Page 445 - Contributed Paper Session (CPS) - Volume 4
P. 445
CPS2564 Tiffany Rizkika et al.
3. Nowcasting
We use the different nowcasting approaches based on the volatility of the
food prices:
a. Volatile food price nowcasting
Historical data-based uses two methods, namely statistical modelling
using Distributed Lag Model (DLM) and machine learning using Neural
Network RPROP (NN RPROP), also using two types of periods daily and
weekly. Before modelling, pre-processing techniques are applied to
prepared data.
The step of data pre-processing for historical data-based approaches
is data cleaning, data transformation, smoothing, and data imputation.
Data cleaning includes filtering to filter data according to time and place
of study, removing incomplete record, extreme prices, and outliers. Data
transformation includes standardizing the unit price and calculating daily
and weekly price. Smoothing is used to minimize the fluctuation pattern,
using smoothing spline. Data imputation is needed to complete the
unavailable data in a certain day, also to get daily price from weekly price
using temporal disaggregation. After pre-processing, data is divided into
training and testing data. Training data is used to making models, while
testing data is used to calculate MAPE to get the best model.
Modelling using DLM involves data in the current and past time of the
independent variable X. According to Baltagi (2011), DLM is a dynamic
model that have a form as follows:
(1)
(2)
where, Y_t is the t-th observation on the dependent variable X, and X_ (t-
s) is an independent variable of observation. α is an intercept, and β_0, β_1,
... β_s are coefficients at the present time and at lag time, and u_t is a
stationary error. The lag is the time required for X independent variables
to influence non-independent variables Y (Supranto, 1995). One type of
distributed lag model is the infinite lag model, where the length of the lag
is known or determined. The amount of lag will affect the number of
observations used as samples. The more number of lags, the lower the
number of sample data.
Modelling using Neural Network solves the problem by learning from
training examples (Michael,2015). The algorithm that applied is Resilient
Backpropagation implements two stages of learning, namely the forward
propagation stage to get the output error and the backward propagation
stage to change the values of weights. Changing the weight and network
bias in NN RPROP, according to the gradient behaviour in each training
434 | I S I W S C 2 0 1 9