Scribd
Upload
110 views6 pages

ARMA Estimation of US Housing Prices

1. The document describes estimating an ARMA model to forecast US housing price indexes. It identifies the best fitting ARMA(1,1) model based on minimizing information criteria. 2. The ARMA(1,1) model estimates monthly housing price returns have an unconditional mean of 1.438%, with the price index influenced by previous month's return and previous month's forecast error. 3. Diagnostic tests show the ARMA(1,1) model adequately captures patterns in returns, improving over a simple mean model. Out-of-sample forecasts are less precise than in-sample but converge to the unconditional mean over longer horizons.

Uploaded by

Shadowfax_u
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
110 views6 pages

ARMA Estimation of US Housing Prices

1. The document describes estimating an ARMA model to forecast US housing price indexes. It identifies the best fitting ARMA(1,1) model based on minimizing information criteria. 2. The ARMA(1,1) model estimates monthly housing price returns have an unconditional mean of 1.438%, with the price index influenced by previous month's return and previous month's forecast error. 3. Diagnostic tests show the ARMA(1,1) model adequately captures patterns in returns, improving over a simple mean model. Out-of-sample forecasts are less precise than in-sample but converge to the unconditional mean over longer horizons.

Uploaded by

Shadowfax_u
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

1

An Example of ARMA Estimation

ARMA estimation can be broken into 3 stages: model identication, model estimation and forecasting. Consider an example of ARMA estimation for the monthly price index of US housing, based on repeat sales. Data was obtained from the Oce of Federal Housing Enterprise Oversight. Figure 1 plots the raw data which ranges from 1975:1 - 2003:4. There is a clear trend upward suggesting nonstationary behaviour. Figure 2 displays the autocorrelation function (ACF).
1. US Housing Price Index 350

300

250

200

150

100

50 0 20 40 60 2. ACF 1 80 100 120

0.5

-0.5

-1 0 5 10 15 20 25 30 35 40 45 50

To remove the nonstationary structure in the data we will consider the growth rate of the series or the percent return. A more formal approach to checking for nonstationary behaviour is to conduct a unit-root test which we omit. If pt is the housing price index for time t, then dene the return in percent as gt = 100 log pt pt1 .

Table 1 provides summary statistics for housing returns while Figure 3 displays the series. Table 1: Summary Statistics, Housing Returns Mean Stdev Min Max 4.3444

1.3899 0.94413 -0.68597

3. US Housing Returns 5

-1 0 20 40 60 80 100 120

Next consider the ACF and PACF for the data.

4. ACF, Housing Returns 1

0.5

-0.5

-1 0 5 10 15 20 25 30 35 40 45 50

5. PACF, Housing Returns 1

0.5

-0.5

-1 0 5 10 15 20 25 30 35 40 45 50

These plots suggest a low order ARMA model may be appropriate. Table 2 reports the Schwarz and Akaike model criteria for various models. The ARMA(1,1) model minimizes both criteria and produces a good R2 . Next consider the ARMA(1,1) model Table 2: Model Selection Criteria Model ARMA(1,0) ARMA(2,0) ARMA(1,1) ARMA(2,1) ARMA(1,2) SCHWARZ AKAIKE -0.44679 -0.47263 -0.53069 -0.48088 -0.48106 -0.49453 -0.54424 -0.60230 -0.57635 -0.57654 R2 0.3450 0.3906 0.4264 0.4264 0.4266

estimates found in Table 3. All parameters are measured accurately except for the coecient. We can see that this model implies an unconditional mean of Eyt = c .11214 = = 1.438 1 1 .92200

which is very close to the mean estimate in Table 1. Also note the reduction in the innovation variance from .89 (from Table 1 .9442 ) to .519. Clearly the ARMA model is a big improvement over just a mean estimate. The second panel of Table 3 displays Table 3: ARMA(1,1) Model Estimates gt = c + gt1 + t1 + t , t N ID(0, 2 ) Coecient c 2 Ljung-Box Statistics k 5 10 20 30 Estimate 0.11214 0.92200 0.57191 0.51972 Q(k ) 3.32 10.15 22.19 28.84 Standard Error 0.0753 0.0510 0.1070 T-statistic 1.488 18.08 5.344

p-value .344 .255 .224 .421

Ljung-Box tests for remaining serial correlation in the model residuals. The p-values 4

suggest that the ARMA(1,1) does an adequate job in accounting the structure in the data (ie. no remaining serial correlation in the residuals). Figure 6 displays the in-sample one step ahead forecast from the model against the real data. the forecast is calculated as Et gt+1 = c + gt +
t1

using the parameters estimates. The last gure displays the out-of-sample forecast. In this case the model is estimated with observations 1 - 75 and thereafter forecast are produced beyond this point. Clearly the out-of-sample forecasts are much less precise than the in-sample forecasts. Also note that the forecast, as it goes out further, is converging to the unconditional value of 1.438. The data and shazam le for calculating these results can be obtained from the class web site.

6. In-sample forecast of ARMA(1,1) 5

-1 0 20 40 60 80 100 120

7. Out-of-Sample forecast of ARMA(1,1) from Observation 75 with 95% confidence interval 4 3.5 3 2.5 2 1.5 1 0.5 0 -0.5 -1 75 80 85 90 95 100 105 110 115 120

You might also like