Exploring Data Patterns: Time Series
Uploaded by
Sylvia CheungExploring Data Patterns: Time Series
Uploaded by
Sylvia CheungExploring Data Patterns
A time series consists of data that are recorded
over regular or irregular intervals of time.
e.g. Daily closing price of a stock
Strength of earthquakes
Denote a time series by
Yt ; (t = 1, 2, .)
or
Y(t) ; (t = 1, 2, .)
Slide 1
Components of a Time Series
Trend - data observations grow or decline over an extended
period of time
Seasonal pattern - change that repeats itself year after year.
Noise - fluctuations that are not part of the other two
components.
Slide 2
Data Patterns
Trend line
Demand for product
Seasonal peaks
Actual
demand curve
Noise
Year
1
Year
2
Year
3
Year
4
Slide 3
Forecast Error
Yt = observed value of time series at period t
t = forecast value of Yt
et
= Yt t = forecast error
Slide 4
Assessing Forecast Accuracy
MAD (Mean Absolute Deviation) : measures forecast
accuracy by averaging the absolute value of the forecast
errors.
1 n
=
MAD
Yt Yt
n t =1
Useful when one wants to measure forecast error in the
same units as the original data.
Slide 5
Assessing Forecast Accuracy
MSE (Mean Squared Error) - each error is squared, then
summed and divided by the number of errors. Magnifies
large forecasting errors because of the squares.
1 n
2
=
MSE
Y
Y
(
)
t
t
n t =1
RMSE (Root Mean Squared Error)
=
RMSE
Y
(
)
t t n
t =1
Slide 6
Nave Methods
Suitable for small data sets
Basic naive method
Naive method adjusting for (linear) trend
Slide 7
Basic Nave
Suitable for data without a trend
Forecast made at time t for time t+1:
t+1 = Yt (for data without seasonal variation)
t+1 = Yt-3 (for seasonal quarterly data)
t+1 = Yt-11 (for seasonal monthly data)
Slide 8
Basic Nave : Example
t
Year
Quarter
Sales
2007
850
600
450
700
Forecast the sales for Q1 2008 :
Y=
Y4+=
Y=
700
5
1
4
Slide 9
Basic Nave : Example
t
Year
Quarter
Sales
2006
500
350
250
400
850
600
450
700
2007
Forecast the sales for Q1 2008 :
Y= Y = Y= 850
9
94
Slide 10
Basic Nave : Whats Wrong ?
Slide 11
Nave : Adjusting for Trend
Slide 12
Averaging Methods
Moving average
Double moving average
Slide 13
Moving Average
The average of the most recent k observations, where k is chosen
to minimize a certain error measure, is taken. Averages computed
this way are called moving averages of order k.
Works best with trendless data
The t-th moving average of order k is calculated according to the
formula :
where t k.
M t := Yt +Yt-1+...+Yt-( k-1) k
To forecasting the next time period, we set :
Yt+1= M t
Slide 14
Moving Average Layout
Slide 15
Forex Moving Averages : GBP/USD
Slide 16
Double Moving Averages (DMA)
Designed to handle data with (linear) trend.
One set of moving averages is calculated and then a
second set is calculated as a moving average of the
first set.
Slide 17
Formulas for DMA
1. Calculate first set of moving averages :
M t = Yt +Yt-1+...+Yt-( k-1) k
2. Calculate second set of moving averages :
M t = M t +M t-1+...+M t-( k-1) k
Slide 18
Formulas for DMA
3. Calculate new level :
'
at = M t + ( M t M t )
4. Calculate new slope :
bt =
2( M t M ) ( k 1)
'
t
5. Forecast p periods ahead :
Yt+p= at + pbt
Slide 19
Meaning of at and bt
Slide 20
