 ***
## polarity_twitter 0.0070015 0.0238181 0.2940 0.768913
## yoy_revenue_growth 0.1687012 0.1014829 1.6624 0.097067 .
## google_trends 0.0666675 0.0237914 2.8022 0.005273 **
## number_of_news 0.0220853 0.0313411 0.7047 0.481339
## number_of_tweets 0.0603640 0.0362501 1.6652 0.096497 .
## sales -0.0543929 0.0243196 -2.2366 0.025754 *
## ROA 0.6609974 0.0258889 25.5321 < 2.2e-16 ***
## fin_leverage -0.0521373 0.0255184 -2.0431 0.041563 *
## buyback_yield -0.0145474 0.0289439 -0.5026 0.615461
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 208.97
## Residual Sum of Squares: 80.697
## R-Squared: 0.61384
## Adj. R-Squared: 0.52811
## F-statistic: 79.4792 on 10 and 500 DF, p-value: < 2.22e-16
ols2 <-plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel_std, year != "2016"), model="within")summary(ols2)# оцениваем вторую спецификацию как обычную множественную линейную регрессию со стандартизированными коэффициентами## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel_std,
## year != "2016"), model = "within")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -1.861411 -0.172475 -0.018288 0.159766 1.607813
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 0.2111822 0.0275253 7.6723 8.910e-14 ***
## polarity_twitter 0.0027737 0.0231268 0.1199 0.9045831
## yoy_revenue_growth 0.1741428 0.0989866 1.7593 0.0791455 .
## google_trends 0.0839150 0.0234737 3.5749 0.0003844 ***
## number_of_news 0.0276843 0.0283873 0.9752 0.3299146
## number_of_likes_twitter 0.1217805 0.0230780 5.2769 1.961e-07 ***
## sales -0.0617157 0.0235184 -2.6241 0.0089523 **
## ROA 0.6754596 0.0253669 26.6276 < 2.2e-16 ***
## fin_leverage -0.0546116 0.0248894 -2.1942 0.0286822 *
## buyback_yield -0.0087167 0.0282713 -0.3083 0.7579650
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 208.97
## Residual Sum of Squares: 76.864
## R-Squared: 0.63218
## Adj. R-Squared: 0.55052
## F-statistic: 85.9362 on 10 and 500 DF, p-value: < 2.22e-16
ols3 <-plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel, year != "2016"), model="within")summary(ols3) # оцениваем первую спецификацию как обычную множественную линейную регрессию## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel,
## year != "2016"), model = "within")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -2.726513 -0.247364 -0.030149 0.246958 2.316698
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 25.11107946 3.56879052 7.0363 6.535e-12 ***
## polarity_twitter 0.48001275 1.63293902 0.2940 0.768913
## yoy_revenue_growth 0.01885132 0.01134010 1.6624 0.097067 .
## google_trends 0.00106223 0.00037908 2.8022 0.005273 **
## number_of_news 0.00965490 0.01370117 0.7047 0.481339
## number_of_tweets 0.00224214 0.00134646 1.6652 0.096497 .
## sales -0.00971785 0.00434496 -2.2366 0.025754 *
## ROA 0.14950853 0.00585571 25.5321 < 2.2e-16 ***
## fin_leverage -0.02371977 0.01160952 -2.0431 0.041563 *
## buyback_yield -0.00594522 0.01182875 -0.5026 0.615461
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 428.76
## Residual Sum of Squares: 165.57
## R-Squared: 0.61384
## Adj. R-Squared: 0.52811
## F-statistic: 79.4792 on 10 and 500 DF, p-value: < 2.22e-16
ols4 <-plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel, year != "2016"), model="within")summary(ols4)# оцениваем вторую спецификацию как обычную множественную линейную регрессию ## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel,
## year != "2016"), model = "within")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -2.666281 -0.247053 -0.026195 0.228849 2.303028
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 26.83779857 3.49801536 7.6723 8.910e-14 ***
## polarity_twitter 0.19016168 1.58554481 0.1199 0.9045831
## yoy_revenue_growth 0.01945939 0.01106114 1.7593 0.0791455 .
## google_trends 0.00133704 0.00037401 3.5749 0.0003844 ***
## number_of_news 0.01210258 0.01240990 0.9752 0.3299146
## number_of_likes_twitter 0.01376069 0.00260772 5.2769 1.961e-07 ***
## sales -0.01102614 0.00420181 -2.6241 0.0089523 **
## ROA 0.15277968 0.00573763 26.6276 < 2.2e-16 ***
## fin_leverage -0.02484544 0.01132336 -2.1942 0.0286822 *
## buyback_yield -0.00356232 0.01155387 -0.3083 0.7579650
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 428.76
## Residual Sum of Squares: 157.71
## R-Squared: 0.63218
## Adj. R-Squared: 0.55052
## F-statistic: 85.9362 on 10 and 500 DF, p-value: < 2.22e-16
МОДЕЛЬ МНОЖЕСТВЕННОЙ ЛИНЕЙНОЙ РЕГРЕССИИ | ВЫБОР СПЕЦИФИКАЦИИ
ols <-lm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel, year != "2016"))
summary(ols)##
## Call:
## lm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel,
## year != "2016"))
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.5872 -0.3956 -0.0876 0.3342 4.3366
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -4.2611883 0.3583478 -11.891 < 2e-16 ***
## polarity_news 47.2520171 3.2026623 14.754 < 2e-16 ***
## polarity_twitter 1.8313043 1.7020330 1.076 0.28238
## yoy_revenue_growth 0.0428334 0.0027884 15.361 < 2e-16 ***
## google_trends 0.0009137 0.0004016 2.275 0.02324 *
## number_of_news -0.0361348 0.0127386 -2.837 0.00471 **
## number_of_tweets -0.0024498 0.0010205 -2.401 0.01667 *
## sales 0.0001444 0.0051885 0.028 0.97780
## ROA 0.1542498 0.0053542 28.809 < 2e-16 ***
## fin_leverage -0.0133935 0.0115023 -1.164 0.24471
## buyback_yield -0.0357095 0.0116653 -3.061 0.00230 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.7735 on 601 degrees of freedom
## Multiple R-squared: 0.7132, Adjusted R-squared: 0.7084
## F-statistic: 149.5 on 10 and 601 DF, p-value: < 2.2e-16
regclass::VIF(ols)## polarity_news polarity_twitter yoy_revenue_growth google_trends
## 1.330886 1.291543 1.304850 1.331091
## number_of_news number_of_tweets sales ROA
## 1.779327 1.581860 1.767337 1.174227
## fin_leverage buyback_yield
## 1.339493 1.707357
Мультиколлинеарности в модели после удаления части переменных нет, так как коэффициенты VIF < 5. Дополнительное подтверждение отсутствию необходимости включения регуляризации. Однако в выборке присутствуют компании и года, значительно отличающиеся друг от друга, проверим дополнительные спецификации.
bptest(ols4, data = dataPanel)##
## studentized Breusch-Pagan test
##
## data: ols4
## BP = 117.69, df = 10, p-value < 2.2e-16
plm::pbgtest(ols4, data = dataPanel)##
## Breusch-Godfrey/Wooldridge test for serial correlation in panel models
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + ...
## chisq = 71.11, df = 6, p-value = 2.42e-13
## alternative hypothesis: serial correlation in idiosyncratic errors
Есть автокорреляция и гетероскедастичность, необходимо использовать скорректированную ковариационную матрицу.
coeftest(ols4, vcovHC(ols4, method = "arellano", type = "HC2"))##
## t test of coefficients:
##
## Estimate Std. Error t value Pr(>|t|)
## polarity_news 26.83779857 7.37249982 3.6403 0.0003007 ***
## polarity_twitter 0.19016168 1.91886821 0.0991 0.9210978
## yoy_revenue_growth 0.01945939 0.01052297 1.8492 0.0650148 .
## google_trends 0.00133704 0.00064068 2.0869 0.0374018 *
## number_of_news 0.01210258 0.02130224 0.5681 0.5701972
## number_of_likes_twitter 0.01376069 0.00479490 2.8699 0.0042804 **
## sales -0.01102614 0.00530562 -2.0782 0.0382006 *
## ROA 0.15277968 0.01074163 14.2231 < 2.2e-16 ***
## fin_leverage -0.02484544 0.00850044 -2.9228 0.0036257 **
## buyback_yield -0.00356232 0.00979264 -0.3638 0.7161795
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
plm::pcdtest(ols4, test = c("cd"))##
## Pesaran CD test for cross-sectional dependence in panels
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield
## z = -0.5697, p-value = 0.5689
## alternative hypothesis: cross-sectional dependence
no cross-sectional dependence
plm::pcdtest(ols4, test = c("lm"))##
## Breusch-Pagan LM test for cross-sectional dependence in panels
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield
## chisq = 6560.3, df = 5151, p-value < 2.2e-16
## alternative hypothesis: cross-sectional dependence
cross-sectional dependence Вывод: результаты тестов разнятся, однако даже в случае наличия кросс-секциональной зависимости это небольшая проблема, так как у нас короткий временной интервал и большая выборка компаний.
dwtest(ols)##
## Durbin-Watson test
##
## data: ols
## DW = 1.0993, p-value < 2.2e-16
## alternative hypothesis: true autocorrelation is greater than 0
fixed <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel, year != "2016"), model="within")
summary(fixed)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel,
## year != "2016"), model = "within")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -2.726513 -0.247364 -0.030149 0.246958 2.316698
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 25.11107946 3.56879052 7.0363 6.535e-12 ***
## polarity_twitter 0.48001275 1.63293902 0.2940 0.768913
## yoy_revenue_growth 0.01885132 0.01134010 1.6624 0.097067 .
## google_trends 0.00106223 0.00037908 2.8022 0.005273 **
## number_of_news 0.00965490 0.01370117 0.7047 0.481339
## number_of_tweets 0.00224214 0.00134646 1.6652 0.096497 .
## sales -0.00971785 0.00434496 -2.2366 0.025754 *
## ROA 0.14950853 0.00585571 25.5321 < 2.2e-16 ***
## fin_leverage -0.02371977 0.01160952 -2.0431 0.041563 *
## buyback_yield -0.00594522 0.01182875 -0.5026 0.615461
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 428.76
## Residual Sum of Squares: 165.57
## R-Squared: 0.61384
## Adj. R-Squared: 0.52811
## F-statistic: 79.4792 on 10 and 500 DF, p-value: < 2.22e-16
Фиксированные эффекты = константы для каждой компании:
fixef(fixed)## AAPL ABNB ADBE ADI ADP ADSK AEP ALGN
## -5.268724 -1.091923 -0.060543 -2.979120 -1.645704 -0.895133 -1.607038 -1.266904
## AMAT AMD AMGN AMZN ANSS ASML ATVI AVGO
## -3.188847 -1.965730 -2.417254 -0.540221 -1.584121 -1.364192 -1.833405 -1.468335
## AZN BIDU BIIB BKNG CDNS CEG CHTR CMCSA
## -1.578474 -1.750082 -2.571405 -2.171583 -1.806305 -1.410195 -1.947064 -2.910469
## COST CPRT CRWD CSCO CSX CTAS CTSH DDOG
## -2.488231 -1.200323 -1.636930 -2.711529 -2.488288 -1.798977 -2.166980 -1.546219
## DLTR DOCU DXCM EA EBAY EXC FAST FISV
## -2.104780 -1.724635 -1.692139 -1.924382 -2.194261 -2.431851 -2.245624 -2.126222
## FTNT GILD GOOG GOOGL HON IDXX ILMN INTC
## -1.744955 -2.105113 -1.974287 -1.895555 -2.403201 -1.704043 -1.415760 -3.245168
## INTU ISRG JD KDP KHC KLAC LCID LRCX
## 0.301209 -1.157126 -3.115394 -2.097613 -2.180186 -2.170309 -1.021640 -2.232596
## LULU MAR MCHP MDLZ MELI META MNST MRNA
## -2.190834 -2.034622 -1.618909 -2.464274 -1.823332 -1.236988 -1.848202 -1.457964
## MRVL MSFT MTCH MU NFLX NTES NVDA NXPI
## -1.637366 -2.262962 -1.970139 -1.610258 -0.130539 -1.848994 0.938387 -2.187915
## ODFL OKTA ORLY PANW PAYX PCAR PDD PEP
## -1.667947 -1.788360 -2.294968 -1.904230 -1.789004 -2.418683 -1.699515 -2.224405
## PYPL QCOM REGN ROST SBUX SGEN SIRI SNPS
## -1.710553 -2.543108 -2.125301 -1.949423 -2.435735 -1.672976 -2.049722 -1.405538
## SPLK SWKS TEAM TMUS TSLA TXN VRSK VRSN
## -2.028079 -2.163760 -1.843044 -2.184615 -0.903707 -2.073659 -2.307856 -2.221563
## VRTX WBA WDAY XEL ZM ZS
## -2.158579 -2.331539 -1.634870 -1.886763 -1.307891 -1.736754
Сравним спецификации с учетом фиксированных эффектов и без:
pFtest(fixed, ols)##
## F test for individual effects
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + ...
## F = 5.7995, df1 = 101, df2 = 500, p-value < 2.2e-16
## alternative hypothesis: significant effects
Нулевая гипотеза: спецификация МНК лучше спецификации с фиксированными эффектами p-value < 2.2e-16 < 0.01, следовательно, на уровне значимости 1% нулевая гипотеза отклоняется, фиксированные эффекты значимы.
Оценим спецификацию со случайными эффектами:
random <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel, year != "2016"), model="random")
summary(random)## Oneway (individual) effect Random Effect Model
## (Swamy-Arora's transformation)
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = filter(dataPanel,
## year != "2016"), model = "random")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Effects:
## var std.dev share
## idiosyncratic 0.3311 0.5755 0.632
## individual 0.1926 0.4388 0.368
## theta: 0.528
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -2.610474 -0.308718 -0.057921 0.260679 3.231782
##
## Coefficients:
## Estimate Std. Error z-value Pr(>|z|)
## (Intercept) -2.82509910 0.37697534 -7.4941 6.674e-14 ***
## polarity_news 34.48548368 3.31286816 10.4096 < 2.2e-16 ***
## polarity_twitter 0.35801533 1.59610592 0.2243 0.822520
## yoy_revenue_growth 0.03597278 0.00405744 8.8659 < 2.2e-16 ***
## google_trends 0.00106144 0.00037318 2.8443 0.004451 **
## number_of_news -0.00967462 0.01294532 -0.7473 0.454855
## number_of_tweets -0.00072379 0.00114784 -0.6306 0.528323
## sales -0.00562326 0.00442311 -1.2713 0.203609
## ROA 0.15024663 0.00546816 27.4766 < 2.2e-16 ***
## fin_leverage -0.02323962 0.01087148 -2.1377 0.032544 *
## buyback_yield -0.02162316 0.01118163 -1.9338 0.053136 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 612.51
## Residual Sum of Squares: 221.69
## R-Squared: 0.63807
## Adj. R-Squared: 0.63205
## Chisq: 1059.54 on 10 DF, p-value: < 2.22e-16
Чтобы выбрать между фиксированными или случайными эффектами, проведем тест Хаусмана, где нулевая гипотеза состоит в том, что предпочтительная модель — это случайные эффекты, а альтернативная — фиксированные эффекты (см. Green, 2008, глава 9). По сути, он проверяет, коррелируют ли уникальные ошибки с регрессорами, но нулевая гипотеза заключается в том, что это не так. Если значение p значимо (например, <0,05), необходимо использовать фиксированные эффекты, если нет, - случайные эффекты.
phtest(fixed, random)##
## Hausman Test
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + ...
## chisq = 120.89, df = 10, p-value < 2.2e-16
## alternative hypothesis: one model is inconsistent
В нашем случае p-value < 2.2e-16, на 1% уровне значимости нулевая гипотеза отвергается, следовательно, релевантнее модель с фиксированными эффектами. Также необходимо проверить наличие значимых временных эффектов.
fixed.time <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield + factor(year), data = filter(dataPanel, year != "2016"), model="within")
summary(fixed.time)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield + factor(year),
## data = filter(dataPanel, year != "2016"), model = "within")
##
## Balanced Panel: n = 102, T = 6, N = 612
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -2.75165 -0.25545 -0.02850 0.24421 2.33601
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 24.92529548 3.67057085 6.7906 3.21e-11 ***
## polarity_twitter 0.42532670 1.64773139 0.2581 0.796415
## yoy_revenue_growth 0.01886452 0.01147667 1.6437 0.100867
## google_trends 0.00101128 0.00038639 2.6173 0.009135 **
## number_of_news 0.00980988 0.01386546 0.7075 0.479586
## number_of_tweets 0.00218942 0.00136210 1.6074 0.108609
## sales -0.01048957 0.00458596 -2.2873 0.022598 *
## ROA 0.14932862 0.00600220 24.8790 < 2.2e-16 ***
## fin_leverage -0.02403491 0.01168224 -2.0574 0.040172 *
## buyback_yield -0.00612786 0.01196459 -0.5122 0.608763
## factor(year)2018 0.04002950 0.08368190 0.4784 0.632610
## factor(year)2019 -0.02957576 0.09070176 -0.3261 0.744504
## factor(year)2020 0.00763063 0.08454971 0.0903 0.928125
## factor(year)2021 0.02453276 0.08370993 0.2931 0.769592
## factor(year)2022 -0.01166094 0.08362002 -0.1395 0.889150
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 428.76
## Residual Sum of Squares: 165.29
## R-Squared: 0.61449
## Adj. R-Squared: 0.52415
## F-statistic: 52.6018 on 15 and 495 DF, p-value: < 2.22e-16
pFtest(fixed.time, fixed)##
## F test for individual effects
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + ...
## F = 0.16857, df1 = 5, df2 = 495, p-value = 0.9741
## alternative hypothesis: significant effects
Нулевая гипотеза: проверяемый эффект не значим или равен 0. p-value = 0.9741, следовательно, на уровне значимости 1% нулевая гипотеза на исследуемой выборке не отвергается, временные эффекты не значимы.
plmtest(fixed, c("time"), type=("bp"))##
## Lagrange Multiplier Test - time effects (Breusch-Pagan)
##
## data: price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + ...
## chisq = 1.076, df = 1, p-value = 0.2996
## alternative hypothesis: significant effects
Результаты теста Бреуша-Пагана аналогичны, наилучшей является спецификация с фиксированными эффектами.
datapanel_train = filter(dataPanel[,-c(8, 9, 20)], year != "2016" & year !="2017")МОДЕЛЬ БУСТИНГ СЛУЧАЙНЫХ ЛЕСОВ
#install.packages('xgboost')
library(xgboost)##
## Attaching package: 'xgboost'
## The following object is masked from 'package:dplyr':
##
## slice
# Create the specification with placeholders
boost_spec <- boost_tree(
trees = 1000,
learn_rate = tune(),
tree_depth = tune()) %>%
set_mode("regression") %>%
set_engine("xgboost")
# Create the tuning grid
tunegrid_boost <- grid_regular(parameters(boost_spec),
levels = 5)## Warning: `parameters.model_spec()` was deprecated in tune 0.1.6.9003.
## ℹ Please use `hardhat::extract_parameter_set_dials()` instead.
tunegrid_boost## # A tibble: 25 × 2
## tree_depth learn_rate
## <int> <dbl>
## 1 1 0.001
## 2 4 0.001
## 3 8 0.001
## 4 11 0.001
## 5 15 0.001
## 6 1 0.00422
## 7 4 0.00422
## 8 8 0.00422
## 9 11 0.00422
## 10 15 0.00422
## # … with 15 more rows
# Create CV folds of training data
folds <- vfold_cv(datapanel_train[,-c(1:5)], v = 6)
# Tune along the grid
tune_results <- tune_grid(boost_spec,
price_to_book ~ .,
resamples = folds,
grid = tunegrid_boost,
metrics = metric_set(rmse, mae, rsq))
# Plot the results
autoplot(tune_results)
# Select the final hyperparameters
best_params <- select_best(tune_results)## Warning: No value of `metric` was given; metric 'rmse' will be used.
# Finalize the specification
final_spec <- finalize_model(boost_spec, best_params)
# Train the final model on the full training data
final_model <- final_spec %>% fit(formula = price_to_book ~ .,
data = datapanel_train[,-c(1:5)])
final_model## parsnip model object
##
## ##### xgb.Booster
## raw: 1.5 Mb
## call:
## xgboost::xgb.train(params = list(eta = 0.0177827941003892, max_depth = 4L,
## gamma = 0, colsample_bytree = 1, colsample_bynode = 1, min_child_weight = 1,
## subsample = 1), data = x$data, nrounds = 1000, watchlist = x$watchlist,
## verbose = 0, nthread = 1, objective = "reg:squarederror")
## params (as set within xgb.train):
## eta = "0.0177827941003892", max_depth = "4", gamma = "0", colsample_bytree = "1", colsample_bynode = "1", min_child_weight = "1", subsample = "1", nthread = "1", objective = "reg:squarederror", validate_parameters = "TRUE"
## xgb.attributes:
## niter
## callbacks:
## cb.evaluation.log()
## # of features: 11
## niter: 1000
## nfeatures : 11
## evaluation_log:
## iter training_rmse
## 1 2.90173070
## 2 2.85470977
## ---
## 999 0.09365539
## 1000 0.09359102
vip::vip(final_model)
best_params## # A tibble: 1 × 3
## tree_depth learn_rate .config
## <int> <dbl> <chr>
## 1 4 0.0178 Preprocessor1_Model12
КАЧЕСТВО МОДЕЛИ МНОЖЕСТВЕННОЙ ЛИНЕЙНОЙ РЕГРЕССИИ С ВКЛЮЧЕНИЕМ ФИКСИРОВАННЫХ ЭФФЕКТОВ
datapanel = filter(dataPanel[,-c(8, 9, 20)], year != "2016")
# Predict new data
predictions_fixed_1 <- predict(ols3,
new_data = datapanel[,-c(1:5)])# Compute the mean absolute error using one single function
mae_fixed_1 = mae(cbind(datapanel[,-c(1:5)], predictions_fixed_1),
truth = price_to_book,
estimate = predictions_fixed_1)
# Compute the RMSE using a function
rmse_fixed_1 = rmse(cbind(datapanel[,-c(1:5)], predictions_fixed_1),
truth = price_to_book,
estimate = predictions_fixed_1)
rsq_fixed_1 = rsq_trad(cbind(datapanel[,-c(1:5)], predictions_fixed_1),
truth = price_to_book,
estimate = predictions_fixed_1)
# Print errors
mae_fixed_1["model"] = "fixed_1_1"
#mae_fixedtime
rmse_fixed_1["model"] = "fixed_1_1"
#rmse_fixedtime
rsq_fixed_1["model"] = "fixed_1_1"
#rsq_fixedtime
fixed_evaluation_1 = rbind(mae_fixed_1, rmse_fixed_1, rsq_fixed_1)
fixed_evaluation_1## # A tibble: 3 × 4
## .metric .estimator .estimate model
## <chr> <chr> <dbl> <chr>
## 1 mae standard 0.359 fixed_1_1
## 2 rmse standard 0.520 fixed_1_1
## 3 rsq_trad standard 0.868 fixed_1_1
# Predict new data
predictions_fixed_2 <- predict(ols4,
new_data = datapanel[,-c(1:5)])# Compute the mean absolute error using one single function
mae_fixed_2 = mae(cbind(datapanel[,-c(1:5)], predictions_fixed_2),
truth = price_to_book,
estimate = predictions_fixed_2)
# Compute the RMSE using a function
rmse_fixed_2 = rmse(cbind(datapanel[,-c(1:5)], predictions_fixed_2),
truth = price_to_book,
estimate = predictions_fixed_2)
rsq_fixed_2 = rsq_trad(cbind(datapanel[,-c(1:5)], predictions_fixed_2),
truth = price_to_book,
estimate = predictions_fixed_2)
# Print errors
mae_fixed_2["model"] = "fixed_1_2"
#mae_fixedtime
rmse_fixed_2["model"] = "fixed_1_2"
#rmse_fixedtime
rsq_fixed_2["model"] = "fixed_1_2"
#rsq_fixedtime
fixed_evaluation_2 = rbind(mae_fixed_2, rmse_fixed_2, rsq_fixed_2)
fixed_evaluation_2## # A tibble: 3 × 4
## .metric .estimator .estimate model
## <chr> <chr> <dbl> <chr>
## 1 mae standard 0.352 fixed_1_2
## 2 rmse standard 0.508 fixed_1_2
## 3 rsq_trad standard 0.874 fixed_1_2
КАЧЕСТВО МОДЕЛИ БУСТИНГА СЛУЧАЙНОГО ЛЕСА
diplom_1802_w_imp = read.csv('/Users/olesyamba/Downloads/data/diplom_1802_w_imp.csv')
dataPanel <- pdata.frame(diplom_1802_w_imp, index=c("ticker","year"))
datapanel_test = filter(dataPanel[,-c(8, 9, 20)], year == "2016" | year =="2017")# Predict new data
predictions_boostforest <- predict(final_model,
new_data = datapanel_test[,-c(1:5)])
# Compute the mean absolute error using one single function
mae_boostforest = mae(predictions_boostforest,
truth = datapanel_test[,-c(1:5)]$price_to_book,
estimate = .pred)
# Compute the RMSE using a function
rmse_boostforest = rmse(predictions_boostforest,
truth = datapanel_test[,-c(1:5)]$price_to_book,
estimate = .pred)
rsq_boostforest = rsq_trad(predictions_boostforest,
truth = datapanel_test[,-c(1:5)]$price_to_book,
estimate = .pred)
# Print errors
mae_boostforest["model"] = "boosted_random_forest"
#mae_fixedtime
rmse_boostforest["model"] = "boosted_random_forest"
#rmse_fixedtime
rsq_boostforest["model"] = "boosted_random_forest"
#rsq_fixedtime
boostforest_evaluation = rbind(mae_boostforest, rmse_boostforest, rsq_boostforest)
boostforest_evaluation## # A tibble: 3 × 4
## .metric .estimator .estimate model
## <chr> <chr> <dbl> <chr>
## 1 mae standard 0.360 boosted_random_forest
## 2 rmse standard 0.596 boosted_random_forest
## 3 rsq_trad standard 0.692 boosted_random_forest
evaluation_metrics = rbind(fixed_evaluation_1, fixed_evaluation_2, boostforest_evaluation)
colnames(evaluation_metrics) = c("Metric", "Estimator", "Estimation", "Model")
evaluation_metrics## # A tibble: 9 × 4
## Metric Estimator Estimation Model
## <chr> <chr> <dbl> <chr>
## 1 mae standard 0.359 fixed_1_1
## 2 rmse standard 0.520 fixed_1_1
## 3 rsq_trad standard 0.868 fixed_1_1
## 4 mae standard 0.352 fixed_1_2
## 5 rmse standard 0.508 fixed_1_2
## 6 rsq_trad standard 0.874 fixed_1_2
## 7 mae standard 0.360 boosted_random_forest
## 8 rmse standard 0.596 boosted_random_forest
## 9 rsq_trad standard 0.692 boosted_random_forest
АНАЛИЗ УСТОЙЧИВОСТИ
diplom_1802_w_imp = read.csv('/Users/olesyamba/Downloads/data/diplom_1802_w_imp.csv')
diplom_1802_w_imp1 = filter(diplom_1802_w_imp, year == "2017"|year == "2018")
diplom_1802_w_imp2 = filter(diplom_1802_w_imp, year == "2019"|year == "2020")
diplom_1802_w_imp3 = filter(diplom_1802_w_imp, year == "2021"|year == "2022")dataPanel1 <- pdata.frame(diplom_1802_w_imp1, index=c("ticker","year"))
dataPanel2 <- pdata.frame(diplom_1802_w_imp2, index=c("ticker","year"))
dataPanel3 <- pdata.frame(diplom_1802_w_imp3, index=c("ticker","year"))МОДЕЛЬ МНОЖЕСТВЕННОЙ ЛИНЕЙНОЙ РЕГРЕССИИ
fixed1_1 <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = dataPanel1, model="within")
summary(fixed1_1)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = dataPanel1,
## model = "within")
##
## Balanced Panel: n = 102, T = 2, N = 204
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -1.0268e+00 -1.6253e-01 4.5146e-16 1.6253e-01 1.0268e+00
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 14.37122909 6.14715589 2.3379 0.021562 *
## polarity_twitter -0.36059325 3.01081848 -0.1198 0.904930
## yoy_revenue_growth 0.07705016 0.11240291 0.6855 0.494764
## google_trends -0.00027423 0.00140911 -0.1946 0.846127
## number_of_news 0.18301852 0.05625504 3.2534 0.001596 **
## number_of_tweets -0.01387097 0.00632290 -2.1938 0.030772 *
## sales -0.01402590 0.01022547 -1.3717 0.173504
## ROA 0.13511819 0.01052098 12.8427 < 2.2e-16 ***
## fin_leverage -0.01282421 0.02920037 -0.4392 0.661561
## buyback_yield -0.00057185 0.02119325 -0.0270 0.978532
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 72.093
## Residual Sum of Squares: 19.526
## R-Squared: 0.72916
## Adj. R-Squared: 0.40238
## F-statistic: 24.7681 on 10 and 92 DF, p-value: < 2.22e-16
fixed1_2 <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield, data = dataPanel1, model="within")
summary(fixed1_2)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter +
## sales + ROA + fin_leverage + buyback_yield, data = dataPanel1,
## model = "within")
##
## Balanced Panel: n = 102, T = 2, N = 204
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -9.6340e-01 -1.6344e-01 -1.6523e-16 1.6344e-01 9.6340e-01
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 14.8841081 6.3465892 2.3452 0.02117 *
## polarity_twitter 0.8663237 3.0445566 0.2845 0.77663
## yoy_revenue_growth 0.0012769 0.1144324 0.0112 0.99112
## google_trends -0.0010243 0.0014198 -0.7214 0.47247
## number_of_news 0.0915475 0.0514224 1.7803 0.07833 .
## number_of_likes_twitter 0.0247709 0.0248682 0.9961 0.32182
## sales 0.0023871 0.0097863 0.2439 0.80783
## ROA 0.1420189 0.0115497 12.2963 < 2e-16 ***
## fin_leverage -0.0062061 0.0298837 -0.2077 0.83594
## buyback_yield 0.0060592 0.0215155 0.2816 0.77887
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 72.093
## Residual Sum of Squares: 20.328
## R-Squared: 0.71803
## Adj. R-Squared: 0.37783
## F-statistic: 23.4277 on 10 and 92 DF, p-value: < 2.22e-16
fixed2_1 <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = dataPanel2, model="within")
summary(fixed2_1)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = dataPanel2,
## model = "within")
##
## Balanced Panel: n = 102, T = 2, N = 204
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -1.3231e+00 -1.3635e-01 -2.8428e-16 1.3635e-01 1.3231e+00
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 1.6854e+01 7.0282e+00 2.3980 0.01850 *
## polarity_twitter -5.2471e+00 3.9080e+00 -1.3427 0.18268
## yoy_revenue_growth -8.5357e-03 2.3717e-02 -0.3599 0.71975
## google_trends -6.4638e-05 7.4014e-04 -0.0873 0.93060
## number_of_news 5.1190e-02 4.5631e-02 1.1218 0.26485
## number_of_tweets -3.5223e-03 3.3878e-03 -1.0397 0.30119
## sales -1.9026e-02 8.5391e-03 -2.2282 0.02831 *
## ROA 1.6573e-01 1.6697e-02 9.9260 3.267e-16 ***
## fin_leverage -7.3582e-02 2.3992e-02 -3.0669 0.00284 **
## buyback_yield 3.0785e-02 2.3702e-02 1.2988 0.19725
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 38.473
## Residual Sum of Squares: 16.686
## R-Squared: 0.5663
## Adj. R-Squared: 0.043031
## F-statistic: 12.0128 on 10 and 92 DF, p-value: 5.1753e-13
fixed2_2 <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield, data = dataPanel2, model="within")
summary(fixed2_2)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter +
## sales + ROA + fin_leverage + buyback_yield, data = dataPanel2,
## model = "within")
##
## Balanced Panel: n = 102, T = 2, N = 204
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -1.3208e+00 -1.2727e-01 -5.6606e-16 1.2727e-01 1.3208e+00
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 17.26980244 7.06806630 2.4434 0.016460 *
## polarity_twitter -5.07417818 4.00467520 -1.2671 0.208331
## yoy_revenue_growth -0.01266937 0.02361684 -0.5365 0.592939
## google_trends -0.00021547 0.00073264 -0.2941 0.769343
## number_of_news 0.03085816 0.04530642 0.6811 0.497520
## number_of_likes_twitter 0.00138078 0.00688275 0.2006 0.841442
## sales -0.01672805 0.00835767 -2.0015 0.048282 *
## ROA 0.16997193 0.01676548 10.1382 < 2.2e-16 ***
## fin_leverage -0.07194695 0.02411252 -2.9838 0.003645 **
## buyback_yield 0.02792485 0.02372115 1.1772 0.242147
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 38.473
## Residual Sum of Squares: 16.874
## R-Squared: 0.5614
## Adj. R-Squared: 0.03221
## F-statistic: 11.7756 on 10 and 92 DF, p-value: 8.3955e-13
fixed3_1 <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_tweets + sales + ROA + fin_leverage + buyback_yield, data = dataPanel3, model="within")
summary(fixed3_1)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_tweets +
## sales + ROA + fin_leverage + buyback_yield, data = dataPanel3,
## model = "within")
##
## Balanced Panel: n = 102, T = 2, N = 204
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -6.3761e-01 -1.3421e-01 1.6827e-16 1.3421e-01 6.3761e-01
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 4.9942e+01 1.1329e+01 4.4081 2.82e-05 ***
## polarity_twitter -2.8625e+00 3.1879e+00 -0.8979 0.3715737
## yoy_revenue_growth -6.1430e-02 1.5659e-02 -3.9230 0.0001684 ***
## google_trends -2.8239e-05 5.3186e-03 -0.0053 0.9957752
## number_of_news 8.9655e-03 3.0480e-02 0.2941 0.7693114
## number_of_tweets -7.5085e-06 3.3123e-03 -0.0023 0.9981962
## sales -2.9688e-02 1.2494e-02 -2.3762 0.0195655 *
## ROA 1.7542e-01 1.0245e-02 17.1230 < 2.2e-16 ***
## fin_leverage -9.7198e-03 1.7936e-02 -0.5419 0.5891945
## buyback_yield 5.5794e-02 2.4152e-02 2.3101 0.0231191 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 48.502
## Residual Sum of Squares: 9.0925
## R-Squared: 0.81253
## Adj. R-Squared: 0.58635
## F-statistic: 39.8753 on 10 and 92 DF, p-value: < 2.22e-16
fixed3_2 <- plm(price_to_book ~ polarity_news + polarity_twitter + yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter + sales + ROA + fin_leverage + buyback_yield, data = dataPanel3, model="within")
summary(fixed3_2)## Oneway (individual) effect Within Model
##
## Call:
## plm(formula = price_to_book ~ polarity_news + polarity_twitter +
## yoy_revenue_growth + google_trends + number_of_news + number_of_likes_twitter +
## sales + ROA + fin_leverage + buyback_yield, data = dataPanel3,
## model = "within")
##
## Balanced Panel: n = 102, T = 2, N = 204
##
## Residuals:
## Min. 1st Qu. Median 3rd Qu. Max.
## -6.3721e-01 -1.3447e-01 1.5873e-16 1.3447e-01 6.3721e-01
##
## Coefficients:
## Estimate Std. Error t-value Pr(>|t|)
## polarity_news 4.9996e+01 1.1189e+01 4.4683 2.241e-05 ***
## polarity_twitter -3.0370e+00 3.3248e+00 -0.9134 0.3634047
## yoy_revenue_growth -6.0963e-02 1.5708e-02 -3.8809 0.0001955 ***
## google_trends 2.2066e-05 5.1682e-03 0.0043 0.9966026
## number_of_news 1.0572e-02 2.3456e-02 0.4507 0.6532559
## number_of_likes_twitter -9.4426e-04 6.3769e-03 -0.1481 0.8826073
## sales -3.0066e-02 1.2236e-02 -2.4572 0.0158772 *
## ROA 1.7514e-01 1.0238e-02 17.1072 < 2.2e-16 ***
## fin_leverage -9.6844e-03 1.7763e-02 -0.5452 0.5869370
## buyback_yield 5.5892e-02 2.4141e-02 2.3152 0.0228244 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Total Sum of Squares: 48.502
## Residual Sum of Squares: 9.0904
## R-Squared: 0.81258
## Adj. R-Squared: 0.58645
## F-statistic: 39.887 on 10 and 92 DF, p-value: < 2.22e-16
datapanel1 = dataPanel1[,-c(8, 9, 16:20)]
datapanel2 = dataPanel2[,-c(8, 9, 16:20)]
datapanel3 = dataPanel3[,-c(8, 9, 16:20)]
# Train the final model on the full training data
final_model1 <- final_spec %>% fit(formula = price_to_book ~ .,
data = datapanel1[,-c(1:5)])
final_model2 <- final_spec %>% fit(formula = price_to_book ~ .,
data = datapanel2[,-c(1:5)])
final_model3 <- final_spec %>% fit(formula = price_to_book ~ .,
data = datapanel3[,-c(1:5)])
vip::vip(final_model1)
vip::vip(final_model2)
vip::vip(final_model3)
h1 = vip::vi_model(final_model1)
h1 = h1 %>%
mutate(model = c(1,1,1,1,1,1,1))
h2 = vip::vi_model(final_model2)
h2 = h2 %>%
mutate(model = c(2,2,2,2,2,2,2))
h3 = vip::vi_model(final_model3)
h3 = h3 %>%
mutate(model = c(3,3,3,3,3,3,3))
h = rbind(h1,h2,h3)
colnames(h) = c("Variable", "Importance", "Model")
h## # A tibble: 21 × 3
## Variable Importance Model
## <chr> <dbl> <dbl>
## 1 number_of_tweets 0.457 1
## 2 polarity_news 0.231 1
## 3 google_trends 0.100 1
## 4 yoy_revenue_growth 0.0653 1
## 5 number_of_likes_twitter 0.0588 1
## 6 polarity_twitter 0.0483 1
## 7 number_of_news 0.0398 1
## 8 number_of_tweets 0.467 2
## 9 polarity_news 0.320 2
## 10 number_of_likes_twitter 0.0611 2
## # … with 11 more rows
p = ggplot(h, aes(Variable, Importance, group = Model))
p + geom_line(aes(colour = Model)) + geom_point(aes(colour = Model))