NEWS.md

In thfuchs/tsRNN: Time Series Forecasting using Recurrent Neural Networks by Keras

tsRNN 0.2.2

Breaking Changes

• Bug in smis() calculation. Function compared prediction interval with prediction instead of actual values. Argument data now requires train set only and actual equivalent to lower/upper bounds. Argument forecast and h therefore removed.
• Bug in mase() calculation. Fixed scaling for forecast horizons not immediately following the train set, e.g. when evaluating only on h = 5:6. Therefore changed arguments to
• data for train set
• actual for actual values, the equivalent of forecast values
• forecast for predictions
• The arguments are finally consistent across all metrics

Minor Changes

• sMIS and MASE fixes with direct impact on cv_arima, cv_baselines and tune_keras_rnn_eval

tsRNN 0.2.1

Breaking Changes

• Finished function tune_keras_rnn_eval incl. internal argument checks, unit testing and examples
• Renamed argument multiple_h to h in following functions:
• cv_arima
• cv_baselines
• tune_keras_rnn_eval
• Fixed smis calculation for differing forecast horizons. Directly affecting output of cv_arima, cv_baselines and tune_keras_rnn_eval

tsRNN 0.2.0

New Features

• Created parse_tf_version to get TensorFlow version

Minor Changes

• (Finished) unit testing by tinytest
• Created internal function check_cv_setting for automated check of cv_setting in functions including cross-validation
• Enhanced functions with (better) intenal argument checks. keras_rnn-function family in particular
• Enhanced functions with examples
• Reconstruct metrics (metrics_dist and metrics_point) documentation
• Remove documentation of internal utils functions

tsRNN 0.1.4

Important Fix

• predict_baselines, predict_arima, tune_keras_rnn_bayesoptim, tune_keras_rnn_predict All functions checked id column for class numeric. Fixed bug to check for class character

Minor Changes

• tune_keras_rnn_bayesoptim Change bayes optimization iterations from 50 to 30
• tune_keras_rnn_bayesoptim and tune_keras_rnn_predict: Fix resample naming once purrr loop finished

tsRNN 0.1.3

Splitting tune_keras_rnn

New Features

• tune_keras_rnn_bayesoptim tuning process by Bayesian Optimization
• tune_keras_rnn_predict Train and forecast based on tuning parameters
• tune_keras_rnn_eval evaluation of trained models

tsRNN 0.1.2

Minor Fixes

• tune_keras_rnn development of grid search for PI dropout rate finished
• py_dropout_model slight bug fixed
• keras_rnn enhanced by "optimizer" argument

tsRNN 0.1.1

Minor Fixes

• predict_arima and predict_baselines: Now using stats::window for subset.ts

tsRNN 0.1.0

First (official) release of package tsRNN (former fcf) for time series recurrent neural network training and estimation

New Features

• check_acf Check for autocorrelation (Ljung-Box tests for 8, 12 and 16 lags)
• forecast_baseline Baseline models to get benchmark results for advanced models
• keras_rnn Train Recurrent Neural Network with Gated Recurrent Unit (GRU) or Long-Short Term Memory (LSTM) using Keras framework
• acd Absolute coverage difference (ACD)
• smis Scaled Mean Interval Score
• mase Mean Absolute Scaled Error (MASE)
• mape Mean Absolute Percentage Error (MAPE)
• smape symmetric Mean Absolute Percentage Error (sMAPE)
• plot_baselines Plot Forecasts by baseline methods
• plot_baselines_samples Plot cross validated samples of forecasts by baseline methods
• plot_prediction Plot timeseries and forecast for single split and company
• plot_prediction_samples Plot multiple splits from list with forecast results
• predict_arima Forecast the next n steps by ARIMA for each split
• predict_baselines Prediction and evaluation for baseline models including cross validation
• predict_keras_rnn Prediction and evaluation for rnn models
• py_dropout_model Change dropout rate in recurrent layer or dropout layer of trained model (python wrapper)
• ts_nn_preparation Timeseries data preparation for neural network Keras models
• ts_normalization Normalize univariate timeseries
• tune_keras_rnn Tune recurrent neural network with Keras functional API and Bayes-Optimization and select best performing model

thfuchs/tsRNN documentation built on April 17, 2021, 11:03 p.m.