| Software | Version |
|---|---|
Python |
>=3.6 |
Tensorflow |
1.12.0 |
smac |
0.8.0 |
Set the PYTHONPATH env variable of the system. Append absolute paths of both the project root directory and the directory of the external_packages/cocob_optimizer into the PYTHONPATH
For R scripts, make sure to set the working directory to the project root folder.
The data files used for the experiments are available as a Google Drive folder at this link.
Three files need to be created for every model, one per training, validation and testing. Example preprocessing scripts to create text data files are in preprocess_scripts directory.
Sample Record of validation file in moving window format:
1|i -0.120404761911659 -0.138029849544217 -0.158262315258994 -0.117573911196581 -0.047514354625692 -0.054921000563831 -0.087502195404757 -0.0468948356427585 -0.0265522120279886 -0.0259454546421436 -0.0149743425531481 -0.0750882944138711 0.0300152959591582 0.037022158965339 0.0168685236725015 |o -0.0487566395401897 -0.00313169841363781 -0.0356365611845675 0.11323901494058 0.0498791083802557 -0.0222170965043569 0.0324163879281905 0.0353096916266837 0.085162256512592 0.0644041024983562 0.0970988030382305 0.100330957527596 |# 6.88534358640275 -0.00313977170055892 -0.0044384039516765 0.00223114486678285 0.00574405742601041 0.00832797755707766 0.00264786188838812 0.00821557645548867 0.0196038788714076 -0.0082329067304395 -0.0136679361428553 -0.00526828286265864 -0.0120231978314266
input_window_size = 15
output_window_size = 12
meta_data_size (normalization factor + forecast horizon seasonality values) = 1 + output_window_size = 13
Sample Record of validation file in non moving window format:
1|i -0.567828815938088 -0.585453903570646 -0.605686369285423 -0.56499796522301 -0.494938408652121 -0.50234505459026 -0.534926249431186 -0.494318889669188 -0.473976266054418 -0.473369508668573 -0.462398396579577 -0.5225123484403 -0.417408758067271 -0.41040189506109 -0.430555530353928 -0.496180693566619 -0.450555752440067 -0.483060615210997 -0.33418503908585 -0.397544945646174 -0.469641150530786 -0.415007666098239 -0.412114362399746 -0.362261797513837 -0.383019951528073 -0.350325250988199 -0.347093096498833 -0.385629851372523 -0.33765528593077 -0.345559091168428 -0.318001882157549 -0.316676017389535 -0.352726522297433 -0.313242921446948 -0.282102294051863 -0.39715699799074 -0.295846540532483 -0.31886420750242 -0.284579499161778 -0.327664544654104 -0.291079231868399 -0.265463866077922 -0.301299949523727 -0.248215274184239 -0.257073832944534 -0.232876174866125 -0.274657317980361 -0.235741883456098 -0.293700188394566 -0.257319750871742 -0.227493242898714 -0.182332265822538 -0.226778708376941 -0.183411643917652 -0.175304442585818 -0.185429704907653 -0.174450279897044 -0.165652924361955 -0.164673044962693 -0.153843029747357 -0.153398758376659 -0.173864098477157 -0.175699109712323 -0.149000031675115 -0.118212150653881 -0.125336123838077 -0.11559266124817 -0.162446928605872 -0.142920770031087 -0.182453963932645 -0.0843017018306496 -0.0400239133080627 -0.123235344673406 -0.108910774407422 -0.0955200648753625 -0.074071004750234 -0.0886897595416567 -0.107996723281453 -0.0489798864064088 -0.0897408886572117 -0.0556907895876506 -0.113434652776987 -0.0662362512991423 -0.0736203504006339 -0.0543147221095195 -0.0214098332257642 -0.00792657237932826 0.00371722858574586 -0.0390213828131403 -0.0522027300921888 -0.129052228394151 -0.072525671787675 0.00452387954346101 -0.00802511272308681 -0.0258166978136511 0.0131909044091891 |o 0.0296250975171493 0.0263897446789843 0.0333614516505714 0.0821482410736927 0.000100754260286884 0.0588968744959812 -0.00878589639917671 0.0440999722838731 0.0424870606078542 0.0765977431805602 -0.00316200785091869 0.0483167388332202 |# 7.33276764042918 -0.0136679361428553 -0.00526828286265864 -0.0120231978314266 -0.00313977170055892 -0.0044384039516765 0.00223114486678285 0.00574405742601041 0.00832797755707766 0.00264786188838812 0.00821557645548867 0.0196038788714076 -0.0082329067304395
Sample Record of validation file in moving window format without STL Decomposition:
1|i -0.376703850726204 -0.385929285078564 -0.41291666576211 -0.363344835568828 -0.294583911249058 -0.295321008368737 -0.324389290650435 -0.28119801075737 -0.26653550281129 -0.260361030858344 -0.238001616353429 -0.325952353816 -0.226283792855386 -0.210877276569009 -0.237785826830614 |o -0.294527563912438 -0.250201255037004 -0.276036568989474 -0.123648080305099 -0.184424066734356 -0.262200387287658 -0.20199918828801 -0.187717582173598 -0.165701802889537 -0.191894986316188 -0.150800632496117 -0.15432339297552 |# 1246.35022471111
meta_data_size (normalization factor) = 1
For faster execution, the text data files are converted to tfrecord binary format. The tfrecords_handler module converts the text data into tfrecords format (using tfrecord_writer.py) as well as reads in tfrecord data (using tfrecord_reader.py) during execution. Example scripts to convert text data into tfrecords format can be found in the preprocess_scripts directory.
Example bash scripts are in the directory utility_scripts/execution_scripts.
The model expects a number of arguments.
- dataset_name - Any unique string for the name of the dataset
- contain_zero_values - Whether the dataset contains zero values(0/1)
- address_near_zero_instability - Whether the dataset contains zero values(0/1) - Whether to use a custom SMAPE function to address near zero instability(0/1). Default value is 0
- integer_conversion - Whether to convert the final forecasts to integers(0/1). Default is 0
- initial_hyperparameter_values_file - The file for the initial hyperparameter range configurations
- binary_train_file_train_mode - The tfrecords file for train dataset in the training mode
- binary_valid_file_train_mode - The tfrecords file for validation dataset in the training mode
- binary_train_file_test_mode - The tfrecords file for train dataset in the testing mode
- binary_test_file_test_mode - The tfrecords file for test dataset in the testing mode
- txt_test_file - The text file for test dataset
- actual_results_file - The text file of the actual results
- original_data_file - The text file of the original dataset with all the given data points
- cell_type - The cell type of the RNN(LSTM/GRU/RNN). Default is LSTM
- input_size - The input size of the moving window. Default is 0 in the case of non moving window format
- seasonality_period - The seasonality period of the time series
- forecast_horizon - The forecast horizon of the dataset
- optimizer - The type of the optimizer(cocob/adam/adagrad)
- model_type - The type of the model(stacking/seq2seq/seq2seqwithdenselayer)
- input_format - Input format(moving_window/non_moving_window)
- without_stl_decomposition - Whether not to use stl decomposition(0/1). Default is 0
- seed - Integer seed to use as the random seed for hyperparameter tuning
The first point of invoking the models is the generic_model_trainer.py. The generic_model_trainer.py parses the external arguments and identifies the required type of model, optimizer, cell etc... The actual models are inside the directory rnn_architectures.
First, the hyperparameter tuning is carried out using the validation errors of the respective model trainer. Example initial hyperparameter ranges can be found inside the directory configs/initial_hyperparameter_values. The found optimal hyperparameter combination is written to a file in the directory results/optimized_configurations.
Then the found optimal hyperparameter combination is used on the respective model tester to generate the final forecasts. Every model is run on 10 Tensorflow graph seeds (from 1 to 10). The forecasts are written to 10 files inside the directory results/rnn_forecasts.
The forecasts from the 10 seeds are ensembled by taking the median. The utility_scripts/error_summary_scripts/ensembling_forecasts.py script does this. The ensembled forecasts are written to the directory results/ensemble_rnn_forecasts.
Example: python ensembling_forecasts.py --dataset_name nn5_adam
The SMAPE and MASE errors are calculated per each series for each model using the error calcualtion scripts in the directory error_calculator. The script perform the post processing of the forecasts to reverse initial preprocessing. The errors of the ensembles are written to the directory results/ensemble_errors.
Example: Rscript error_calculator/moving_window/final_evaluation.R results/ensemble_rnn_forecasts/nn5_ensemble /results/ensemble_errors/ /results/ensemble_processed_rnn_forecasts/ nn5_ensemble_error datasets/text_data/NN5/moving_window/nn5_test12i15.txt datasets/text_data/NN5/nn5_test.txt datasets/text_data/NN5/nn5_train.txt 70 56 0 0 1 7 0
For datasets that have different clusters such as the M3, M4 and CIF 2016, the utility_scripts/error_summary_scripts/cluster_results_merger.py script merges the results from different clusters into one file.
The utility_scripts/error_summary_scripts/error_summary_generator.py script creates the error summaries given the name of the dataset. Calculates the mean, median and ranked error values.
The results from our experiments are available as csv files in the Google Drive folder at this link
When using this repository, please cite:
@article{HEWAMALAGE2020,
title = "Recurrent Neural Networks for Time Series Forecasting: Current status and future directions",
journal = "International Journal of Forecasting",
year = "2020",
issn = "0169-2070",
doi = "https://doi.org/10.1016/j.ijforecast.2020.06.008",
url = "http://www.sciencedirect.com/science/article/pii/S0169207020300996",
author = "Hansika Hewamalage and Christoph Bergmeir and Kasun Bandara",
keywords = "Big data, Forecasting, Best practices, Framework",
abstract = "Recurrent Neural Networks (RNNs) have become competitive forecasting methods, as most notably shown in the winning method of the recent M4 competition. However, established statistical models such as exponential smoothing (ETS) and the autoregressive integrated moving average (ARIMA) gain their popularity not only from their high accuracy, but also because they are suitable for non-expert users in that they are robust, efficient, and automatic. In these areas, RNNs have still a long way to go. We present an extensive empirical study and an open-source software framework of existing RNN architectures for forecasting, and we develop guidelines and best practices for their use. For example, we conclude that RNNs are capable of modelling seasonality directly if the series in the dataset possess homogeneous seasonal patterns; otherwise, we recommend a deseasonalisation step. Comparisons against ETS and ARIMA demonstrate that (semi-) automatic RNN models are not silver bullets, but they are nevertheless competitive alternatives in many situations."
}