Identifying ARIMA Models

45
1 Identifying ARIMA Identifying ARIMA Models Models What you need to know What you need to know

description

Identifying ARIMA Models. What you need to know. Autoregressive of the second order. X(t) = b 1 x(t-1) + b 2 x(t-2) + wn(t) b 2 is the partial regression coefficient measuring the effect of x(t-2) on x(t) holding x(t-1) constant - PowerPoint PPT Presentation

Transcript of Identifying ARIMA Models

Page 1: Identifying ARIMA Models

11

Identifying ARIMA ModelsIdentifying ARIMA Models

What you need to knowWhat you need to know

Page 2: Identifying ARIMA Models

22

Autoregressive of the second orderAutoregressive of the second order

• X(t) = b1 x(t-1) + b2 x(t-2) + wn(t)

• b2 is the partial regression coefficient measuring the effect of x(t-2) on x(t) holding x(t-1) constant

• Since x(t) is regressed on itself lagged, b2

can also be interpreted as a partial autoregression coefficient of x(t) regressed on itself lagged twice.

Page 3: Identifying ARIMA Models

33

continuedcontinued• In one more step b2 can be defined as the

partial autocorrelation coefficient at lag 2, b2 = pacf(2)

• Solving the yule-Walker equations:

• b2 = {acf(2) – [acf(1)]2 }/[1 – [acf(1)]2

• We know that if the process is autoregressive of the first order, then acf(2) = [acf(1)]2 and so b2 = 0

Page 4: Identifying ARIMA Models

44

So now we are back to So now we are back to autoregressive of the first oderautoregressive of the first oder

• x(t) = b x(t-1) + wn(t)

• There is only one regression coefficient, b, so acf(1) = pacf(1) = b

Page 5: Identifying ARIMA Models

55

In summaryIn summary• The partial autocorrelation function, pacf(u)

indicates the order of the autregressive process. If only pacf(1) is significantly different from zero, then the autoregressive process is of order one. If the pacf(2) is significantly different from zero, then the autoregressive process is of order two, and so on.

• Thus we use the partial autocorrelation function to specify the order of the autoregressive process to be estimated

Page 6: Identifying ARIMA Models

66

The autocorrelation functionThe autocorrelation function

• The autocorrelation function, acf(u) is used to determine the order of the moving average process

• If acf(1) is significantly different from zero and there are no other significant autocorrelations, then we specify a first order MA process to be estimated

Page 7: Identifying ARIMA Models

77

Cont.Cont.

• If there is a significant autocorrelation at lag two and none at higher lags, then we specify a second order moving average process

Page 8: Identifying ARIMA Models

88

Moving Average ProcessMoving Average Process

• X(t) = wn(t) + a1wn(t-1) + a2wn(t-2) + a3wn(t-3)

• Taking expectations the mean function is zero, Ex(t) = m(t) = o

• Multiplying by x(t-1) and taking expectations, E[x(t)x(t-1)] =

• EX(t) = wn(t) + a1wn(t-1) + a2wn(t-2) + a3wn(t-3)

X(t-1) = wn(t-1) + a1wn(t-2) + a2wn(t-3) + a3wn(t-4), γx,x (1) = [a1 + a1 a2 + a2 a3 ] σ2

Page 9: Identifying ARIMA Models

99

ContinuingContinuing• The autocovariance at lag 2, γx,x (2) = E x(t) x(t-2)• EX(t) = wn(t) + a1wn(t-1) + a2wn(t-2) + a3wn(t-3)

X(t-2) = wn(t-2) + a1wn(t-3) + a2wn(t-4) + a3wn(t-5), γx,x (2) = [a2 + a1 a3 ] σ2

• The autocovariance at lag 3, γx,x (3) = E x(t) x(t-4)• EX(t) = wn(t) + a1wn(t-1) + a2wn(t-2) + a3wn(t-3)

X(t-3) = wn(t-3) + a1wn(t-4) + a2wn(t-5) + a3wn(t-6), γx,x (3) = [a3 ] σ2

• The autocovariance at lag 4 is zero, E x(t)x(t-4) = 0, so the autocovariance function determines the order of the MA process

Page 10: Identifying ARIMA Models

1010

Specifying ARMA ProcessesSpecifying ARMA Processes• x(t) = A(z)/B(z)• The autocovariance function divided by the variance,

i.e. the autocorrelation function, acf(u), indicates the order of A(z) and the partial autocorrelation function, pacf(u) indicates the order of B(z)

• In Eviews specify x(t) c ar(1) ar(2) ….ar(u) for a u th order B(z) and include ma(1) ma(2) ….ma(u) for a uth order A(z),

• i.e. X(t) c ar(1) ar(2) …ar(u) ma(1) ma(2) …ma(u)

Page 11: Identifying ARIMA Models

1111

Summary of IdentificationSummary of Identification• Spreadsheet

• Trace: Is it stationary?

• Histogram: is it normal?

• Correlogram: order of A(z) and B(z)

• Unit root test: is it stationary?1111

• Specification

• estimation

Page 12: Identifying ARIMA Models

1212

ARMA ProcessesARMA Processes• Identification• Specification and Estimation• Validation

– Significance of estimated parameters and DW– Actual, fitted and residual– Residual tests

• Correlogram: are they orthogonal? Also the Breusch-Godfrey test for serial correlation

• Histogram; are they normal?

• Forecasting

Page 13: Identifying ARIMA Models

1313

Example: Capacity utilization mfg.Example: Capacity utilization mfg.

Page 14: Identifying ARIMA Models

1414

SpreadsheetSpreadsheet

Page 15: Identifying ARIMA Models

1515

HistogramHistogram

0

10

20

30

40

50

60

66 68 70 72 74 76 78 80 82 84 86 88

Series: MCUMFNSample 1972:01 2010:03Observations 459

Mean 78.98322Median 79.30000Maximum 88.50000Minimum 65.20000Std. Dev. 4.647918Skewness -0.603442Kurtosis 3.051589

Jarque-Bera 27.90780Probability 0.000001

Page 16: Identifying ARIMA Models

1616

CorrelogramCorrelogram

Page 17: Identifying ARIMA Models

1717

Unit root testUnit root test

Page 18: Identifying ARIMA Models

1818

Pre-WhitenPre-Whiten

Gen dmcumfn =mcumfn – mcumfn(-1)

Page 19: Identifying ARIMA Models

1919

SpreadsheetSpreadsheet

Page 20: Identifying ARIMA Models

2020

TraceTrace

-4

-3

-2

-1

0

1

2

75 80 85 90 95 00 05 10

DMCUMFN

Page 21: Identifying ARIMA Models

2121

histogramhistogram

0

20

40

60

80

100

120

-4 -3 -2 -1 0 1 2

Series: DMCUMFNSample 1972:02 2010:03Observations 458

Mean -0.023799Median 0.000000Maximum 1.800000Minimum -3.900000Std. Dev. 0.660698Skewness -1.109196Kurtosis 7.351042

Jarque-Bera 455.1915Probability 0.000000

Page 22: Identifying ARIMA Models

2222

CorrelogramCorrelogram

Page 23: Identifying ARIMA Models

2323

Unit root testUnit root test

Page 24: Identifying ARIMA Models

2424

SpecificationSpecification

Dmcumfn c ar(1) ar(2)

Page 25: Identifying ARIMA Models

2525

Estimation

Page 26: Identifying ARIMA Models

2626

ValidationValidation

-4

-2

0

2

4

-4

-2

0

2

75 80 85 90 95 00 05 10

Residual Actual Fitted

Page 27: Identifying ARIMA Models

2727

Correlogram of the residualsCorrelogram of the residuals

Page 28: Identifying ARIMA Models

2828

Breusch-Godfrey Serial correlation testBreusch-Godfrey Serial correlation test

Page 29: Identifying ARIMA Models

2929

Re-SpecifyRe-Specify

Page 30: Identifying ARIMA Models

3030

EstimationEstimation

Page 31: Identifying ARIMA Models

3131

ValidationValidation

-4

-2

0

2

4

-4

-2

0

2

75 80 85 90 95 00 05 10

Residual Actual Fitted

Page 32: Identifying ARIMA Models

3232

Correlogram of the ResidualsCorrelogram of the Residuals

Page 33: Identifying ARIMA Models

3333

Breusch-Godfrey Serial correlation testBreusch-Godfrey Serial correlation test

Page 34: Identifying ARIMA Models

3434

Histogram of the residualsHistogram of the residuals

0

20

40

60

80

100

-3 -2 -1 0 1 2

Series: ResidualsSample 1972:05 2010:03Observations 455

Mean 8.09E-05Median 0.010460Maximum 2.669327Minimum -2.954552Std. Dev. 0.583609Skewness -0.316126Kurtosis 6.875041

Jarque-Bera 292.2558Probability 0.000000

Page 35: Identifying ARIMA Models

3535

Forecasting: Procs. Workfile rangeForecasting: Procs. Workfile range

Page 36: Identifying ARIMA Models

3636

Forecasting: Equation Forecasting: Equation window.forecastwindow.forecast

Page 37: Identifying ARIMA Models

3737

ForecastingForecasting

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

2.0

10:04 10:05 10:06 10:07 10:08 10:09 10:10 10:11 10:12

DMCUMFNF ± 2 S.E.

Page 38: Identifying ARIMA Models

3838

Forecasting: Quick, showForecasting: Quick, show

Page 39: Identifying ARIMA Models

3939

ForecastingForecasting

Page 40: Identifying ARIMA Models

4040

Forecasting: show, view, graph-lineForecasting: show, view, graph-line

-3

-2

-1

0

1

2

00 01 02 03 04 05 06 07 08 09 10

DMCUMFNFORECAST

+2*SEFFORECAST-2*SEF

Page 41: Identifying ARIMA Models

4141

ReintegrationReintegration

Page 42: Identifying ARIMA Models

4242

Forecasting mcumfnForecasting mcumfn

Page 43: Identifying ARIMA Models

4343

Forecast mcumfn, quick, showForecast mcumfn, quick, show

Page 44: Identifying ARIMA Models

4444

Forecasting mcumfnForecasting mcumfn

64

68

72

76

80

84

00 01 02 03 04 05 06 07 08 09 10

MCUMFNMCUMFNF

MCUMFNF+2*SEFMCUMFNF-2*SEF

Page 45: Identifying ARIMA Models

4545

What can we learn from this forecast?What can we learn from this forecast?• If, in the next nine months, mcumfn grows

beyond the upper bound, this is new information indicating a rebound in manufacturing

• If, in the next nine months, mcumfn stays within the upper and lower bounds, then this means the recovery remains sluggish

• If mcumfn goes below the lower bound, run for the hills!