# Change Logs¶

## Changes since 4.0¶

• Enable user to specify a specific value of the backcast in place of the automatically generated value.
• Fixed a big where parameter-less models where incorrectly reported as having constant variance (GH248)
• Added support for MIDAS volatility processes using Hyperbolic weighting (GH233)
• Added a parameter to forecast that allows a user-provided callable random generator to be used in place of the model random generator. (GH225)
• Added a low memory automatic lag selection method that can be used with very large time-series.
• Improved performance of automatic lag selection in ADF and related tests.
• Added named parameters to Dickey-Fuller regressions.
• Removed use of the module-level NumPy RandomState. All random number generators use separate RandomState instances.
• Fixed a bug that prevented 1-step forecasts with exogenous regressors
• Added the Generalized Error Distribution for univariate ARCH models
• Fixed a bug in MCS when using the max method that prevented all included models from being listed
• Added FixedVariance volatility process which allows pre-specified variances to be used with a mean model. This has been added to allow so-called zig-zag estimation where a mean model is estimated with a fixed variance, and then a variance model is estimated on the residuals using a ZeroMean variance process.
• Fixed a bug that prevented fix from being used with a new model (GH156)
• Added first_obs and last_obs parameters to fix to mimic fit
• Added ability to jointly estimate smoothing parameter in EWMA variance when fitting the model
• Added ability to pass optimization options to ARCH model estimation (GH195)

## Changes since 3.0¶

• Added forecast code for mean forecasting
• Added fix to arch models which allows for user specified parameters instead of estimated parameters.
• Added Hansen’s Skew T distribution to distribution (Stanislav Khrapov)
• Updated IPython notebooks to latest IPython version
• Bug and typo fixes to IPython notebooks
• Changed MCS to give a pvalue of 1.0 to best model. Previously was NaN
• Removed hold_back and last_obs from model initialization and to fit method to simplify estimating a model over alternative samples (e.g., rolling window estimation)
• Redefined hold_back to only accept integers so that is simply defined the number of observations held back. This number is now held out of the sample irrespective of the value of first_obs.