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