A Decision Support System for Municipal Budget Plan Decisions - New Contributions in Information Systems and Technologies - page 127

Table 2. Several models for the direct taxation indicator

3.2 Discussion

The applied forecasting models are based on annual values because we intend to
understand their behavior, creating knowledge concerning forecasts, which is used for
the preparation of the local authority budget plan. The municipality only has data for
12 years. However, identification of seasonality involves patterns in constant periods,
given the size of the available series it is not surprising that it was not possible to
identify neither seasonality nor economic cycles.
Three methods were applied for each indicator to build predictive models, as
shown in Table 3: exponential smoothing, classical decomposition modeling the trend
with linear regression, and ARIMA. We calculate the respective prediction errors
using the measures mean root square error (RMSE) and average percentage error
(APE). Table 4 shows the main indicators errors used in accordance with each
prediction model. Values in bold are the error values closer to zero. For the case
where the two error measures are not consistent, the decision on which prediction to
use is let to the decision maker, according to his\her domain knowledge.
It was also considered the use of splinef() function. However, this function reveled
to be less useful because of too high error measurements calculated, which results in
poor fit for all indicators. For example, for the indicator of Current Transfers, the
value obtained for RMSE was 355,911.7. For indicator of Current Revenue, the value
obtained for RMSE was 1,498,773 and the APE was -7.0. For this reason, and also
because that predictions in these models only use a linear regression between the last
two values, this feature was not included in the DSS.