Vous êtes sur la page 1sur 5

Forecasting – Time Series

Time Series Analysis using SPSS

Planning for the future is an essential ingredient of any good household, company,
corporate or government. In order to do this, it is important to forecast events to ensure
smooth transition. In many situations, forecasts are made on ad-hoc basis. But a
scientific and fairly foolproof forecasting can be made by creating a catalogue of past and
present information on any factor that is the center of study and see whether there are any
patterns can be identified in this data. This type of forecasting is called time series
analysis.

A time series is a variable whose values represent equally spaced observations of a


phenomenon of time. Examples – Annual sale of cars, quarterly interest rate, yearly
admission of patients in a hospital, daily share prices etc.

Time series models Vs. Econometric models

Time series models explain the movement of any variable in terms of its own past
movement and its position with respect to time. It does not study the possible impact of
any other variable on it.

Time series Data

In time series data each column corresponds to a particular variable. The important point
to be noticed for time series data is that each row of data should correspond to a
particular period of time. The data should be collected , or at least be summarized over
equal time periods. In the time series data file it is essential that the data should be on
equally spaced time periods. If data is missing for any one of the time period, a row
should be included for that time period at least with a missing value for the period.

Sequence Chart

The sequence chart is the simplest way to have a graphic representation of the data on
hand. It will be the initial study to understand the basic pattern that is noticed in the data.
The sequence chart can be used to detect the trend, seasonal and cyclical patterns that can
be seen in time series.

A time series model is a tool used to predict future values of a series by analyzing the
relationship between the values observed in the series and the time of their occurrence.
There are many techniques used in time series analysis like Exponential smoothening,
autoregression, ARIMA and seasonal decomposition .

Types of time series models

Pure time series models use information solely from the series itself without attempting
to find out the effect of other factors which affect the movement of the time series.

Ansi Techniques, Chennai  2826 2373 / 98840 63145 / ansitech.in


Forecasting – Time Series

Exponential smoothening is an example of this type of a model. It is easy and fast to use
and forecast. But the drawback of this model is that it can not identify the factors which
affect the series. As a natural extension, this model can not study the impact of any
decision made by the company / Govt of the series.

Causal time series models like Regression and Arima predict future values of a time
series data wherein a dependend variable ( the time series data is our example ) is studied
based on the impact of time and a set of independent variables. The benefit of such a
model is that it takes more variables and details into account while forecasting. That itself
becomes a drawback of the model in that lot of time energy and resources are to be used
to collect the independent variables. Also decisions are to be made as to which of the
variables are to be used in the model. Also the model has to forecast the future values of
the independent variables before using them as an input to predict the dependant
variable.

Seasonal adjustment models such as seasonal decomposition automatically find out the
seasonal, trend and error components in the series.

Steps in creating a time series forecasting model.

1. Use basic plots to identify patterns in the data.


2. Use appropriate time series model
3. See how well each model fitted the data
4. Compare the performance of different models.
5. Choose the one which has fitted the actual data better.
6. Use this model to forecast.

Data Requirements

Defining the periodicity of a time series. This periodicity is linked to the idea of
seasonality in time series. The number of periods that occur during the completion of a
seasonal pattern is referred to as the series periodicity. How often the time series data is
collected generally depends on the seasonality.

Two questions that are to be answered while collecting time series data are

a) How many data points to collect and


b) How often should the data be collected.

Though there is no clear cut formula to find a solution for the above mentioned questions,
the purpose of research, the forecasting needs and the seasonality seen in the data form an
important component in deciding on the size and periodicity of collection of data of the
time series.

Ansi Techniques, Chennai  2826 2373 / 98840 63145 / ansitech.in


Forecasting – Time Series

Data Smoothing

Once we have collected the data , the next step will be to look at the data and look out for
any patterns, smoothen the data to clear it of all noises that it may have.Smoothing the
data in order to reduce the noise in the data is achieved by using moving averages,
running medians etc.

Exponential Smoothing

The next technique that we study regarding forecasting time series is Exponential
smoothing. This procedure produces forecast values and residuals for one or more time
series using an algorithm that smooths out irregular components of the time series data.
A variety of models differing in trend (none, linear or exponential) and seasonality (none,
additive or multiplicative ) are available. This is a pure time series technique in that the
forecast is based on the data collected of the series itself. It takes the approach that the
most recent observations should have greater weight in forecasting the series than distant
observations.

The model performance can be tested by

a. See how well the model fits the actual data by seeing the graph of both these data in the
sequence chart

b. If the model fits the data well, then the error value will be nearer to zero. Also
constant variance of the error term ( heteroscedasticity) over the time can be analysed by
looking at the sequence chart of the error term

c. Analysing the error sequence chart to see whether any pattern is seen in the residuals
which will denote that the model had not captured a few patterns in the data.
( autocorrelation and partial correlation charts)

d. Testing whether the error are normally distributed.

e. Split data into training and testing component and check the efficacy of the model.

Seasonal Decomposition.

The seasonal decomposition procedure decomposes a series into seasonal component, a


combined trend and cycle component and an “error” component.

The seasonal decomposition procedure removes periodic fluctuations from time series ,
such as annual or seasonal highs or lows. Thus if we have data in which seasonality is
present, then applying a smoother , such as moving average , whose span is equal to the
number of periods in the season, should largely smooth out the seasonal variation.

Ansi Techniques, Chennai  2826 2373 / 98840 63145 / ansitech.in


Forecasting – Time Series

The seasonal decomposition method creates another series which is called ‘seasonally
adjusted series’ This is the original series from which seasonal variations are removed. In
other words, the new series has variations due to trend, cyclical and error components
only.

Autoregression

Autoregression procedure is an extension of ordinary least square regression analysis


specifically designed for time series.

When we try and study the impact of one (or more) variables on a dependant variable,
and if the variables are collected over a period of time, a standard regression model may
not be a good solution. Most time series have some trend, either up or down and any
two trending series will correlate simply because of the trends , regardless of whether
they are causally related or not. Autoregression helps to remove the autocorrelation
inherent in many time series and ascertain any statistically significant relationship
between the variables. The Autoregression procedure accounts for first order
autocorrelated residuals and provides reliable estimates of both goodness of fit measures
and significance levels of chosen predictor variables. It can not be used when a series
contains missing values.

At times , the seasonal decomposition procedure is used on the series to remove the
seasonal effects before the autoregression procedure is used.

ARIMA

ARIMA is a more complex modeling technique. It is also called Box-Jenkins model since
they were the developers of this model. ARIMA stands for AutoRegressive Integrated
Moving Average and the assumption of this model is that the common variation in a
dependent variable can be divided into three components namely

1. Autoregressive
2. Integrated or difference and
3. Moving Average.

The final model may consist of any one of the above, a combination of any two or
presence of all the three components. ARIMA models provide more sophisticated
methods for modeling trend and seasonal components than do exponential smoothing
models , and they allow the added benefit of including predictor variables in the model.

The general form of ARIMA model is ARIMA (p,d,q) where

“p” refers to the order of autoregressive process incorporated in the ARIMA model
“d” refers to the order of integration or differencing and
“q” refers to the order of the moving average process incorporated in the model.

Ansi Techniques, Chennai  2826 2373 / 98840 63145 / ansitech.in


Forecasting – Time Series

The determination of the values of p, d and q representing the autoregressive ,


differencing and Moving average component is a very crucial step in building the model.
For a seasonal model, the seasonal counterparts to these components should be specified.
Since the three types of random processes in ARIMA models are closely related, there is
no algorithm to determine the correct model. Instead there is a model building procedure
that allows to build the best possible model for the series.

The residuals calculated are used to diagnose which of the model is the best. If the model
is to be a good fit of the series, then the residual has to be random. The autocorrelation
and partial autocorrelation function of the residual should be significantly different from
zero .The residuals should be without pattern . The Box – Ljung Q statistic is the best test
for this.

Ansi Techniques, Chennai  2826 2373 / 98840 63145 / ansitech.in

Vous aimerez peut-être aussi