What's New in TSP 4.4

H O M E / I N F O / P R O D U C T S / O R D E R / S U P P O R T

TSP / What is TSP? / Features / Requirements

New in TSP 4.4 4.3 4.2b 4.2a Older versions

TSP 4.4 was released in May 1997. If you are a user of a version of TSP before 4.3, or bought 4.3 before the summer of 1996, you should also look at the what’s new in 4.3. Many features were added to 4.3 in the year and a half since it was first released.

· Regression output has been improved by adding more default tests and P-values. The new diagnostics are a general purpose heteroskedasticity test LMHET, the Shapiro-Wilk normality test, Ramsey's RESET test, and a "deviance" R-squared for Probit and Logit.

· ML PROC: A powerful enhancement to the maximum likelihood procedure (ML) that allows you to evaluate the log likelihood function in a TSP procedure (PROC) rather than as a single equation.

· PANEL procedure enhanced to handle panels with missing values, lags, and leads. A new PANEL option has been added to FREQ in order to facilitate future language development in the handling of PANEL data. The Durbin-Watson and LMHET diagnostics have been added to PANEL.

· The instrumental variable procedures (INST, 3SLS, GMM) have been revised to speed up computation in the case where there are large numbers of instruments and observations.

· Long variable names (up to 64 characters) are now allowed.

· A new VERT option has been added to MMAKE that stacks the matrices or series vertically.

· FORM (VARPREF=<prefix>) can be used to give the coefficients in the created equation names formed by concatenating <prefix> with the variable name.

· Bundled visual interface utility, TSP through the Looking Glass (now replaced by Oxmetrics interface). Windows 95 and NT versions include TSP through the Looking Glass 1.10 while Windows 3.1 includes TLG 1.01.

· New integrated installer installs both TSP, TSP through the Looking Glass with an option to install the TSP examples library as well.

New in TSP 4.3

TSP 4.3 was released in 1995.The following summarizes the improvements to TSP from version 4.2b to 4.3. Highlights include full adjustable memory, a new COINT command and many enhancements to existing commands.

Adjustable Memory: On all machines, TSP 4.3 adjusts working space to fit available memory (RAM or virtual memory). There is no longer a fixed limit on the number of observations in a series (now only limited by overall working space). Large series and matr ices are processed more quickly.
On-screen editing: in interactive mode, use the arrow keys to repeat and/or edit commands. (DOS/Windows only)
Batch loop: a feature that lets you iterate between running a batch TSP job, looking at the output, editing the input, and running the job again. This is useful for developing complex batch programs. (for task-switching operating systems)
Subscripts: Matrices may now be doubly subscripted with numbers or variable names(up to two characters in length).
login.tsp: you can create this special file to contain all the TSP commands you normally use when the program starts up (setting OPTIONS, suppressing some printout, specifying the databank, and so forth.)

CLEAR restarts interactive sessions.
COINT performs unit root and cointegration tests on a set of series.
DBDEL databank varlist deletes the variables in varlist from the databank.

ANALYZ works after the procedures that have a linear regression model (AR1, OLSQ, INST, LIML, and PROBIT) with series names used in place of coefficient names to refer to their coefficients (added to 4.2b in 12/92). Unnormalized equations are labeled by the equation names (rather than not labeled at all).
ARCH (HEXP=value) estimates an ARCH model with h(t) raised to an arbitrary power given by the HEXP option. For HEXP=0, the log(h) is used. The default is still HEXP=.5 (square root of the variance). The UNCOND option holds the presample h(t) fixed at the current estimate of the unconditional residual variance.
BJEST has been integrated into TSP's nonlinear estimation system in order to improve convergence behavior and produce standard nonlinear output.
COVA (PAIRWISE) computes a pairwise covariance matrix, that is, a matrix where each cell is computed over all the available data for the two variables involved.
CORR (PAIRWISE) computes a pairwise correlation matrix, that is, a matrix where each cell is computed over all the available data for the two variables involved.
CUSUM, CUSUMSQ -- These new regression output options produce plots, test statistics, and P-values for tests of structural stability (already implemented in some releases of 4.2b).
Dates Y:P -- P can now be bigger than 100, but must be smaller than 32,768 (see FREQ number). Y and P must be numbers. See the new BASEYEAR option for dates between 0 A.D. and 200 A.D.
DOT has several new options and the limit on the number of DOT elements has been increased to 2000. Strings such as '01' are allowed as DOT names. DOT(INDEX=varname) creates a variable varname that is incremented over the DOT loop (this elimina tes the need to create such an index manually). DOT(VALUE=varname) creates a variable that is filled with the actual values of the DOT list (if they are numeric). DOT (CHAR=char) defines a character that will be used together with . to index that DOT list . This is useful if you want to use a single dot on variables in an inner loop, or if you want to reverse the order of the DOT expansion.
FREQ W or number -- W is weekly and assumes exactly 52 weeks per year. W converts to quarterly (Q) but not to monthly (M) because TSP does not yet have a true calendar built in. FREQ number where number is a number smaller then 32,768 defines a frequency of number periods per year.
GMM (MASK=) counts the 1's in the mask as the number of orthogonality conditions (because the OC's corresponding to zeroes in the mask are dropped in estimation). SMPL gaps are properly handled when NMA>0 (GMM skips over the gaps).
MOMENT (PAIRWISE) computes a pairwise moment matrix, that is, a matrix where each cell is computed over all the available data for the two variables involved.
MSD (BYVAR) computes univariate statistics for a list of variables using all the data available for each variable (even if there is missing data for other variables).
RANDOM (DRAW=matrix name) produces multivariate random numbers with an empirical distribution by drawing from the rows of a matrix. The random vector is #columns long, The matrix may have a number of rows not equal to the number of observations in the current SMPL.
SIML has been integrated into TSP's nonlinear system in order to improve convergence behavior and produce standard nonlinear output. It is much faster for linear models with several observations.
SOLVE has been integrated into TSP's nonlinear system in order to improve convergence behavior and produce standard nonlinear output.

AR1 handles series with missing data (already in some 4.2b versions).
BJIDENT stores the autocorrelations, partial autocorrelations, and inverse autocorrelations under @AC, @PAC, and @IAC respectively(already in some 4.2b versions).
CDF has a new TSQ option when using a squared trend in the Dickey-Fuller or Engle-Granger regressions.
CNORMI() returns less extreme answers for the arguments 0 and 1 (the true answer would be +/- infinity).
COMPRESS (NOPRINT) -- when the user requests compression of TSP's data storage, the default is for no message to be printed.
CONVERT now goes from a lower frequency to a higher (e.g., annual to quarterly), with the options interpreted appropriately. (AVERAGE uses the average over the period, SUM causes division by the ratio of the frequencies to occur, and so forth). (already in some 4.2b versions).
DBCOPY -- The .TSP file produced is more concise, and faster to upload/download.
DIVIND -- when the quantity is zero, DIVIND now drops that good temporarily from the indices.
DUMMY can take a user-specified list of names as the variables to be created (rather than just making up the names).
FETCH now reads DOCumentation from microTSP (.DB) databanks. N/A's (missing values) are handled properly.
GENR -- dynamic GENRS can now be done in reverse time (if leads are on the righthandside instead of lags).
INPUT works in batch mode.
KALMAN (XFIXED=X) has improved labels when there is more than one state variable. When the prior variance based on the first few observations is singular, it is reestimated with an additional observation.
LIST (FIRST=number,LAST=number,PREFIX=name) listname creates a LIST of the form name1, name2,and so forth (using the obvious set of numbers), and stores it under listname.
LOGIT checks for univariate complete and quasi-complete separation of the data; in this instance the coefficients will not be identified and maximization behavior will be poor.
MATRIX -- EIGVAL works on non-symmetric square matrices; in this case the imaginary part of the eigenvalues are stored in @EIGVALI. EIGVEC causes @EIGVAL to be stored along with @EIGVEC. INV stores the determinant of the matrix before inversio n as well as the inverse.
MFORM (DIAG) works with series or general matrices.
MMAKE works with column vectors as well as series.
MODEL -- the endogenous variable list is now optional if they all appear on the lefthandside of the equations. The PRINT and SILENT options can be used with MODEL to control output.
MODEL(FILE=filename) specifies an external file on which to write the simultaneous model after ordering, for input to CAUSOR, an external program that analyzes the structure of large econometric models.
NONLINEAR options -- The default TOL has been changed from .01 to .001 and STEP=GOLDEN has been improved.
OLSQ -- when there are missing values, @RES and @FIT are stored correctly. This affects all procedures that allow missing values (AR1, INST, LIML, and PROBIT). (already in some 4.2b versions).
OPTIONS BASEYEAR=1900/2000/0 for two digit years allows the dates 0-200 A.D. to be used, and is designed to move TSP into the next millennium!
PANEL -- the R-squares for WITHIN and VARCOMP estimation now reflect the original dependent variable, rather than the variable with individual means removed. Output for F and Hausman tests cleaned up (reduced to 80-column width). The residuals from all models are now stored under the names @REST, @RESB, @RESW, and @RESV.
PROBIT checks for univariate complete and quasi-complete separation of the data; in this instance the coefficients will not be identified and maximization behavior will be poor.
PROC now prints the original (calling program) variable name if the variable is missing (instead of the dummy name inside the PROC). Also, for any error, a traceback consisting of the list of nested PROCs being called and the line numbers from which they were called is printed.
READ will probably support the Excel 5.0 format for spreadsheet data at the time of release. OPTIONS CHARID allows character ID variables to be read.
SHOW prints the internal limits for the program (added in 4.2b, not needed as much in 4.3 because of increased storage flexibility).
STORE now writes DOCumentation to microTSP databanks. N/A's (missing values) are handled properly.
TOBIT uses a globally concave parametrization that greatly improves the speed of convergence in some cases. (already in some 4.2b versions).

New in TSP 4.2b

TSP 4.2b was released in 1993, with the following enhancements.

386/486 TSP (for DOS) prints graphics.
Paging of interactive regression output (no longer runs off the screen).
ANALYZ can follow any estimation procedure (including OLSQ, etc.). This makes it easy to test hypotheses, even for linear models.
A CNORMI function which gives the inverse of the cumulative normal distribution function has been added. This is useful for truncation models with non-normal distributions.
KALMAN uses orthonormalized variables in computation (improves numerical accuracy).

MAT stores complex eigenvalues/ eigenvectors for general square matrices for further computation by user.
MODEL is substantially revised with improved output.
In PROBIT, the computed Mills ratios are negative when Y=0, making them the correct expected value of the latent residual conditional on Y=0 and the Xs. Previously, they were computed conditional on Y=1. In both PROBIT and LOGIT, variable-by-variable checks for complete or quasi-complete separation are performed and a warning printed if the check fails.

LAD depvar list-of-indepvars; estimates a LAD (least absolute deviations) regression, also known as L1 or Laplace regression.

BJIDENT stores the autocorrelations, partial autocorrelations, and inverse autocorrelations.
READ reads free format files of arbitrary width and does it 50% faster. It can also read Lotus (V.1-3 and Japanese) and Excel (V.2-4) files, even on Unix systems.

GMM stores the vector of orthogonality conditions. The NMA= option handles gaps in the SMPL correctly. The MASK= option excludes instruments from some equations.

AR1 handles missing values.

The last (innermost) of a series of DOT statements is used when there is a single dot (.) in a variable referenced in the innermost loop. DOTs with strings like '01' are allowed also.

KEEP ALL; statement forces all variables to be stored in any open databanks.

RANDOM has many new distributions: CAUCHY (no parameters), EXPON (for exponential with option LAMBDA=a, where the expectation is 1/a), LAPLACE (double exponential with the option LAMBDA=a), and T (Student's t distribution with option DF=n). RANDOM ta kes multiple arguments for any distribution; the random variables generated will be independent except for the multivariate normal distribution with VCOV= a non-diagonal matrix. The
DRAW= option for the empirical distribution function now samples only from the non-missing values of the supplied series.

PANEL computes a Hausman test for correlated effects (within vs. varcomp), and stores the residuals from the different estimations. It also stores the fixed effects, and the chi-squared, p-value, and degrees of freedom for the Hausman test.

SHOW ; with no arguments lists internal array dimensions in TSP.

New in TSP 4.2a

TSP 4.2a was released in 1990, with the following enhancements.

· ARCH estimates the GARCH-M class of models, which allows the variance of the disturbance in a regression equation to evolve as an ARMA process whose mean is a function of a set of variables. Special cases of this model are ordinary ARCH and weighted least squares.

· GMM (Generalized Method of Moments) uses instrumental variables or minimum distance estimators popularized by Hansen, Singleton, and others whereby the weighting matrix is (optionally) a heteroskedastic- and autocorrelation- consistent estimate. GMM can be used to estimate parameters of a set of orthogonality conditions implied by a rational expectations model. GMM is more general than that, however, and has applications outside of time series data.

· KALMAN estimates linear regressions using the Kalman filter, as well as more elaborate state space models (time- varying parameters, adaptive regression, stochastically convergent parameters).

· VAR estimates vector autogressions (including exogenous variables), calculates impulse response functions, and forcasts error variance decompositions.

· CDF calculates tail probabilities (p-values) or critical values for the F, t, chi-squared, normal, and bivariate normal. CDF can also calculate the significance levels for unit root testing (using the Dickey-Fuller methodology) and cointegration tests (using the Engle-Granger methodology).

· DOC creates and stores documentation for series or other variables in a TSP databank.

· HELP gives basic command syntax and is now available on all computers.

· MAT calculates a matrix algebra equation, using natural notation such as B = (X'X)"X'Y. Operators available include addition, subtraction, multiplication, inversion, transpose, Kronecker product, and Hadamard product. MAT provides many matrix functions, such as CHOL, IDENT, EIGVAL, EIGVEC, VEC, VECH, DIAG, TR, DET, LOGDET, MIN, MAX, SUM, NROW, NCOL, and RANK.

· PANEL estimates linear regression models for balanced or unbalanced time series-cross section data. PANEL can handle pooled, between, within (fixed effects), and variance components (random effects) models. Grouped means and tests for equality are also provided.

· REGOPT controls the calculation and display of regression diagnostics. Since a large number of new diagnostics have been added, REGOPT enables the user to customize his program to eliminate irrelevant diagnostics. P-values and stars can also be automatically computed for most statistics.

· SORT sorts one or more series, and is useful for reordering panel data among other things.

· ANALYZ - Names of calculated coefficients, their estimated values, standard errors, and variance-covariance are stored under the names @RNMSA, @COEFA, @SESA, and @VCOVA so that several ANALYZs may be performed following a single estimation (without destroying the input data required by ANALYZ).

· CAPITL - END option causes end-of-period capital stock to be computed rather than beginning-of-period.

· FORM - Linear equations (FRMLs) for use in LSQ and other procedures can now be formed automatically after OLSQ, INST, or AR1. The NAR option allows easy estimation of AR(p) models with LSQ.

· LOGIT - The @DPDX and @DPDZ matrices are now printed by default and labelled with the variable names and corresponding values of the dependent variable.

· MFORM has the addition of a BLOCK option for forming block diagonal matrices.

· MSD's ALL option computes skewness and kurtosis in addition to other statistics.

· Output with PDL variables has been rewritten to report unscrambled coefficients only. Thus, there is no longer any difficulty in using PDL equations for forecasting, since coefficients of unscrambled variables are stored in the equation. A new option for Shiller distributed lags has been added to PDL variables (in OLSQ only). Shiller lags relax the polynomial constraint slightly.

· OPTIONS - LIMWARN= (suppresses printing of warning messages), DOUBLE (specifies that all series should be stored in double precision), DISPLAY= (for PC Graphics version), and SECONDS= (tells how often to update screen message in long procedures).

· PRIN - Addition of MAXIT option to limit or extend the number of iterations used to find the principal components.

· READ/WRITE - Addition of BYSER option for reading data matrix in transposed form (by series instead of observation), FORMAT=RB8 for double precision binary format, and FORMAT=LOTUS for PC version to read Lotus or other spreadsheet files (already released in PC TSP 4.1C version).

· All scalars (CONSTs or PARAMs) are now stored as double precision for improved accuracy.

· Previous limits on matrix size (&lt16384; elements) or time series length (&lt32768;) have been removed, as have the limits on the number of instruments or endogenous variables. There is still a limit on the number of observations in TSP (4.2, but not 4.3), but now the limit can be set &gt32768; when making a custom version.

· Improved list processing, with expressions expressions like A0-A2 (starting with zeroes), X(-1)-X(-1) (beginning and ending the same), L1A-L10A, and M1A1-M10A1 (internal index), being allowed.

· Improved error-handling and reporting of numerical errors in nonlinear procedures. In SOLVE, the equation number is reported when there is a numerical error.

· Equations non-differentiable at a finite number of points (such as those involving logical expressions) are now allowed for estimation or simulation. The derivative is defined to be zero at those points.

· Improved MISSING VALUE report names the series with the missing values and how many are in each series.

If you have any questions or comments about TSP please send an email to info@tspintl.com.

Comments or questions about this website should also be sent to the info@tspintl.com. Lost? Please consult the site map.