The Econometric Modeler app is an interactive tool for analyzing univariate time series data. The app is well suited for visualizing and transforming data, performing statistical specification and model identification tests, fitting models to data, and iterating among these actions. When you are satisfied with a model, you can export it to the MATLAB® Workspace to forecast future responses or for further analysis. You can also generate code or a report from a session.
Start Econometric Modeler by entering econometricModeler
at the
MATLAB command line, or by clicking Econometric Modeler
under Computational Finance in the apps gallery
(Apps tab on the MATLAB Toolstrip).
The following workflow describes how to find a model with the best in-sample fit to time series data using Econometric Modeler. The workflow is not a strict prescription—the steps you implement depend on your goals and the model type. You can easily skip steps and iterate several steps as needed. The app is well suited to the Box-Jenkins approach to time series model building [1].
Prepare data for Econometric Modeler — Select a response variable to analyze and from which to build a predictive model. Optionally, select explanatory variables to include in the model.
You can import only one variable from the MATLAB Workspace into Econometric Modeler. Therefore, at the command line, you must synchronize and concatenate multiple series into one variable.
Import time series variables — Import Data into Econometric Modeler from the MATLAB Workspace or a MAT-file. After importing data, you can adjust variable properties or the presence of variables.
Perform exploratory data analysis — View the series in various ways, stabilize a series by transforming it, and detect time series properties by performing statistical tests.
Visualize time series data — Supported plots include time series plots and correlograms (for example, the autocorrelation function (ACF)).
Perform specification and model identification hypothesis tests — Test series for stationarity, heteroscedasticity, autocorrelation, and collinearity among multiple series. For ARIMA and GARCH models, this step can include determining the appropriate number of lags to include in the model. Supported tests include the augmented Dickey-Fuller test, Engle's ARCH test, the Ljung-Box Q-test, and Belsley collinearity diagnostics.
Transform time series — Supported transformations include the log transformation and seasonal and nonseasonal differencing.
Fit candidate models to the data — Choose model parametric forms for a univariate response series based on the exploratory data analysis or dictated by economic theory. Then, estimate the model. Supported models include seasonal and nonseasonal conditional mean (for example, ARIMA), conditional variance (for example, GARCH), and multiple linear regression models (optionally containing ARMA errors).
Conduct goodness-of-fit checks — Ensure that the model adequately describes the data by performing residual diagnostics.
Visualize the residuals to check whether they are centered on zero, normally distributed, homoscedastic, and serially uncorrelated. Supported plots include quantile-quantile and ACF plots.
Test the residuals for homoscedasticity and autocorrelation. Supported tests include the Ljung-Box Q-test and Engle's ARCH test on the squared residuals.
Find the model with the best in-sample fit — Estimate multiple models within the same family, and then choose the model that yields the minimal fit statistic, for example, Akaike information criterion (AIC).
Export session results — After you find a model or models that perform adequately, summarize the results of the session. The method you choose depends on your goals. Supported methods include:
Export variables — Econometric Modeler exports selected variables to the MATLAB Workspace. If a session in the app does not complete your analysis goal, such as forecasting responses, then you can export variables (including estimated models) for further analysis at the command line.
Generate a function — Econometric Modeler generates a MATLAB function that returns a selected model given the imported data. This method helps you understand the command-line functions that the app uses to create predictive models. You can modify the generated function to accomplish your analysis goals.
Generate a report — Econometric Modeler produces a document, such as, a PDF, describing your activities on selected variables or models. This method provides a clear and convenient summary of your analysis when you complete your goal in the app.
You can import only one variable from the MATLAB Workspace into Econometric Modeler. Therefore, before importing data, concatenate the response series and any predictor series into one variable.
Econometric Modeler supports these variable data types.
MATLAB timetable — Variables must be double-precision numeric vectors. A best practice is to import your data in a timetable because Econometric Modeler:
Names variables by using the names stored in the
VariableNames
field of the
Properties
property.
Uses the time variable values as tick labels for any axis that represents time. Otherwise, tick labels representing time are indices.
Enables you to overlay recession bands on time series
plots (see recessionplot
)
MATLAB table — Variables must be double-precision numeric
vectors. Variable names are the names in the
VariableNames
field of the
Properties
property.
Numeric vector or matrix — For a matrix, each column is a separate
variable named variableNamej
, where
j
is the corresponding column.
Regardless of variable type, Econometric Modeler assumes that rows correspond to time points (observations).
The data set can exist in the MATLAB Workspace or in a MAT-file that you can access from your machine.
To import a data set from the Workspace, on the Econometric
Modeler tab, in the Import section,
click . In the Import
Data dialog box, click the check box in the
Import? column for the variable containing the
data, and then click Import. All variables in the
Workspace of the supported data type appear in the dialog box, but you
can choose one only.
To import data from a MAT-file, in the Import
section, click Import, then select
Import From MAT-file
. In the
Select a MAT-file dialog box, browse to the
folder containing the data set, then double-click the MAT-file.
After you import data, the name of each variable (column) in the data set appears
in the Time Series section of the Data
Browser. Also, a time series plot including all variables appears in
the Time Series Plot(VariableName
)
figure window, where VariableName
is the name of one of
the variables in the Data Browser.
You can interact with the variables in the Data Browser in several ways.
To select a variable to perform a statistical test or create a plot, for example, click the variable in the Data Browser. If you double-click the variable instead, then the app also plots it in a separate time series plot.
To change a variable name or delete a variable, right-click it in the Data Browser. Then, from the context menu, choose the desired action.
To operate on multiple time series simultaneously, press Ctrl and click each variable you want to use.
Consider importing the data in the Data_USEconModel
MAT-file.
At the command line, load the data into the Workspace.
load Data_USEconModel
In Econometric Modeler, in the Import section of
the Econometric Modeler tab, click . The Import Data
dialog box appears.
Data_USEconModel
stores several variables.
DataTable and
Data contain the same data, but
DataTable is a timetable that attributes
names to variables and sampling times to rows. Import
DataTable by selecting the corresponding
Import? check box, then clicking
Import.
All variables in DataTable appear in the Data
Browser. Suppose that you want to retain
COE
, FEDFUNDS
, and
GDP
only. Select all other variables, right-click one
of them, and select Delete.
After working in the app, you can import another data set. After you click Import, Econometric Modeler displays the following dialog box.
If you click OK, then Econometric Modeler deletes all variables from the Data Browser, and closes all documents in the right pane.
An exploratory data analysis includes determining characteristics of your variables and relationships between them, with the formation of a predictive model in mind. For time series data, identify series that grow exponentially, contain trends, or are nonstationary, and then transform them appropriately. For ARIMA models, to identify the model form and significant lags in the serial correlation structure of the response series, use the Box-Jenkins methodology [1]. If you plan to create GARCH models, then assess whether the series contain volatility clustering and significant lags. For multiple regression models, identify collinear predictors and those predictors that are linearly related to the response.
For time series data analysis, an exploratory analysis usually includes iterating among visualizing the data, performing statistical specification and model identification tests, and transforming the data.
After you import a data set, Econometric Modeler selects all variables in the
imported data and displays a time series plot of them in the right pane by
default. For example, after you import DataTable
in
the Data_USEconModel
data set, the app displays this time
series plot.
To create your own time series plot:
In the Data Browser, select the appropriate number of series for the plot.
Click the Plots tab in the toolstrip.
Click the button for the type of plot you want.
Econometric Modeler supports the following time series plots.
Plot | Goals |
---|---|
Time series |
|
Autocorrelation function (ACF) |
|
Partial ACF (PACF) |
|
Correlations |
|
You can interact with an existing plot by:
Right-clicking it
Using the plot buttons that appear when you pause on the plot
Using the options on the figure window
Supported interactions vary by plot type.
Save a figure — Right-click the figure, then select Export. Save the figure that appears.
Add or remove time series in a plot — Right-click the figure, point to Show Time Series menu, then select the time series to add or remove.
Plot recession bands — Right-click a time series plot, then select Show Recessions.
Zoom — Pause on the figure. To zoom in, click . To zoom out, click
. For more details, see Zoom, Pan, and Rotate Data (MATLAB).
Pan — Pause on the figure, then click . For more details on panning, see
Zoom, Pan, and Rotate Data (MATLAB).
Toggle legend — Pause on the figure, then click .
Show grid lines — Pause on the figure, then click .
For serial correlation function plots, additional options exist in the ACF or PACF tabs. You can specify the:
Number of lags to display
Number of standard deviations for the confidence bands
MA or AR order in which the theoretical ACF or PACF, respectively, is effectively zero
Econometric Modeler updates the plot in real time as you adjust the parameters.
To view multiple plots simultaneously, you can orient plot tabs by dragging
them into different sections of the right pane. As you drag a plot, the app
highlights possible sections in which to place it. To undo the last document or
figure window positioning, pause on the dot located in the middle of the
partition, then click when it appears.
Consider an ARIMA model for the effective federal funds rate
(FEDFUNDS
). To identify the model characteristics (for
example, the number of AR or MA lags), plot the time series, ACF, and PACF side-by-side.
In the Data Browser, double-click
FEDFUNDS
.
Add recession bands to the plot by right-clicking the plot in the Time Series Plot(FEDFUNDS) figure window, then selecting Show Recessions.
On the Plots tab, click ACF.
Click PACF.
Click the Time Series Plot(FEDFUNDS) figure window and drag it to the left side of the right pane. Click the PACF(FEDFUNDS) figure window and drag it to the bottom right of the pane.
The ACF dies out slowly and the PACF cuts off after the first lag. The behavior of the ACF suggests that the time series must be transformed before you choose on the form of the ARIMA model.
In the right pane, observe the dot in the middle of the horizontal partition
between the correlograms (below the Lag
x axis label of the ACF). To undo this correlogram
positioning, that is, separate the correlograms by tabs, pause on the dot and
click when it appears.
You can perform hypothesis tests to confirm time series properties that you obtain visually or test for properties that are difficult to see. Econometric Modeler supports these tests for univariate series.
Test | Hypotheses |
---|---|
Augmented Dickey-Fuller | H0: Series has a unit root. H1: Series is stationary. For details on the
supported parameters, see |
Kwiatkowski, Phillips, Schmidt, Shin (KPSS) | H0: Series is trend stationary. H1: Series has a unit root. For details on the
supported parameters, see |
Leybourne-McCabe | H0: Series is a trend stationary AR(p) process. H1: Series is an ARIMA(p,1,1) process. To specify p,
adjust the Number of Lags
parameter. For details on the supported parameters, see
|
Phillips-Peron | H0: Series has a unit root. H1: Series is stationary. For details on the
supported parameters, see |
Variance ratio | H0: Series is a random walk. H1: Series is not a random walk. For details on
the supported parameters, see |
Engle's ARCH | H0: Series exhibits no conditional heteroscedasticity (ARCH effects). H1: Series is an ARCH(p) model, with p > 0. To specify
p, adjust the Number of
Lags parameter. For details on the
supported parameters, see |
Ljung-Box Q-test | H0: Series exhibits no autocorrelation in the first m lags, that is, corresponding coefficients are jointly zero. H1: Series has at least one nonzero autocorrelation coefficient ρj, j ∈ {1,…,m}. To specify
m, adjust the Number of
Lags parameter. For details on the
supported parameters, see |
Before conducting tests, Econometric Modeler removes leading and
trailing missing values (NaN
values) in the series.
Engle's ARCH test does not support missing values within the series,
that is, NaN
values preceded and succeeded by
observations.
The stationarity test results suggest whether you should transform a series to stabilize it, and which transformation is appropriate. For ARIMA models, stationarity test results suggest whether to include degrees of integration. Engle's ARCH test results indicate whether the series exhibits volatility clustering and suggests lags to include in a GARCH model. Ljung-Box Q-test results suggest how many AR lags are required in an ARIMA model.
To perform a univariate test in Econometric Modeler:
Select a variable in the Data Browser.
On the Econometric Modeler tab, in the Tests section, click New Test.
In the test gallery, click the test you want to conduct. A new tab for the test type appears in the toolstrip, and a new document for the test results appears in the right pane.
On the test type tab, in the Parameters section, adjust parameters for the test. For example, consider performing an Engle's ARCH test. On the ARCH tab, in the Parameters section, select the number of lags in the test statistic using the Number of Lags spin box, or the significance level (that is, the value of α) using the Significance Level spin box.
On the test-type tab, in the Tests section, click Run Test. The test results, including whether to reject the null hypothesis, the p-value, and the parameter settings, appear in a new row in the Results table of the test results document. If the null hypothesis has been rejected, then the app highlights the row in yellow.
You can adjust parameters and run the test multiple times. The results from each test run for a particular variable appear as a new row in the Results table. To remove a row from the Results table, select the corresponding check box in the Select column, then click Clear Tests in the test-type tab.
Multiple testing inflates the false discovery rate. One conservative way to maintain an overall false discovery rate of α is to apply the Bonferroni correction to the significance level of each test. That is, for a total of t tests, set Significance Level value to α/t.
For multiple series, you can assess the strength and sources of collinearity
among the series using the Belsley collinearity diagnostics . To perform Belsley correlation diagnostics:
Select at least two variables in the Data Browser.
On the Econometric Modeler tab, in the Tests section, click New Test.
In the tests gallery, in the Collinearity section, click Belsley Collinearity Diagnostics. A new tab for Belsley collinearity diagnostics appears in the toolstrip, and a new document for the results appears in the right pane.
On the Collinearity tab, in the Tolerances section, adjust parameters for the test. The app performs diagnostics in real time as you adjust parameter values.
Econometric Modeler returns a table of the singular values, condition indices, and the variance-decomposition proportions for each variable. Rows Econometric Modeler highlights in yellow have a condition index greater than the tolerance specified by the Condition Index parameter value in the Tolerances section of the Collinearity tab. Also, Econometric Modeler plots the variance-decomposition proportions for the highlighted rows for each variable.
In the highlighted rows, those variables with variance-decomposition greater
than the tolerance (or, those variables with red markers in the plot) exhibit
multicollinearity. For more details on the Belsley collinearity diagnostics
results and multicollinearity, see collintest
and Time Series Regression II: Collinearity and Estimator Variance.
To add or remove time series from the diagnostics:
In the test-results document, right-click the Results table or plot.
Point to Show Time Series. A list of all variables appears.
Click a variable to add it to the diagnostics, or click a selected variable to remove it from the diagnostics.
Consider a predictive model containing Canadian inflation and interest rates
as predictor variables. Determine whether the variables are collinear. The
Data_Canada
data set contains the time series.
Import the DataTable
variable in the
Data_Canada
data set into Econometric
Modeler (see Import Time Series Variables).
The time series plot appears in the right pane.
All series appear to contain autocorrelation. Although you should remove autocorrelation from predictor variables before creating a predictive model, this example proceeds without removing autocorrelation.
In the Tests section, click New Test. In the Collinearity section, click Belsley Collinearity Diagnostics.
Econometric Modeler creates a document containing the Belsley collinearity diagnostics results.
The condition index and variance-decomposition proportion
tolerance defaults are 30
and 0.5
,
respectively. Because their variance-decomposition proportions are above the
tolerance for the condition index, the collinear predictors are
INT_L
, INT_M
, and
INT_S
.
The Box-Jenkins methodology [1] for ARIMA model selection assumes that the response series is stationary, and spurious regression models can result from a model containing nonstationary predictors and response variables (for more details, see Time Series Regression IV: Spurious Regression). To stabilize your series, Econometric Modeler supports these transformations in the Transforms section of the Econometric Modeler tab.
Transformation | Use When Series ... | Notes |
---|---|---|
Log | Has an exponential trend or variance that grows with its levels | All values in the series must be positive. |
Linear detrend | Has a linear deterministic trend that can be identified using least squares | When Econometric Modeler detrends the series, it
ignores leading or trailing missing
( If any
missing values occur between observed values, then the
app returns a vector of |
First-order difference | Has a stochastic trend | Econometric Modeler prepends the differenced series with
a NaN value. This action ensures that the
differenced series has the same length and time base as the
original series. |
Seasonal difference | Has a seasonal, stochastic trend | You can specify the period in a season using the
spin box. For example, Econometric
Modeler prepends the differenced series with
|
For more details, see Data Transformations.
To transform a variable, select the variable in the Data Browser, then click the desired transformation. After you transform a series, a new variable representing the transformed series appears in the Data Browser. Also, Econometric Modeler plots and selects the new variable. To create the variable name, the app appends the transformation name to the end of the variable name. You can rename the transformed variable by right-clicking it in the Data Browser, selecting Rename in the context menu, then typing the new name.
For example, suppose that the GDP series in
Data_USEconModel
has an exponential trend and a
stochastic trend. Stabilize the GDP by applying the log transformation and then
applying the second difference.
Import the DataTable
variable in the
Data_USEconModel
data set into the
Econometric Modeler (see Import Time Series Variables).
In the Data Browser, select
GDP
.
On the Econometric Modeler tab, in the
Transforms section, click
Log. The app creates a variable named
GDPLog
, which appears in the
Data Browser, and displays a plot for the
time series.
In the Transforms section, click
Difference. The app creates a
variable named GDPLogDiff
and displays a
plot for the time series.
In the Transforms section, click
Difference. The app creates a variable
called GDPLogDiffDiff
and displays a plot
for the time series.
GDPLogDiffDiff
is the stabilized
GDP.
The results of an exploratory data analysis can suggest several candidate models. To choose a model, in the Data Browser, select a time series variable for the response, then, on the Econometric Modeler tab, in the Models section, click a model or click one in the models gallery. Econometric Modeler supports the following models.
Model | Type |
---|---|
Conditional mean: ARMA/ARIMA Models section | Stationary autoregressive (AR) For details,
see Autoregressive Model, |
Stationary moving average (MA) For details,
see Moving Average Model, | |
Stationary ARMA For details,
see Autoregressive Moving Average Model, | |
Nonstationary, integrated ARMA (ARIMA) For details,
see ARIMA Model, | |
Seasonal (Multiplicative) ARIMA (SARIMA) For details,
see Multiplicative ARIMA Model, | |
ARIMA including exogenous predictors (ARIMAX) For details,
see ARIMA Model Including Exogenous Covariates, | |
Seasonal ARIMAX | |
Conditional variance: GARCH Models section | Generalized autoregressive conditional heteroscedastic
(GARCH) For details,
see GARCH Model, |
Exponential GARCH (EGARCH) For details,
see EGARCH Model, | |
Glosten, Jagannathan, and Runkle (GJR) | |
Multiple linear regression: Regression Models section | Multiple linear regression For details,
see Time Series Regression I: Linear Models, |
Regression model with ARMA errors For details,
see Regression Models with Time Series Errors, |
For conditional mean model estimation, SARIMA and SARIMAX are the most flexible models. You can create any conditional mean model that excludes exogenous predictors by clicking SARIMA, or you can create any conditional mean model that includes at least one exogenous predictor by clicking SARIMAX.
After you select a model, the app displays the Type
Model
Parameters dialog box, where Type
is the model
type. This figure shows the SARIMAX Model Parameters dialog box.
Adjustable parameters in the Type
Model
Parameters window depend on Type
. In
general, adjustable parameters include:
A model constant (offset or intercept) and linear regression coefficients corresponding to predictor variables (see Adjusting Model Constant and Regression Component Parameters)
Time series component parameters, which include seasonal and nonseasonal lags and degrees of integration (see Adjusting Time Series Component Parameters)
The innovation distribution (see Adjusting Innovation Distribution Parameters)
As you adjust parameter values, the equation in the Model Equation section changes to match your specifications. Adjustable parameters correspond to input and name-value pair arguments described in corresponding model creation reference pages. For details, see the function reference page for a specific model. Regardless of the model you choose, all coefficients in the model are unknown and estimable, including the t-distribution degrees of freedom parameter (when you specify a t innovation distribution).
Econometric Modeler does not support:
Optimization option adjustments for estimation.
Composite conditional mean and variance models. For details, see Specify Conditional Mean and Variance Models.
Applying equality constraints to specified parameters during estimation (except for holding parameters fixed at zero during estimation).
To adjust optimization options, estimate composite conditional mean and variance models, or apply equality constraints, use the MATLAB command line.
To include a model constant (offset or intercept) term, select the
Include Constant Term or Include Offset
Term check box. To remove the model constant (that is, constrain
it to zero during estimation), clear the check box. The location and type of the
check box in the Type
Model
Parameters dialog box depends on the model type. By default,
Econometric Modeler includes a model constant in all model types except
conditional variance models.
To select predictors for the regression component, in the Predictors list, select the check box in the Include? column corresponding to the predictors you want to include in the model. By default, the app does not include a regression component in any model type. If you select ARIMAX, SARIMAX, MLR, or RegARMA, then you must choose at least one predictor.
Consider a linear regression model of GDP onto CPI and the unemployment rate. To specify the regression:
Import the DataTable
variable in the
Data_USEconModel
data set into the
Econometric Modeler (see Import Time Series Variables).
In the Data Browser, select the response
variable GDP
.
On the Econometric Modeler tab, in the Models section, click the arrow to display the models gallery.
In the models gallery, in the Regression Models section, click MLR.
In the MLR Model Parameters dialog box, in the Include? column, select the CPIAUCSL and UNRATE check boxes.
Click the Estimate button.
In general, time series component parameters contain lags to include in the seasonal and nonseasonal lag operator polynomials, and seasonal and nonseasonal degrees of integration.
For conditional mean models, you can specify seasonal and nonseasonal autoregressive lags, and seasonal and nonseasonal moving average lags. You can also adjust seasonal and nonseasonal degrees of integration.
For conditional variance models, you can specify ARCH and GARCH lags. EGARCH and GJR models also support leverage lags.
For regression models with ARMA errors, you can specify nonseasonal autoregressive and moving average lags. For models containing seasonal lags or degrees of seasonal or nonseasonal integration, use the command line instead.
Econometric Modeler supports two options to adjust the
parameters. The adjustment options are on separate tabs of the
Type
Model Parameters
dialog box: the Lag Order and Lag
Vector tabs. On the Lag Order tab, you can
specify the orders of lag operator polynomials. This
feature enables you to include all lags efficiently, from 1 through the
specified order, in a lag operator polynomial. On the Lag
Vector tab, you can specify the individual
lags that comprise a lag operator polynomial. This
feature is well suited for creating flexible models. For more details, see Specifying Lag Operator Polynomials Interactively.
For all models, you can specify that the distribution of the innovations is Gaussian. For all models, except multiple linear regression models, you can specify the Student's t instead to address leptokurtic innovation distributions (for more details, see Maximum Likelihood Estimation for Conditional Mean Models, Maximum Likelihood Estimation for Conditional Variance Models, or Maximum Likelihood Estimation of regARIMA Models). If you specify the t distribution, then Econometric Modeler estimates its degrees of freedom parameter using maximum likelihood.
By default, Econometric Modeler uses the Gaussian distribution for the
innovations. To change the innovation distribution, in the
Type
Model Parameters
dialog box, from the Innovation Distribution button, select
a distribution in the list.
Econometric Modeler treats all parameters in the model as unknown and
estimable. After you specify a model, fit it to the data by clicking
Estimate in the
Type
Model Parameters
dialog box.
Econometric Modeler requires initial values for estimable
parameters and presample observations to initialize the model for
estimation. Econometric Modeler always chooses the default initial
and presample values as described in the estimate
reference page of the model you want to estimate.
If Econometric Modeler issues an error during estimation, then:
The specified model poorly describes the data. Adjust model parameters, then estimate the new model.
At the command line, adjust optimization options and estimate the model. For details, see Optimization Settings for Conditional Mean Model Estimation, Optimization Settings for Conditional Variance Model Estimation, or Optimization Settings for regARIMA Model Estimation.
After you estimate a model:
A new variable that describes the estimated model appears in the
Models section of the Data
Browser with the name
Type_response
.
Type
is the model type and
response
is the response variable to
which Econometric Modeler fit the model, for example,
ARIMA_FEDFUNDS
.
You operate on an estimated model in the Data
Browser by right-clicking it. In addition to the
options available for time series variables (see Import Time Series Variables),
the context menu includes the Modify
option, which enables you to modify and re-estimate a model. For
example, right-click a model and select
Modify
. Then, in the
Type
Model
Parameters dialog box, adjust parameters and click
Estimate.
The Model
Summary(Type_response
)
document summarizing the estimation results appears in the right
pane. Results shown depend on the model type. For conditional mean
and regression models, results include:
Model Fit — A time series plot of the response series and the fitted values
Parameters — An estimation summary table containing parameter estimates, standard errors, and t statistics and p-values for testing the null hypothesis that the corresponding parameter is 0
Residual Plot — A time series plot of the residuals
Goodness of Fit — Akaike information criterion (AIC) and Bayesian information criterion (BIC) model fit statistics
For conditional variance models, the results also include an estimation summary table and goodness-of-fit statistics, but Econometric Modeler plots:
Conditional Variances — A time series plot of the inferred conditional variances
Standardized Residuals — A time series plot of the standardized residuals , where c is the estimated offset
You can interact with individual plots by pausing on one and selecting an interaction (see Visualizing Time Series Data). You can also interact with the summary by right-clicking the document. Options include:
Export — Place plot in a separate figure window.
Show Model — Display the summary of another estimated model by pointing to Show Model, then selecting a model in the list.
Show Recessions — Plot recession bands in time series plots.
Consider a SARIMA(0,1,1)×(0,1,1)12 for the monthly
international airline passenger numbers from 1949 to 1960 in the
Data_Airline
data set. To estimate this model using the
Econometric Modeler:
Import the DataTable
variable in the
Data_Airline
data set into Econometric
Modeler (see Import Time Series Variables).
On the Econometric Modeler tab, in the Models section, click the arrow > SARIMA.
In the SARIMA Model Parameters dialog box, on the Lag Order tab:
Nonseasonal section
Set Degrees of Integration to
1
.
Set Moving Average Order to
1
.
Clear the Include Constant Term check box.
Seasonal section
Set Period to 12
to
indicate monthly data.
Set Moving Average Order to
1
.
Select the Include Seasonal Difference check box.
Click Estimate.
A variable named SARIMA_PSSG
appears in
the Models section of the Data
Browser, and this estimation summary appears in the new
Model Summary(SARIMA_PSSG) document.
After you estimate a model, a good practice is to determine the adequacy of the fitted model (see Goodness of Fit). Econometric Modeler is well suited for visually assessing the in-sample fit (for all models except conditional variance models) and performing residual diagnostics.
Residual diagnostics include evaluating the model assumptions and investigating whether you must respecify the model to address other properties of the data. Model assumptions to assess include checking whether the residuals are centered on zero, normally distributed, homoscedastic, and serially uncorrelated. If the residuals do not demonstrate all these properties, then you must determine the severity of the departure, whether to transform the data, and whether to specify a different model. For more details on residual diagnostics, see Time Series Regression VI: Residual Diagnostics and Residual Diagnostics.
To perform goodness-of-fit checks using Econometric Modeler, in the Models section of the Data Browser, select an estimated model. Then complete the following steps:
To visually assess the in-sample fit for all models (except conditional variance models), inspect the Model Fit plot in the Model Summary document.
To visually assess whether the residuals are centered on zero, autocorrelated, and heteroscedastic, inspect the Residual Plot in the Model Summary document.
On the Econometric Modeler tab, in the Diagnostics section, click Residual Diagnostics. The diagnostics gallery provides these residual plots and tests.
Method | Diagnostic |
---|---|
Residual histogram | Visually assess normality |
Residual quantile-quantile plot | Visually assess normality and skewness |
ACF | Visually assess whether residuals are autocorrelated |
Ljung-Box Q-test | Test residuals for significant autocorrelation |
ACF of squared residuals | Visually assess whether residuals have conditional heteroscedasticity |
Engle's ARCH test | Test residuals for conditional heteroscedasticity (significant ARCH effects) |
Alternatively, to plot a histogram, quantile-quantile plot, or ACF of the residuals of an estimated model:
Select a model in the Data Browser.
Click the Plots tab.
In the Plots section, click the arrow and then click one of the plots in the Model Plots section of the gallery.
Another important goodness-of-fit check is predictive-performance assessment. To assess the predictive performance of several models:
Fit a set of models to the data using Econometric Modeler.
Perform residual diagnostics on all models.
Choose a subset of models with desirable residual properties and minimal fit statistics (see Finding Model with Best In-Sample Fit).
Export the chosen models to the MATLAB Workspace (see Export Session Results).
Perform a predictive performance assessment at the command line (see Check Predictive Performance).
For an example, see Compare Predictive Performance After Creating Models Using Econometric Modeler App.
Consider performing goodness-of-fit checks on the estimated SARIMA(0,1,1)×(0,1,1)12 model for the airline counts data in Estimating a Model.
In the right pane, on the Model Summary(SARIMA_PSSG) document:
Model Fit suggests that the model fits to the data fairly well.
Residual Plot suggests that the residuals have a mean of zero. However, the residuals appear heteroscedastic and serially correlated.
On the Econometric Modeler tab, in the Diagnostics section, click Residual Diagnostics. In the diagnostics gallery:
Click Residual Q-Q Plot. The right pane display a figure window named QQPlot(SARIMA_PSSG) containing a quantile-quantile plot of the residuals.
The plot suggests that the residuals are approximately normal, but with slightly heavier tails.
Click Autocorrelation Function. In the toolstrip, the ACF tab appears and contains plot options. The right pane displays a figure window named ACF(SARIMA_PSSG) containing the ACF of the residuals.
Because almost all the sample autocorrelation values are below the confidence bounds, the residuals are likely not serially correlated.
Click Engle's ARCH Test. On the ARCH tab, in the Tests section, click Run Test to run the test using default options. The right pane displays the ARCH(SARIMA_PSSG) document, which shows the test results in the Results table.
The results suggest rejection of the null hypothesis that the residuals exhibit no ARCH effects at a 5% level of significance. You can try removing heteroscedasticity by applying the log transformation to the series.
Econometric Modeler enables you to fit multiple related models to a data set efficiently. After you estimate a model, you can estimate other models by iterating the methods in Perform Exploratory Data Analysis, Fitting Models to Data, and Conducting Goodness-of-Fit Checks. After each iteration, a new model variable appears in the Models section of the Data Browser.
For models in the same parametric family that you fit to the same response series, you can determine the model with the best parsimonious, in-sample fit among the estimated models by comparing their fit statistics. From a subset of candidate models, to determine the model of best fit using Econometric Modeler:
In the Models section of the Data
Browser, double-click an estimated model. In the right
pane, estimation results of the model appear in the Model
Summary(Model
) document,
where Model
is the name of the selected
model.
On the Model
Summary(Model
) document, in
the Goodness of Fit table, choose a fit statistic
(AIC or BIC) and record its value.
Iterate the previous steps for all candidate models.
Choose the model that yields the minimal fit statistic.
For more details on goodness-of-fit statistics, see Information Criteria.
Consider finding the best-fitting SARIMA model, with a period of 12, for the log
of the airline passenger counts in the Data_Airline
data set. Fit
a subset of SARIMA models, considering all combinations of models that include up to
two seasonal and nonseasonal MA lags.
Import the DataTable
variable in the
Data_Airline
data set into Econometric
Modeler (see Import Time Series Variables).
Apply the log transformation to PSSG
(see
Transforming Time Series).
Fit a
SARIMA(0,1,q)×(0,1,q12)12
to PSSGLog
, where all unknown orders are 0
(see Estimating a Model).
In the right pane, on the Model Summary(SARIMA_PSSGLog) document, in the Goodness of Fit table, record the AIC value.
In the Data Browser, select
PSSGLog
.
Iterate steps 4 and 5, but adjust q and
q12 to cover the nine
permutations of q ∈
{0
,1
,2
}
and q12 ∈
{0
,1
,2
}.
Econometric Modeler distinguishes subsequent models of the same type by
appending consecutive digits to the end of the variable name.
The resulting AIC values are in this table.
Model | Variable Name | AIC |
---|---|---|
SARIMA(0,1,0)×(0,1,0)12 | SARIMA_PSSGLog1 | -491.8042 |
SARIMA(0,1,0)×(0,1,1)12 | SARIMA_PSSGLog2 | -530.5327 |
SARIMA(0,1,0)×(0,1,2)12 | SARIMA_PSSGLog3 | -528.5330 |
SARIMA(0,1,1)×(0,1,0)12 | SARIMA_PSSGLog4 | -508.6853 |
SARIMA(0,1,1)×(0,1,1)12 | SARIMA_PSSGLog5 | -546.3970 |
SARIMA(0,1,1)×(0,1,2)12 | SARIMA_PSSGLog6 | -544.6444 |
SARIMA(0,1,2)×(0,1,0)12 | SARIMA_PSSGLog7 | -506.8027 |
SARIMA(0,1,2)×(0,1,1)12 | SARIMA_PSSGLog8 | -544.4789 |
SARIMA(0,1,2)×(0,1,2)12 | SARIMA_PSSGLog9 | -542.7171 |
Because it yields the minimal AIC, the SARIMA(0,1,1)×(0,1,1)12 model is the model with the best parsimonious, in-sample fit.
Econometric Modeler offers several options for you to share your session results. The option you choose depends on your analysis goals.
The options for sharing your results are in the Export section of the Econometric Modeler tab. This table describes the available options.
Option | Description |
---|---|
Export Variables | Export time series and model variables to the MATLAB Workspace. Choose this option to perform further analysis at the MATLAB command line. For example, you can generate forecasts from an estimated model or check the predictive performance of several models. |
Generate Function | Generate a MATLAB function that creates a model. Choose this option to:
|
Generate Report | Generate a report that summarizes the session. Choose this option when you achieve your analysis goals in Econometric Modeler, and you want to share a summary of the results. |
To export time series and estimated model variables from the Data Browser to the MATLAB Workspace:
On the Econometric Modeler tab, in the
Export section, click or Export
> Export Variables.
In the Export Variables dialog box, all time
series variables in the Data Browser appear in
the left pane and all model variables appear in the right pane.
Choose time series and model variables to export by selecting the
corresponding check boxes in the Select column.
The app selects the check box of all time series or model variables
that are selected in the Data Browser. Clear
any check boxes for variables to you do not want to export. For
example, this figure shows how to select the
PSSGLog
time series and the
SARIMA_PSSGLog
SARIMA model.
Click Export.
The selected variables appear in the MATLAB Workspace. Time series variables are double-precision column
vectors. Estimated models are objects of type depending on the model (for
example, an exported ARIMA model is an arima
object).
Alternatively, you can export variables from the Data Browser by selecting at least one variable, right-clicking a selected variable, and selecting Export.
To export a MATLAB function that creates an estimated model:
In the Data Browser, select an estimated model.
On the Econometric Modeler tab, in the Export section, click Export > Generate Function.
In the MATLAB Editor, an untitled, unsaved function appears containing the code that estimates the model.
By default, the function name is
modelTimeSeries
.
The function accepts the originally imported data set as input.
Before the function estimates the model, it extracts the variables from the input data set used in estimation, and applies the same transformations to those variables that you applied in the Econometric Modeler.
The function returns the selected estimated model.
Consider generating a function that returns SARIMA_PSSGLog
,
the SARIMA(0,1,1)×(0,1,1)12 model fit to the log of the
airline passenger data (see Estimating a Model). This figure
shows the generated function.
Econometric Modeler can produce a report describing your activities on selected time series and model variables. The app organizes the report into chapters corresponding to selected time series and model variables. Chapters describe session activities that you perform on the corresponding variable.
Chapters on time series variables describe transformations, plots, and tests that you perform on the selected variable in the session. Estimated model chapters contain an estimation summary, that is, elements of the Model Summary document (see Estimating a Model), and residual diagnostics plots and tests.
You can export the report as one of the following document types:
Hypertext Markup Language (HTML)
Microsoft® Word XML Format Document (DOCX)
Portable Document Format (PDF)
To export a report:
On the Econometric Modeler tab, in the Export section, click Export > Generate Report.
In the Select items to include in report dialog box, all time series variables in the Data Browser appear in the left pane and all model variables appear in the right pane. Choose variables to include the report by selecting their check boxes in the Select column.
Select a document type by clicking Report Format and selecting the format you want.
Click OK.
In the Select File to Write window:
Browse to the folder in which you want to save the report.
In the File name box, type a name for the report.
Click Save.
Consider generating an HTML report for the analysis of the airline passenger data (see Conducting Goodness-of-Fit Checks). This figure shows how to select all variables and the HTML format.
This figure shows a sample of the generated report.
[1] Box, G. E. P., G. M. Jenkins, and G. C. Reinsel. Time Series Analysis: Forecasting and Control. 3rd ed. Englewood Cliffs, NJ: Prentice Hall, 1994.