A Decision Support System for Municipal Budget Plan Decisions - New Contributions in Information Systems and Technologies - page 125

3.1 R Functions to Forecast Financial Indicators

One of the main applications of the DSS is a group of predictive models, which, in
addition to data queries, using the new database, now a centralized system, are the
system's core. To construct these models we use R software, due to the possibility of
using functions that allows to get models with minimal input from the user and also
possibilities to construct data views and easy to understand predictions on an
appealing way.
Taking into account the interests of the decision(s) maker(s), annually aggregated
data are used as corresponding to short series of 12 values, at most, showing no
pronounced seasonality or global trends, although level variations are frequent.
The R application has many features to model this type of series. In this study we
chose to use the functions described in Table 1, all from the forecast package [8].
Selection criteria were based on the ease of use or degree of automation, opting for
functions that optimize not only the model parameters but also the structural
components to include (level, trend, seasonal factors), and viewing capabilities
offered.
All these functions have as an argument a time series created using the function
ts(). This function has an array of values and the time points definition for the start to
end and respective intervals as arguments.
The exponential smoothing model considers that a forecast is based on a weighted
mean, given a lower weight to an older observation. It uses only one parameter (in the
simpler form), but it can use more parameters to accommodate trend and seasonal
factors. It is estimated using a recursive function [9].
Both the classical decomposition method and the splinef use a simple model for the
trend, namely linear regression and a third degree polynomial at each data interval
respectively and use mean differences to the trend for seasonal factors.

Table 1. R functions from the package forecast used in this work (own construction)

Function

Description

Provides an optimized exponential smoothing model, which
can include level variations, trend and seasonal factors.

ets()

Gives a classical decomposition model, modeling the trend as
a linear regression and seasonal factors as mean differences
between the estimated trend and the time series values.

tslm()

Sets a third degree polynomial at each data interval and
provides the historical data to obtain a seamless line, allowing
shaping local level variations [10].

splinef()

Adjusts the ARIMA model that fits the data, optimizing not
only the orders of the autoregressive and moving average
components but also the need for differentiation of the series.

auto.arima()