Package 'PRISM.forecast'

Out-of-sample prediction for whole period

Description

Out-of-sample prediction for whole period

Usage

back_test(n.lag = 1:52, s.window = 52, n.history = 700, stl = TRUE,
  n.training = 156, UseGoogle = T, alpha = 1, nPred = 0,
  discount = 0.01, sepL1 = F)

Arguments

n.lag	the number of lags to be used as regressor in Stage 2 of PRISM (by default = 1:52 for weekly data)
s.window	seasonality span in seasonal decomposition (by default = 52 for weekly data)
n.history	length of training period (e.g. in weeks) for seasonal decomposition.
stl	if TRUE, use STL seasonal decomposition; if FALSE, use classic additive seasonal decomposition.
n.training	length of training period in Stage 2, penalized linear regression (by default = 156)
UseGoogle	boolean variable indicating whether to use Google Trend data.
alpha	penalty between lasso and ridge. alpha=1 represents lasso, alpha=0 represents ridge, alpha=NA represents no penalty (by default alpha = 1).
nPred	the number of periods ahead for forecast. nPred = 0,1,2,3.
discount	exponential weighting: (1-discount)^lag.
sepL1	if TRUE, use separate L1 regularization parameters for time series components and exogenous variables (Goolgle Trend data)

Value

prediction nPred week ahead prediction of the whole periods (07 - 20).

Examples

claim_data = load_claim_data()

# It may take a few minutes.
prism_prediction = back_test()
# evaluate the out-of-sample prediction error as a ratio to naive method
evaluation_table(claim_data, prism_prediction)

---

Out-of-sample prediction evaluation

Description

Out-of-sample prediction evaluation

Usage

evaluation_table(claim_data, prism_prediction)

Arguments

claim_data	the output of load_claim_data().
prism_prediction	the output of back_test().

---

Load Goolge Trends data and initial claims data

Description

Load weekly unemployment initial claim data and related Google Trend data over 5-year span (each week ends on the Saturday). The list of Google search terms is the same as in paper.

Usage

load_5y_search_data(folder = "0408")

Arguments

folder

foldernames for a certain periods of Google Trends data. The foldernames are "0408", "0610", "0812", "1014", "1216", "1418", "1620". For example, the folder "0408" is for 2004-2008.

Value

A list of following named xts objects

• claim.data unemployment initial claim data of the same span as Google Trend data.

• claim.all load all unemployment initial claim data since 1967

• claim.early unemployment initial claim data from 1980-01-06 to the start of claim.data.

• allSearch Google Trends data of a span over five years. It is in the scale of 0 – 100.

---

Load unemployment initial claims data

Description

Load weekly unemployment initial claim data (each week ends on the Saturday).

Usage

load_claim_data(GT.startDate = "2004-01-03", GT.endDate = "2016-12-31")

Arguments

GT.startDate	start date of claim data
GT.endDate	end date of claim data

Value

A list of following named xts objects

• claim.data unemployment initial claim data from GT.startDate to GT.endDate.

• claim.all load all unemployment initial claim data since 1967

• claim.early unemployment initial claim data prior to GT.startDate

---

PRISM function

Description

A function for nowcasting and forecasting time series.

Usage

prism(data, data.early, GTdata, stl = TRUE, n.history = 700,
  n.training = 156, alpha = 1, UseGoogle = T, nPred.vec = 0:3,
  discount = 0.01, sepL1 = F)

Arguments

data	time series of interest as xts, last element can be NA. (e.g., unemployment initial claim data in the same period as GTdata).
data.early	historical time series of response variable before contemporaneous exogenous data, GTdata is available. (e.g., unemployment initial claim prior to 2004)
GTdata	contemporaneous exogenous data as xts. (e.g., Google Trend data)
stl	if TRUE, use STL seasonal decomposition; if FALSE, use classic additive seasonal decomposition.
n.history	training period for seasonal decomposition. (by default = 700 wks)
n.training	length of regression training period (by default = 156)
alpha	penalty between lasso and ridge. alpha=1 represents lasso, alpha=0 represents ridge, alpha=NA represents no penalty.
UseGoogle	boolean variable indicating whether to use Google Trend data.
nPred.vec	the number of periods ahead for forecast. nPred.vec could be a vector of intergers. e.g. nPred.vec=0:3 gives results from nowcast to 3-week ahead forecast.
discount	exponential weighting: (1-discount)^lag (by default = 0.01).
sepL1	if TRUE, use separate L1 regularization parameters for time series components and exogenous variables (Goolgle Trend data)

Value

A list of following named objects

• coef coefficients for Intercept, z.lags, seasonal.lags and exogenous variables.

• pred a vector of prediction with nPred.vec weeks forward.

Examples

prism_data = load_5y_search_data('0610')
data = prism_data$claim.data[1:200] # load claim data from 2006-01-07 to 2009-10-31
data[200] = NA # delete the data for the latest date and try to nowcast it.

data.early = prism_data$claim.earlyData # load claim prior to 2006
GTdata = prism_data$allSearch[1:200] # load Google trend data from 2006-01-07 to 2009-10-31

result = prism(data, data.early, GTdata) # call prism method
result$pred # output 0-3wk forward prediction

---

PRISM stage 2 by batch

Description

PRISM penalized linear regression function for a range of time (only used internally for back testing)

Usage

prism_batch(data, GTdata, var, n.training = 156, UseGoogle = T, alpha = 1,
  nPred.vec = 0:3, start.date = NULL, n.weeks = NULL, discount = 0.01,
  sepL1 = F)

Arguments

data	time series of interest as xts, last element can be NA. (e.g., unemployment initial claim data in the same period as GTdata).
GTdata	contemporaneous exogenous data as xts. (e.g., Google Trend data)
var	generated regressors from stage 1.
n.training	length of regression training period (by default = 156)
UseGoogle	boolean variable indicating whether to use Google Trend data.
alpha	penalty between lasso and ridge. alpha=1 represents lasso, alpha=0 represents ridge, alpha=NA represents no penalty.
nPred.vec	the number of periods ahead for forecast. nPred.vec could be a vector of intergers. e.g. nPred.vec=0:3 gives results from nowcast to 3-week ahead forecast.
start.date	the starting date for forecast. If NULL, the forecast start at the earliest possible date.
n.weeks	the number of weeks in the batch. If NULL, the forecast end at the latest possible date.
discount	exponential weighting: (1-discount)^lag (by default = 0.01)
sepL1	if TRUE, use separate L1 regularization parameters for time series components and exogenous variables (Goolgle Trend data)

Value

A list of following named objects

• coef coefficients for Intercept, z.lags, seasonal.lags and exogenous variables.

• pred prediction results for n.weeks from start.date.

---

PRISM regressors generator

Description

Stage 1 of PRISM. The function generates prism seasonal components and seasonally adjusted lag components.

Usage

var_generator(data, data.early, stl = TRUE, n.lag = 1:52, s.window = 52,
  n.history = 700)

Arguments

data	time series of interest as xts, last element can be NA.
data.early	historical time series of response variable before Google Trend data is available. (e.g., unemployment initial claim prior to 2004)
stl	if TRUE, use STL seasonal decomposition; if FALSE, use classic additive seasonal decomposition.
n.lag	the number of lags to be used as regressor in Stage 2 of PRISM (by default = 1:52 for weekly data)
s.window	seasonality span (by default = 52 for weekly data)
n.history	training period for seasonal decomposition. (by default = 700 wks)

Value

A list of following named objects

• y.lags seasonally adjusted components, z_lag, and seasonal components, s_lag.

---