<Previous version: PcGive 14>

PcGive 15: enhancements and solved problems

Improved in PcGive 15.2

• Graphic analysis, cointegration graphics: options shifted by one (relations shows actual and fitted, actual and fitted show components, relations not shown).
• Level forecast was omitted when regressors with long lags were used.
• Index test low degrees-of-freedom cutoff was a bit too stringent.
• Printing (centered) R^2 also when there is no intercept.
• Autometrics 2.1, minor changes for more variables than observations:
  • Changed handling when almost out of data for blocking: can continue if there are at least 4 observations left.
  • Added option to switch off backtesting in block search for N>T (in advanced Autometrics settings). This gives a gauge that is closer to the nominal significance level.
  • Created block method 2: reduced. This switches backtesting off in the final Autometrics run after the block search. This will reduce the number of insignificant variables retained because of backtesting, so lead to a somewhat tighter gauge.
    This option can be found under Advanced settings/Block identification when there are too many parameters.

    The default method is 1: standard, which is unchanged. The 0: quick method now also has backtesting switched off in the final selection.

    The reduced method can be set in Ox code using AutometricsSet("block_method", "reduced");

Fixed in PcGive 15.1

• TIS: now also dropping penultimate term, because the difference between that and a trend is constant (and it is assumed that usually the Constant is U)
• Better error message handling in PcGive::InvertSymGen
• Making a lagged dependent variable a U implied that it was not treated as a lagged dependent variable in lag analysis and roots of polynomials.
• Cointegration with lags>1 put coefficients in the wrong place: forecasting and part of dynamic analysis wrong.
• FIML+Autometrics from batch: gave runtime error.
• Test/Test (multivariate): gave runtime error.
• Had a runtime error in heterotest: undercounted no of regressors.
• Recursive estimation from batch code is not processed properly.

Fixed in PcGive 15.01

• Option to store forecast Standard errors and robust forecasts in test dialog
• Estimate empty model, withholding forecasts: gave runtime error

Fixed in PcGive 15

• Fixed index error with level forecasts
• RESET test: check for number of observations (previously runtime error if too small). Now requiring at least 5 df.
• Logit predictions: observations with state outside the choice set cause a crash.
• Batch code to map to I(0) for UKM1 writes Z=,DOUT,DOIL;
• PcGive Autometrics output says: outlier detection no Changed to large residuals: no
• For X12arima to work, an X12arima licence was required, instead it should work with a PcGive licence
• Reordering for cointegration put the Y variables in the database order. Interactively, this is always as specified, but may be different for the generated Ox code that uses the actual data set. That can be a problem for the formulated restrictions on the cointegrating space. Now always using specified order.
• Hedgehog function returns fully transformed variables (previously: still had to scale and take anti-log or anti-logit).
• Multivariate robust Hedgehog plots had variables scrambled.
• Hedgehog plots use a different color in the forecast period
• Hedgehog Levels forecasts beyond estimation sample: not integrated
• Levels forecasts with gap: created cGap extra forecasts Also used wrong levels if cGap > 0.
• Recursive hedgehog would omit early part when there are dummies
• Fixed Ox issue with find, affecting forecasting
• Arfima package distributed with OxMetrics 7.0 has out of date oxo file (prerelease - not compatible with Ox 7).

New and improved in PcGive 15

• Markov Switching: removed recursive and forecasts from estimation dlg.
• Growing database for forecasting: if the last observation is zero, extending by zero: useful after saturation.
• Added graph of intercept adjustments (Constant excluded, but Trend included)
• Improved Autometrics for large models: better handling of search for contrasting terminals and using some blocking in the lag presearch when the lag length is very large (beyond 12). BigT10 data gives an example where root-search gets in the way if there is no lag presearch: there are just too many at the root.

  Without lag presearch:	time 13:39:54.59	Autometrics version 1.6
  with:lag presearch	time 7:39.07	Autometrics version 1.6
  new without presearch:	time 20.54	Autometrics version 2.0
  new with presearch:	time 1:13.65	Autometrics version 2.0

• Include options in the Progress menu to "select all models" and "deselect all models".
• More compact output for T>k
• New algorithms for T>k; old one still available under Advanced settings. Three aspects in the expansion stage were slowing the algorithm down in some cases:
  1. Block size shrank as the selected model was growing. Now setting blocksize as a fraction of the sample size (0.2T by default, selected model must be <0.6T to consider up to 0.8T), so it stays constant.
  2. Continuation if the model was too small or large: this could create many extra steps for little gain. Now limiting to one continuation and then selecting by p-value if too many omitted detected.
  3. Stage A. Now looking for 1 terminal (bad blocks could mean search for further terminals too slow, and perhaps pulling in rubbish). Stage B and onwards search for further terminals, except in quick mode.
  4. Stage C now expands at 2*p_\alpha, then does reduction at 0.5*p\_alpha and running D at p_\alpha (previously this was 2p_\alpha,p_\alpha,and 4p_\alpha for start of D.)
  Now estimates GPS data (T=1000,n=3) SIS in under two seconds rather than almost three minutes.
• System T>k: wrongly penalizing block size by dividing it by the no of eqns (counting pars multiplied by no of eqns)
• Recursive estimation: recursive estimation now done when entering the graph functions NB doing a full re-estimation, which is not strictly needed (prints the eqn again), but is easiest way.
• Dropped RALS
• Dropped single eqn IV/OLS choice in settings, because it wasn't really a choice.
• Changed saturation selection in dialog
• Added TIS to saturation selection (note construction of trends! None extend into the future - need a forced Trend for that - check)
• SEM t-probabilities, standard errors of recursive residuals, degrees of freedom of recursive Chow tests now based on T-c instead of T-k, where c is the average number of coefficients in an equation, whereas k is the total number of Z and U variables in the system (not all need to appear in the model).
• Added recursive estimation for 1SLS,2SLS,3SLS.
• Added AR test for 1SLS,2SLS,3SLS.
• Added SEM ARCH test (performed on RF residuals)
• Changed SEM AR test for FIML: it is based now on the 3SLS likelihoods. This avoids a numerical maximization, which is particularly useful when doing automatic model selection.
• Recursive hedgehog for SEM
• Autometrics to implement automatic model selection for SEM (FIML,3SLS, 2SLS, 1SLS, CFIML)
• Added line search to switching algorithm for restricted cointegration (as in paper)