Changelog

All notable changes to this project will be documented in this file. We keep track of changes in this file since v0.1.8. The source code for all releases is available on GitHub.

[0.6.4] - 2022-02-25

Added

  • added residuals to the function export_fitted_vals(). Now gives dates, actuals, fitted vals, and residuals

  • added multiseries.export_model_summaries() function

Changed

Fixed

  • notebook.results_vis() broke after last update and has now been fixed

[0.6.2] - 2022-02-25

Added

Changed

  • the plot(), plot_test_set(), and plot_fitted() functions now return figures instead of automatically plotting for you. This means more customization is now possible. Took out to_png and related args from these functions since that can be done now with matplotlib

Fixed

  • changed the insample metric evaluations for RNN and LSTM to be on the full training set instead of just the last few observations.

[0.6.1] - 2022-02-23

Added

Changed

Fixed

  • fixed an issue when calling RNN or LSTM models after a combo model that saved attributes of the combo model illogically in history

  • fixed an issue that caused test-set evaluation of ARIMA models to be inaccurate (causing severe underperformance)

[0.6.0] - 2022-02-09

Added

  • Added drop argument to several adder functions, giving the user the option to drop the original regressors after making certain transformations on them

Changed

  • add_ar_terms() now accepts 0 as an argument but it doesn’t do anything

Fixed

  • fixed issue with pop_using_criterion() function that wasn’t dropping models correctly in some instances

  • fixed the default parameters for the rnn model which weren’t working due to mislabeling of one of the parameters in TensorFlow. updated the docstring accordingly

[0.5.9] - 2022-02-01

Added

  • Added CurrentEstimator to __repr__() function.

Changed

Fixed

  • Made it impossible to forecast without adding future dates first

[0.5.8] - 2022-01-27

Added

Changed

  • Made it impossible to pass tune argument to manual_forecast()

Fixed

  • Fixed results_vis in notebook (wasn’t displaying one of the widgets)

[0.5.7] - 2022-01-26

Added

Changed

  • Cleaned up a lot of documentation

Fixed

  • Fixed links in readme

[0.5.6] - 2022-01-25

Added

  • Added documentation on Read the Docs

  • Added cilvel information to export functions

Changed

Fixed

[0.5.5] - 2022-01-20

Added

  • Added CILevel info to export model_summaries function

Changed

Fixed

  • Fixed an issue where plots were diplaying incorrect confidence levels if cilevel had been changed since training it

[0.5.4] - 2022-01-20

Added

  • Added the rnn estimator

Changed

  • plot_loss argument no longer considered a hyperparameter value for LSTM and RNN models

Fixed

  • Fixed an issue where “==” wasn’t being accepted in the evaluated_as argument in the pop_using_criterion() function

  • Scaler in history saved as ‘minmax’ instead of None for LSTM and RNN models

[0.5.3] - 2022-01-18

Added

  • EvaluatedModels to __repr__()

Changed

  • Does not call infer_freq() as often, making code more efficient

Fixed

  • sometimes the attribute ci_bootstrap_samples was being called bootstrap_samples, changed everything to bootstrap_samples only

[0.5.2] - 2022-01-12

Added

Changed

Fixed

  • Fixed an error that occured when calling the __repr__() method if no models had been evaluated first

[0.5.1] - 2022-01-12

Added

  • Added export_fitted_vals() function

  • Added ci option to the results_vis() function in notebook

  • Added the get_funcs() function

Changed

  • No Xvars in LSTM model, changed to lags (now model will only look at its own history)

  • No normalizer in LSTM model (always uses a minmax scaler now)

  • LSTM model can no longer be tuned

  • Got rid of all lstm model grids

  • changed __str__() and __repr__() so that they now offer better info

Fixed

  • Fixed the LSTM model by scaling the dependent variable and unscaling it (minmax) when it comes out and getting rid of other Xvars

[0.5.0] - 2022-01-10

Added

  • Added confidence intervals using bootstrapping

    • set_cilvel() function (default .95)

    • set_bootstrap_samples() function (default 100)

  • Added ci parameter to plot() and plot_test_set() function

  • Added UpperCI, LowerCI, TestSetUpperCI, TestSetLowerCI keys to history dict

  • Added export_forecasts_with_cis() and export_test_set_preds_with_cis() functions

  • Added source code (commented out) to get level confidence intervals – when I tested, the intervals were too large to implement but maybe in the future this will be revisited

Changed

Fixed

[0.4.4] - 2022-01-07

Added

  • added lstm grid in example grids

  • added EarlyStopping callback functionality for the LSTM model

  • added get_expanded_lstm_grid() to GridGenerator module which gives an example of a grid with early stopping

Changed

  • changed default paramaters for the lstm model

  • added **kwargs to the lstm model forecast function that are passed to the fit() function in TensorFlow, got rid of epochs and batch_size args consequently

Fixed

[0.4.25] - (quick fix) 2022-01-06

Added

Changed

Fixed

  • source code was using f instead of self when when calling pop_using_criterion()

[0.4.2] - 2022-01-06

Added

  • lstm estimator

  • added pop_using_criterion() function

Changed

  • Fixed an issue where sklearn models were being fit on the same data twice – does not change outcomes but the models run faster now

  • Output from _scale() function is now always a numpy matrix (could have been either that or pandas dataframe before)

  • sorted _estimators_ list

  • changed the error message for when importing a grid fails to account for one other possible reason the failure occured

Fixed

[0.4.1] - 2021-12-30

Added

  • Can now sort by metric value in export_all_validation_grids_to_excel()

Changed

Fixed

  • Fixed an issue where sometimes the incorrect AR terms were being propogated for test-set evaluation only

[0.4.0] - 2021-12-30

Added

Changed

  • deleted the ‘scale’ normalizer from the mlp grid

Fixed

  • Fixed an issue with the PowerTransformer normalizer that failed because of a DivideByZero error, now defaults to a StandardScaler when this issue is encountered and logs a warning

[0.3.9] - 2021-12-30

Added

  • Added init_dates and levely attributes

  • Added 'Observations' info to history and export()

  • Added 'lvl_test_set_predictions' to export dataframes

Changed

  • Got rid of first_obs and first_dates attributes and wrote more efficient code to do what they were there for

  • More information available when __str__() is called

Fixed

  • Fixed what became an issue with the last update in which when calling add_diffed_terms() or add_lagged_terms(), the level series wasn’t accurate due to how undifferencing was being executed. After examining this issue, it became evident that the previous way to undifference forecasts was less efficient than it should have been, this update fixed the issues from the last update and made the code more efficient

  • Fixed an issue where AR terms were manipulating the underlying xreg structures so that each forecast were using its own test-set propogated AR values instead of the correct AR values

  • Fixed the export_Xvars_df() method which wasn’t working correctly if at least one forecast hadn’t been called first

[0.3.8] - 2021-12-29

Added

  • added the following functions that can each add additional Xvars to forecast with:

    • add_exp_terms() - for non polynomial exponential transformations

    • add_logged_terms() - for log of any base transformations

    • add_pt_terms() - for individual variable power transformations (box cox and yeo johnson available)

    • add_diffed_terms() - to difference non-y terms

    • add_lagged_terms() - to lag non-y terms

  • added the ‘pt’ normalizer for yeo-johnson normalization (in addition to ‘minmax’, ‘normalize’, and ‘scale’)

  • added the drop_Xvars() function that is identical to the drop_regressors() function

Changed

  • imports all sklearn models as soon as scalecast is imported

  • src code cleanup with better coding practices when it comes to forecasting sklearn models (no more copying and pasting new functions)

  • changed several set data types to lists in src code

  • changed the names of some hidden functions

  • other src code cleanup for readability and minor efficiency gains

  • better in-line comments and docstring documentation

  • got rid of quiet paramater in save_summary_stats() and save_feature_importance() and now these simply log any problems as warnings

  • time trends now start at 1 instead of 0 (makes log transformations possible)

  • observation dropping for AR terms in sklearn models now based on the number of N/A values in each AR term instead of just the AR number

  • changed some example grids to include the pt normalizer

Fixed

  • now logs all warnings

[0.3.7] - 2021-12-27

Added

  • dynamic_testing argument to manual_forecast() and auto_forecast() functions – this is True by default (makes all testing comparable between sklearn/non-sklearn models)

  • dynamic_tuning argument to tune() function – this is False by default to majorly improve speed in some applications

Changed

  • native Forecaster warnings will be logged

Fixed

[0.3.6] - 2021-12-14

Added

  • added tune_test_forecast() function to notebook module to create a progress bar when using a notebook

Changed

Fixed

  • fixed an issue with Forecaster.ingest_Xvars_df() when use_future_dates=False causing an error to be raised

[0.3.5] - 2021-12-07

Added

  • added include_traing parameter to notebook.results_vis() function

Changed

Fixed

  • fixed print_attr parameter default in notebook.results_vis()

[0.3.4] - 2021-12-07

Added

  • added results_vis() notebook function (requires ipywidgets)

  • added Forecaster.export_Xvars_df() function

  • added max_integration argument to the Forecaster.integrate() function

Changed

Fixed

[0.3.3] - 2021-11-26

Added

Changed

  • Now reloads Grids file each time ingest_grid() is called so that notebooks do not have to be rerun when a grid cannot be found

Fixed

  • Fixed an issue with some sklearn estimators that occurs when passing a subset of regressors in a list to the forecast function

[0.3.2] - 2021-11-01

Added

Changed

Fixed

  • Found an issue when using floor in Prophet

[0.3.1] - 2021-10-29

Added

  • Added the eCommerce example

  • In limit_grid_size(), users can now set random_seed parameter for consistent results

Changed

Fixed

  • Scikit-learn models were not accepting Xvars='all' as an arguments

  • Fixed an issue causing models run at different levels to error out sometimes when plotted

  • Fixed a plotting error that occured sometimes when setting models parameter to None

[0.3.0] - 2021-10-15

Added

  • Added an option to save to png in plot(), plot_test_set(), and plot_fitted() methods using plt.savefig() from matplotlib and calling with to_png = True

Changed

  • Made errors more descriptive, stripping out AssertionError types

Fixed

  • fixed typos in doc strings

[0.2.9] - 2021-09-27

Added

Changed

  • In plot() method, models=None is now accepted and will plot only actual values

  • Example grids are modified to prevent overfitting in some models

Fixed

  • Fixed the add_time_trend() method to not skip a time step in the first observation

[0.2.8] - 2021-08-27

Added

  • Added a descriptive error when all_feature_info_to_excel() or all_validation_grids_to_excel() fails

Changed

  • Using pd.shift() instead of np.roll() to create AR terms to avoid further issues with AR terms

  • Prophet, silverkite, and ARIMA have better Xvar validation mechanisms to ensure that autoregressive terms aren’t fed to them, which could cause errors and doesn’t add anything to the models that isn’t already built into them. Now, even if a user tries to feed AR terms only, it will pass no Xvars to these models

Fixed

  • AR terms were not dropping the correct first observations before being estimated with SKLEARN models, so we fixed that but it didn’t seem to make a noticeable difference in any of the examples

[0.2.7] - 2021-08-20

Added

  • added reset() function that deletes all regressors and resets the object to how it was initiated

  • added documentation and hints in the source code

Changed

  • changed readme documentation to be more concise

Fixed

  • in the documentation, it was stated that the ‘scale’ value passed to the ‘normalize’ parameter when calling manual_forecast() or auto_forecast() would use a StandardScaler from sklearn, but a Normalizer was actually being applied. Now, you can pass ‘scale’ to get the StandardScaler, ‘normalize’ to get the Normalizer, or ‘minmax’ to get the MinMaxScaler (unchanged from previous distributions). ‘minmax’ is still the default for all estimators that accept this argument

[0.2.6] - 2021-08-12

Added

  • added train_only argument to following functions to reduce data leakage in eda/preprocessing steps: integrate, adf_test, plot_acf, plot_pacf, plot_periodogram, seasonal_decompose – default argument is still False for these. Now it is suggested to set a test length before running any one of these methods and only examine the training set correlations to prevent data leakage

Changed

Fixed

[0.2.5] - 2021-08-09

Added

  • added integrate() method that can be used to automatically find the series appropriate level to achieve stationarity, according to augmented dickey fuller test

Changed

Fixed

[0.2.4] - 2021-08-03

Added

  • added tune_test_forecast() function that allows what used to take four lines to be aggregated into one, also allows more easy saving of feature information by setting feature_importance or summary_stats parameters to True

  • added all_feature_info_to_excel() function

  • added all_validation_grids_to_excel() function

Changed

Fixed

  • removed a duplicate column from the dataframe created when calling the export() method

[0.2.3] - 2021-07-19

Added

Changed

  • changed removed pandas-datareader from imports in setup.py since it is not a package dependency (change having to do with installation only and should not affect anything when applying the library)

Fixed

[0.2.2] - 2021-07-16

Added

  • added GridGenerator module so user can more easily create grids in working directory

Changed

  • changed all functions with diffy parameter (plot_acf, plot_pacf, seasonal_decompose) now accept True, False, 0, 1, or 2 as possible values

Fixed

  • fixed issues with two-level undifferencing where it was adding values exponentially because the level was being added to the first and second-level differences

  • fixed issues with two-level undifferencing where dates were being mixed up

  • fixed issues with one-level test-set evaluation where the incorrect initial value was set to undifference values in the test-set only, causing miscalculation of metrics, although the bias was in both directions so when rerunning avocados.ipynb, for example, the results were virtually the same with different models now outperforming others but metrics remaining more or less the same on average; forecasted values did not change

[0.1.9] - 2021-07-09

Added

  • added lightgbm and silverkite as estimators

Changed

  • changed ‘which’ parameter in set_valiation_metric() to ‘metric’ for clarity

  • changed ‘which’ parameter in set_estimator() to ‘estimator’ for clarity

Fixed

[0.1.8] - 2021-07-05

Added

Changed

Fixed

  • fixed an error in combo modeling that was causing incorrect applications of weights in weighted averaging – the weights were generating correctly but not being applied to the best-performing models in the correct order