sentometrics (version 0.2)

sento_model: Optimized and automated sparse regression

Description

Linear or nonlinear penalized regression of any dependent variable on the wide number of sentiment measures and potentially other explanatory variables. Either performs a regression given the provided variables at once, or computes regressions sequentially for a given sample size over a longer time horizon, with associated one-step ahead prediction performance metrics.

Usage

sento_model(sentomeasures, y, x = NULL, ctr)

Arguments

sentomeasures

a sentomeasures object created using sento_measures. There should be at least two explanatory variables including the ones provided through the x argument.

y

a one-column data.frame or a numeric vector capturing the dependent (response) variable. In case of a logistic regression, the response variable is either a factor or a matrix with the factors represented by the columns as binary indicators, with the second factor level or column as the reference class in case of a binomial regression. No NA values are allowed.

x

a named data.frame with other explanatory variables as numeric, by default set to NULL.

ctr

output from a ctr_model call.

Value

If ctr$do.iter = FALSE, a sentomodel object which is a list containing:

reg

optimized regression, i.e. a model-specific glmnet object.

model

the input argument ctr$model, to indicate the type of model estimated.

x

a matrix of the values used in the regression for all explanatory variables.

alpha

calibrated alpha.

lambda

calibrated lambda.

trained

output from train call (if ctr$type = "cv").

ic

a list composed of two elements: the information criterion used in the calibration under "criterion", and a vector of all minimum information criterion values for each value in alphas under "opts" (if ctr$type != "cv").

dates

sample reference dates as a two-element character vector, being the earliest and most recent date from the sentomeasures object accounted for in the estimation window.

nVar

the sum of the number of sentiment measures and other explanatory variables inputted.

discarded

a named logical vector of length equal to the number of sentiment measures, in which TRUE indicates that the particular sentiment measure has not been considered in the regression process. A sentiment measure is not considered when it is a duplicate of another, or when at least 25% of the observations are equal to zero.

If ctr$do.iter = TRUE, a sentomodeliter object which is a list containing:

models

all sparse regressions, i.e. separate sentomodel objects as above, as a list with as names the dates from the perspective of the sentiment measures at which predictions for performance measurement are carried out (i.e. one date step beyond the date sentomodel$dates[2]).

alphas

calibrated alphas.

lambdas

calibrated lambdas.

performance

a data.frame with performance-related measures, being "RMSFE" (root mean squared forecasting error), "MAD" (mean absolute deviation), "MDA" (mean directional accuracy, in which's calculation zero is considered as a positive; in percentage points), "accuracy" (proportion of correctly predicted classes in case of a logistic regression; in percentage points), and each's respective individual values in the sample. Directional accuracy is measured by comparing the change in the realized response with the change in the prediction between two consecutive time points (omitting the very first prediction, resulting in NA). Only the relevant performance statistics are given depending on the type of regression. Dates are as in the "models" output element, i.e. from the perspective of the sentiment measures.

Details

Models are computed using the elastic net regularization as implemented in the glmnet package, to account for the multidimensionality of the sentiment measures. Additional explanatory variables are not subject to shrinkage. Independent variables are normalized in the regression process, but coefficients are returned in their original space. For a helpful introduction to glmnet, we refer to the vignette. The optimal elastic net parameters lambda and alpha are calibrated either through a to specify information criterion or through cross-validation (based on the "rolling forecasting origin" principle, using the train function). In the latter case, the training metric is automatically set to "RMSE" for a linear model and to "Accuracy" for a logistic model. We suppress many of the details that can be supplied to the glmnet and train functions we rely on, for the sake of user-friendliness.

See Also

ctr_model, glmnet, train, retrieve_attributions

Examples

# NOT RUN {
data("usnews")
data("lexicons")
data("valence")
data("epu")

# construct a sentomeasures object to start with
corpusAll <- sento_corpus(corpusdf = usnews)
corpus <- quanteda::corpus_subset(corpusAll, date >= "2004-01-01" & date < "2014-10-01")
l <- setup_lexicons(lexicons[c("LM_eng", "HENRY_eng")], valence[["valence_eng"]])
ctr <- ctr_agg(howWithin = "tf-idf", howDocs = "proportional",
               howTime = c("equal_weight", "almon"),
               by = "month", lag = 3, ordersAlm = 1:2,
               do.inverseAlm = TRUE, do.normalizeAlm = TRUE)
sentomeasures <- sento_measures(corpus, l, ctr)

# prepare y and other x variables
y <- epu[epu$date >= sentomeasures$measures$date[1], ]$index
length(y) == nrow(sentomeasures$measures) # TRUE
x <- data.frame(runif(length(y)), rnorm(length(y))) # two other (random) x variables
colnames(x) <- c("x1", "x2")

# a linear model based on the Akaike information criterion
ctrIC <- ctr_model(model = "gaussian", type = "AIC", do.iter = FALSE, h = 0)
out1 <- sento_model(sentomeasures, y, x = x, ctr = ctrIC)

# some post-analysis (attribution)
attributions1 <- retrieve_attributions(out1, sentomeasures,
                                       refDates = sentomeasures$measures$date[20:40])

# }
# NOT RUN {
# a cross-validation based model
cl <- makeCluster(detectCores() - 2)
registerDoParallel(cl)
ctrCV <- ctr_model(model = "gaussian", type = "cv", do.iter = FALSE,
                   h = 0, alphas = c(0.10, 0.50, 0.90), trainWindow = 70,
                   testWindow = 10, oos = 0, do.progress = TRUE)
out2 <- sento_model(sentomeasures, y, x = x, ctr = ctrCV)
stopCluster(cl)
summary(out2)
# }
# NOT RUN {
# }
# NOT RUN {
# a cross-validation based model but for a binomial target
yb <- epu[epu$date >= sentomeasures$measures$date[1], ]$above
ctrCVb <- ctr_model(model = "binomial", type = "cv", do.iter = FALSE,
                    h = 0, alphas = c(0.10, 0.50, 0.90), trainWindow = 70,
                    testWindow = 10, oos = 0, do.progress = TRUE)
out3 <- sento_model(sentomeasures, yb, x = x, ctr = ctrCVb)
summary(out3)
# }
# NOT RUN {
# an example of an iterative analysis
ctrIter <- ctr_model(model = "gaussian", type = "BIC", do.iter = TRUE,
                     alphas = c(0.25, 0.75), h = 0, nSample = 100, start = 21)
out4 <- sento_model(sentomeasures, y, x = x, ctr = ctrIter)
summary(out4)

# }
# NOT RUN {
# a similar iterative analysis, parallelized
cl <- makeCluster(detectCores() - 2)
registerDoParallel(cl)
ctrIter <- ctr_model(model = "gaussian", type = "Cp", do.iter = TRUE,
                     h = 0, nSample = 100, do.parallel = TRUE)
out5 <- sento_model(sentomeasures, y, x = x, ctr = ctrIter)
stopCluster(cl)
summary(out5)
# }
# NOT RUN {
# some more post-analysis (attribution and prediction)
attributions2 <- retrieve_attributions(out4, sentomeasures)
plot_attributions(attributions2, "features")

nx <- ncol(sentomeasures$measures) - 1 + ncol(x) # don't count date column
newx <- runif(nx) * cbind(sentomeasures$measures[, -1], x)[30:50, ]
preds <- predict(out1, newx = as.matrix(newx), type = "link")

# }