NEWS.md

timetk 2.9.0

Anomalize Integration:

anomalize R package is now available in timetk:

1. anomlize(): 1 function that breaks down, identifies, and cleans anomalies
2. plot_anomalies(): Visualize the anomalies and anomaly bands
3. plot_anomalies_decomp(): Visualize the time series decomposition. Make adjustments as needed.
4. plot_anomalies_cleaned(): Visualize the before/after of cleaning anomalies.

Note - anomalize(.method): Only the .method = "stl" is supported at this time. The "twitter" method is also planned.

Other Changes:

• Removed dependency on tidymodels. (#154, @olivroy).

timetk 2.8.4

Update forecasting vignette: Use glmnet for time series forecasting.

CRAN Fixes:

• tzdata time zone fixes:
  • GB -> Europe/London
  • NZ -> Pacific/Auckland
  • US/Eastern -> America/New_York
  • US/Pacific -> America/Los_Angeles
• Add @aliases to timetk-package

timetk 2.8.3

• remove support for robets
• remove tidyquant from examples
• remove tidyverse from examples
• add FANG dataset to timetk (port from tidyquant)
• cran: fix return, dontrun -> donttest, options(max.print)

timetk 2.8.2

New Features

• plot_time_series(): Gets new arguments to specify .x_intercept and .x_intercept_color. #131

Fixes

• Fix error in plot_time_series() when .group_names is not found. #121
• Merge variable checking update needed for recipes >= 1.0.3 #132

timetk 2.8.1

Trelliscope Plotting

• Expose the facet_trelliscope() plotting parameters.
  • plot_time_series()
  • plot_time_series_boxplot()
  • plot_anomaly_diagnostics()

timetk 2.8.0

New Features

Many of the plotting functions have been upgraded for use with trelliscopejs for easier visualization of many time series.

• plot_time_series():

  • Gets a new argument trelliscope: Used for visualizing many time series.
  • Gets a new argument .facet_strip_remove to remove facet strips since trelliscope is automatically labeled.
  • Gets a new argument .facet_nrow to adjust grid with trelliscope.
  • The default argument for facet_collapse = TRUE was changed to FALSE for better compatibility with Trelliscope JS. This may cause some plots to have multiple groups take up extra space in the strip.

• plot_time_series_boxplot():

  • Gets a new argument trelliscope: Used for visualizing many time series.
  • Gets a new argument .facet_strip_remove to remove facet strips since trelliscope is automatically labeled.
  • Gets a new argument .facet_nrow to adjust grid with trelliscope.
  • The default argument for .facet_collapse = TRUE was changed to FALSE for better compatibility with Trelliscope JS. This may cause some plots to have multiple groups take up extra space in the strip.

• plot_anomaly_diagnostics():

  • Gets a new argument trelliscope: Used for visualizing many time series.
  • Gets a new argument .facet_strip_remove to remove facet strips since trelliscope is automatically labeled.
  • Gets a new argument .facet_nrow to adjust grid with trelliscope.
  • The default argument for .facet_collapse = TRUE was changed to FALSE for better compatibility with Trelliscope JS. This may cause some plots to have multiple groups take up extra space in the strip.

Updates & Bug Fixes

• Recipes steps (e.g. step_timeseries_signature()) use the new recipes::print_step() function. Requires recipes >= 0.2.0. #110

• Offset parameter in step_log_interval() was not working properly. Now works. #103

Potential Breaking Changes

• The default argument for .facet_collapse = TRUE was changed to FALSE for better compatibility with Trelliscope JS. This may cause some plots to have multiple groups take up extra space in the strip.

timetk 2.7.0

New Features

• tk_tsfeatures(): A new function that makes it easy to generate time series feature matrix using tsfeatures. The main benefit is that you can pipe time series data in tibbles with dplyr groups. The features will be produced by group. #95 #84

• plot_time_series_boxplot(): A new function that makes plotting time series boxplots simple using a .period argument for time series aggregation.

New Vignettes

• Time Series Clustering: Uses the new tk_tsfeatures() function to perform time series clustering. #95 #84

• Time Series Visualization: Updated to include plot_time_series_boxplot() and plot_time_series_regression().

Improvements

Improvements for point forecasting when the target is n-periods into the future.

• time_series_cv(), time_series_split(): New parameter point_forecast. This is useful for testing / assessing the n-th prediction in the future. When set to TRUE, will return a single point that returns on the last value in assess.

Fixes

• Updates for rlang > 0.4.11 (dev version) #98
• plot_time_series(): Smoother no longer fails when time series has 1 observation #106

timetk 2.6.2

Improvements

• summarize_by_time(): Added a .week_start argument to allow specifying .week_start = 1 for Monday start. Default is 7 for Sunday Start. This can also be changed with the lubridate by setting the lubridate.week.start option.

• Plotting Functions:

  • Several plotting functions gain a new .facet_dir argument for adjusting the direction of facet_wrap(dir). #94
  • Plot ACF Diagnostics (plot_acf_diagnostics()): Change default parameter to .show_white_noise_bars = TRUE. #85
  • plot_timeseries_regression(): Can now show_summary for group-wise models when visualizing groups

• Time Series CV (time_series_cv()): Add Label for tune_results

• Improve speed of pad_by_time(). #93

Bug Fixes

• tk_make_timeseries() and tk_make_future_timeseries() are now able to handle end of months. #72

• tk_tbl.zoo(): Fix an issue when readr::type_convert() produces warning messages about not having character columns in inputs. #89

• plot_time_series_regression(): Fixed an issue when lags are added to .formula. Pads lags with NA.

• step_fourier() and fourier_vec(): Fixed issue with step_fourier failing with one observation. Added scale_factor argument to override date sequences with the stored scale factor. #77

timetk 2.6.1

Improvements

• tk_augment_slidify(), tk_augment_lags(), tk_augment_leads(), tk_augment_differences(): Now works with multiple columns (passed via .value) and tidyselect (e.g. contains()).

Fixes

• Reduce "New names" messages.

#> New names: 
#> * NA -> ...1 

• Remove dependency on lazyeval. #24
• Fix deprecated functions: select_() used with tk_xts_(). #52

timetk 2.6.0

New Functions

• filter_period() (#64): Applies filtering expressions within time-based periods (windows).
• slice_period() (#64): Applies slices within time-based periods (windows).
• condense_period() (#64): Converts a periodicity from a higher (e.g. daily) to lower (e.g. monthly) frequency. Similar to xts::to.period() and tibbletime::as_period().
• tk_augment_leads() and lead_vec() (#65): Added to make it easier / more obvious on how to create leads.

Fixes

• time_series_cv(): Fix bug with Panel Data. Train/Test Splits only returning 1st observation in final time stamp. Should return all observations.
• future_frame() and tk_make_future_timeseries(): Now sort the incoming index to ensure dates returned go into the future.
• tk_augment_lags() and tk_augment_slidify(): Now overwrite column names to match the behavior of tk_augment_fourier() and tk_augment_differences().

timetk 2.5.0

Improvements

• time_series_cv(): Now works with time series groups. This is great for working with panel data.
• future_frame(): Gets a new argument called .bind_data. When set to TRUE, it performs a data binding operation with the incoming data and the future frame.

Miscellaneous

• Tune startup messages (#63)

timetk 2.4.0

• step_slidify_augment() - A variant of step slidify that adds multiple rolling columns inside of a recipe.

Bug Fixes

• Add warning when %+time% and %-time% return missing values
• Fix issues with tk_make_timeseries() and tk_make_future_timeseries() providing odd results for regular time series. GitHub Issue 60

timetk 2.3.0

New Functionality

• tk_time_series_cv_plan() - Now works with k-fold cross validation objects from vfold_cv() function.

• pad_by_time() - Added new argument .fill_na_direction to specify a tidyr::fill() strategy for filling missing data.

Bug Fixes

• Augment functions (e.g. tk_augment_lags()) - Fix bug with grouped functions not being exported
• Vectorized Functions - Compatabiliy with ts class

timetk 2.2.1

New Functions

• step_log_interval_vec() - Extends the log_interval_vec() for recipes preprocessing.

Parallel Processing

• Parallel backend for use with tune and recipes

Bug Fixes

• log_interval_vec() - Correct the messaging
• complement.ts_cv_split - Helper to show time series cross validation splits in list explorer.

timetk 2.2.0

New Functions

• mutate_by_time(): For applying mutates by time windows
• log_interval_vec() & log_interval_inv_vec(): For constrained interval forecasting.

Improvements

• plot_acf_diagnostics(): A new argument, .show_white_noise_bars for adding white noise bars to an ACF / PACF Plot.
• pad_by_time(): New arguments .start_date and .end_date for expanding/contracting the padding windows.

timetk 2.1.0

New Functions

• plot_time_series_regression(): Convenience function to visualize & explore features using Linear Regression (stats::lm() formula).
• time_series_split(): A convenient way to return a single split from time_series_cv(). Returns the split in the same format as rsample::initial_time_split().

Improvements

• Auto-detect date and date-time: Affects summarise_by_time(), filter_by_time(), tk_summary_diagnostics
• tk_time_series_cv_plan(): Allow a single resample from rsample::initial_time_split or timetk::time_series_split
• Updated Vignette: The vignette, "Forecasting Using the Time Series Signature", has been updated with modeltime and tidymodels.

Plotting Improvements

• All plotting functions now support Tab Completion (a minor breaking change was needed to do so, see breaking changes below)
• plot_time_series():
  • Add .legend_show to toggle on/off legends.
  • Permit numeric index (fix issue with smoother failing)

Breaking Changes

• Tab Completion: Replace ... with .facet_vars or .ccf_vars. This change is needed to improve tab-completion. It affects :
  • plot_time_series()
  • plot_acf_diagnostics()
  • plot_anomaly_diagnostics()
  • plot_seasonal_diagnostics()
  • plot_stl_diagnostics()

Bug Fixes

• fourier_vec() and step_fourier_vec(): Add error if observations have zero difference. Issue #40.

timetk 2.0.0

New Interactive Plotting Functions

• plot_anomaly_diagnostics(): Visualize Anomalies for One or More Time Series

New Data Wrangling Functions

• future_frame(): Make a future tibble from an existing time-based tibble.

New Diagnostic / Data Processing Functions

• tk_anomaly_diagnostics() - Group-wise anomaly detection and diagnostics. A wrapper for the anomalize R package functions without importing anomalize.

New Vectorized Functions:

• ts_clean_vec() - Replace Outliers & Missing Values in a Time Series
• standardize_vec() - Centers and scales a time series to mean 0, standard deviation 1
• normalize_vec() - Normalizes a time series to Range: (0, 1)

New Recipes Preprocessing Steps:

• step_ts_pad() - Preprocessing for padding time series data. Adds rows to fill in gaps and can be used with step_ts_impute() to interpolate going from low to high frequency!
• step_ts_clean() - Preprocessing step for cleaning outliers and imputing missing values in a time series.

New Parsing Functions

• parse_date2() and parse_datetime2(): These are similar to readr::parse_date() and lubridate::as_date() in that they parse character vectors to date and datetimes. The key advantage is SPEED. parse_date2() uses anytime package to process using C++ Boost.Date_Time library.

Improvements:

• plot_acf_diagnostics(): The .lags argument now handles time-based phrases (e.g. .lags = "1 month").
• time_series_cv(): Implements time-based phrases (e.g. initial = "5 years" and assess = "1 year")
• tk_make_future_timeseries(): The n_future argument has been deprecated for a new length_out argument that accepts both numeric input (e.g. length_out = 12) and time-based phrases (e.g. length_out = "12 months"). A major improvement is that numeric values define the number of timestamps returned even if weekends are removed or holidays are removed. Thus, you can always anticipate the length. (Issue #19).
• diff_vec: Now reports the initial values used in the differencing calculation.

Bug Fixes:

• plot_time_series():
  • Fix name collision when .value = .value.
• tk_make_future_timeseries():
  • Respect timezones
• time_series_cv():
  • Fix incorrect calculation of starts/stops
  • Make skip = 1 default. skip = 0 does not make sense.
  • Fix issue with skip adding 1 to stops.
  • Fix printing method
• plot_time_series_cv_plan() & tk_time_series_cv_plan():
  • Prevent name collisions when underlying data has column "id" or "splits"
• tk_make_future_timeseries():
  • Fix bug when day of month doesn't exist. Lubridate period() returns NA. Fix implemented with ceiling_date().
• pad_by_time():
  • Fix pad_value so only inserts pad values where new row was inserted.
• step_ts_clean(), step_ts_impute():
  • Fix issue with lambda = NULL

Breaking Changes:

These should not be of major impact since the 1.0.0 version was just released.

• Renamed impute_ts_vec() to ts_impute_vec() for consistency with ts_clean_vec()
• Renamed step_impute_ts() to step_ts_impute() for consistency with underlying function
• Renamed roll_apply_vec() to slidify_vec() for consistency with slidify() & relationship to slider R package
• Renamed step_roll_apply to step_slidify() for consistency with slidify() & relationship to slider R package
• Renamed tk_augment_roll_apply to tk_augment_slidify() for consistency with slidify() & relationship to slider R package
• plot_time_series_cv_plan() and tk_time_series_cv_plan(): Changed argument from .rset to .data.

timetk 1.0.0

New Interactive Plotting Functions:

• plot_time_series() - A workhorse time-series plotting function that generates interactive plotly plots, consolidates 20+ lines of ggplot2 code, and scales well to many time series using dplyr groups.
• plot_acf_diagnostics() - Visualize the ACF, PACF, and any number of CCFs in one plot for Multiple Time Series. Interactive plotly by default.
• plot_seasonal_diagnostics() - Visualize Multiple Seasonality Features for One or More Time Series. Interactive plotly by default.
• plot_stl_diagnostics() - Visualize STL Decomposition Features for One or More Time Series.
• plot_time_series_cv_plan() - Visualize the Time Series Cross Validation plan made with time_series_cv().

New Time Series Data Wrangling:

• summarise_by_time() - A time-based variant of dplyr::summarise() for flexible summarization using common time-based criteria.
• filter_by_time() - A time-based variant of dplyr::filter() for flexible filtering by time-ranges.
• pad_by_time() - Insert time series rows with regularly spaced timestamps.
• slidify() - Make any function a rolling / sliding function.
• between_time() - A time-based variant of dplyr::between() for flexible time-range detection.
• add_time() - Add for time series index. Shifts an index by a period.

New Recipe Functions:

Feature Generators:

• step_holiday_signature() - New recipe step for adding 130 holiday features based on individual holidays, locales, and stock exchanges / business holidays.
• step_fourier() - New recipe step for adding fourier transforms for adding seasonal features to time series data
• step_roll_apply() - New recipe step for adding rolling summary functions. Similar to recipes::step_window() but is more flexible by enabling application of any summary function.
• step_smooth() - New recipe step for adding Local Polynomial Regression (LOESS) for smoothing noisy time series
• step_diff() - New recipe for adding multiple differenced columns. Similar to recipes::step_lag().
• step_box_cox() - New recipe for transforming predictors. Similar to step_BoxCox() with improvements for forecasting including "guerrero" method for lambda selection and handling of negative data.
• step_impute_ts() - New recipe for imputing a time series.

New Rsample Functions

• time_series_cv() - Create rsample cross validation sets for time series. This function produces a sampling plan starting with the most recent time series observations, rolling backwards.

New Vector Functions:

These functions are useful on their own inside of mutate() and power many of the new plotting and recipes functions.

• roll_apply_vec() - Vectorized rolling apply function - wraps slider::slide_vec()
• smooth_vec() - Vectorized smoothing function - Applies Local Polynomial Regression (LOESS)
• diff_vec() and diff_inv_vec() - Vectorized differencing function. Pads NA's by default (unlike stats::diff).
• lag_vec() - Vectorized lag functions. Returns both lags and leads (negative lags) by adjusting the .lag argument.
• box_cox_vec(), box_cox_inv_vec(), & auto_lambda() - Vectorized Box Cox transformation. Leverages forecast::BoxCox.lambda() for automatic lambda selection.
• fourier_vec() - Vectorized Fourier Series calculation.
• impute_ts_vec() - Vectorized imputation of missing values for time series. Leverages forecast::na.interp().

New Augment Functions:

All of the functions are designed for scale. They respect dplyr::group_by().

• tk_augment_holiday_signature() - Add holiday features to a data.frame using only a time-series index.
• tk_augment_roll_apply() - Add multiple columns of rolling window calculations to a data.frame.
• tk_augment_differences() - Add multiple columns of differences to a data.frame.
• tk_augment_lags() - Add multiple columns of lags to a data.frame.
• tk_augment_fourier() - Add multiple columns of fourier series to a data.frame.

New Make Functions:

Make date and date-time sequences between start and end dates.

• tk_make_timeseries() - Super flexible function for creating daily and sub-daily time series.
• tk_make_weekday_sequence() - Weekday sequence that accounts for both stripping weekends and holidays
• tk_make_holiday_sequence() - Makes a sequence of dates corresponding to business holidays in calendars from timeDate (common non-working days)
• tk_make_weekend_sequence() - Weekday sequence of dates for Saturday and Sunday (common non-working days)

New Get Functions:

• tk_get_holiday_signature() - Get 100+ holiday features using only a time-series index.
• tk_get_frequency() and tk_get_trend() - Automatic frequency and trend calculation from a time series index.

New Diagnostic / Data Processing Functions

• tk_summary_diagnostics() - Group-wise time series summary.
• tk_acf_diagnostics() - The data preparation function for plot_acf_diagnostics()
• tk_seasonal_diagnostics() - The data preparation function for plot_seasonal_diagnostics()
• tk_stl_diagnostics() - Group-wise STL Decomposition (Season, Trend, Remainder). Data prep for plot_stl_diagnostics().
• tk_time_series_cv_plan - The data preparation function for plot_time_series_cv_plan()

New Datasets

• M4 Competition - Sample "economic" datasets from hourly, daily, weekly, monthly, quarterly, and yearly.
• Walmart Recruiting Retail Sales Forecasting Competition - Sample of 7 retail time series
• Web Traffic Forecasting (Wikipedia) Competition - Sample of 10 website time series
• Taylor's Energy Demand - Single time series with 30-minute interval of energy demand
• UCI Bike Sharing Daily - A time series consisting of Capital Bikesharing Transaction Counts and related time-based features.

Improvements:

• tk_make_future_timeseries() - Now accepts n_future as a time-based phrase like "12 seconds" or "1 year".

Bug Fixes:

• Don't set timezone on date - Accommodate recent changes to lubridate::tz<- which now returns POSIXct when used Date objects. Fixed in PR32 by @vspinu.

Potential Breaking Changes:

• tk_augment_timeseries_signature() - Changed from data to .data to prevent name collisions when piping.

timetk 0.1.3

New Features:

• recipes Integration - Ability to apply time series feature engineering in the tidymodels machine learning workflow.
  • step_timeseries_signature() - New step_timeseries_signature() for adding date and date-time features.
• New Vignette - "Time Series Machine Learning" (previously forecasting using the time series signature)

Bug Fixes:

• xts::indexTZ is deprecated. Use tzone instead.
• Replace arrange_ with arrange.
• Fix failing tests due to tidyquant 1.0.0 upagrade (single stocks now return an extra symbol column).

timetk 0.1.2

• Compatability with tidyquant v0.5.7 - Removed dependency on tidyverse
• Dependency cleanup - removed devtools and other unncessary dependencies.

timetk 0.1.1

• Added timeSeries to Suggests to satisfy a CRAN issue.

timetk 0.1.0

• Renamed package timetk. Was formerly timekit.
• Improvements:
  • Fixed issue with back-ticked date columns
  • Update pkgdown
  • support for robets