CEEMDAN_LSTM

CEEMDAN_LSTM

GitHub: https://github.com/FateMurphy/CEEMDAN_LSTM
Future work: CFS

Background

CEEMDAN_LSTM is a Python module for decomposition-integration forecasting models based on EMD methods and LSTM. It aims at helping beginners quickly make a decomposition-integration forecasting by CEEMDAN, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (Torres et al. 2011), and LSTM, Long Short-Term Memory recurrent neural network (Hochreiter and Schmidhuber, 1997). If you use or refer to the content of this module, please cite the paper: (F. Zhou, Z. Huang, C. Zhang, Carbon price forecasting based on CEEMDAN and LSTM, Applied Energy, 2022, Volume 311, 118601, ISSN 0306-2619.).

Flowchart

Note, as it decomposes the entire series first, there is some look-ahead bias.

Install

(1) PyPi (recommended)

The quickest way to install the package is through pip.

pip install CEEMDAN_LSTM

(2) From the package

Download the package CEEMDAN_LSTM-1.2.tar.gz by clicking Code -> Download ZIP. After unzipping, move the package where you like.

pip install .(your file path)/CEEMDAN_LSTM-1.2.tar.gz

(3) From source

If you want to modify the code, you should download the code and build the package yourself. The source is publically available and hosted on GitHub: https://github.com/FateMurphy/CEEMDAN_LSTM. To download the code, you can either go to the source code page and click Code -> Download ZIP, or use the git command line.
After modifying the code, you can install the modified package by using the command line:

python setup.py install

Or, you can link to the path for convenient modification, eg. sys.path.append(.your file path/), and then import.

Import and quickly predict

import CEEMDAN_LSTM as cl
cl.quick_keras_predict(data=None) # default dataset: sse_index.csv

Load dataset

data = cl.load_dataset() # some built-in dataset eg. sp500.csv hsi.csv ftse.csv nsdaq.csv n225.csv
# data = pd.read_csv(your_file_path + its_name + '.csv', header=0, index_col=['date'], parse_dates=['date'])

Help and example

You can use the code to call for help. You can copy the code from the output of cl.show_keras_example() to run forecasting and help you learn more about the code.

cl.help()
cl.show_keras_example()
cl.show_keras_example_model()
cl.details_keras_predict(data=None)

Start to Forecast

Take Class: keras_predictor() as an example.

Brief summary and forecast

data = cl.load_dataset()
series = data['close'] # choose a DataFrame column 
cl.statis_tests(series)
kr = cl.keras_predictor()
df_result = kr.hybrid_keras_predict(data=series, show=True, plot=True, save=True)

0. Statistical tests (not necessary)

The code will output the results of the ADF test, Ljung-Box Test, and Jarque-Bera Test, and plot ACF and PACF figures to evaluate stationarity, autocorrelation, and normality.

cl.statis_tests(series=None)

1. Declare the parameters

Note, when declaring the PATH, folders will be created automatically, including the figure and log folders.

kr = cl.keras_predictor(PATH=None, FORECAST_HORIZONS=30, FORECAST_LENGTH=30, KERAS_MODEL='GRU', 
                        DECOM_MODE='CEEMDAN', INTE_LIST='auto', REDECOM_LIST={'co-imf0':'ovmd'},
                        NEXT_DAY=False, DAY_AHEAD=1, NOR_METHOD='minmax', FIT_METHOD='add', 
                        USE_TPU=False, **kwargs))

2. Forecast

You can try the following forecasting methods. Note, kr. is the class defined in step 1, necessary for the code.

df_result = kr.single_keras_predict(data, show=True, plot=True, save=True)
# df_result = kr.ensemble_keras_predict(data, show=True, plot=True, save=True)
# df_result = kr.respective_keras_predict(data, show=True, plot=True, save=True)
# df_result = kr.hybrid_keras_predict(data, show=True, plot=True, save=True)
# df_result = kr.multiple_keras_predict(data, show=True, plot=True, save=True)

3. Validate

(1) Plot heatmap

You need to install seaborn first, and the input should be 2D-array.

cl.plot_heatmap(data, corr_method='pearson', fig_path=None)

(2) Diebold-Mariano-Test (DM test)

DM test will output the DM test statistics and its p-value. You can refer to https://github.com/johntwk/Diebold-Mariano-Test.

rt = cl.dm_test(actual_lst, pred1_lst, pred2_lst, h=1, crit="MSE", power=2)

4. Next-day Forecast

Set NEXT_DAY=True.

kr = cl.keras_predictor(NEXT_DAY=True)
df_result = kr.hybrid_keras_predict(data, show=True, plot=True, save=True)
# df_result = kr.rolling_keras_predict(data, predict_method='single')

Sklearn Forecast

You can try the following forecasting methods. Note, sr. is the defined class, necessary for the code.

# SKLEARN_MODEL = LASSO, SVM, or LGB(LightGBM); OPTIMIZER = Bayes, GS(GridSearch)
sr = cl.sklearn_predictor(PATH=path, FORECAST_HORIZONS=30, FORECAST_LENGTH=30,
                          SKLEARN_MODEL='LASSO', OPTIMIZER='Bayes',
                          DECOM_MODE='OVMD', INTE_LIST='auto')
df_result = sr.single_sklearn_predict(data, show=True, plot=True, save=True)
# df_result = sr.respective_sklearn_predict(data, show=True, plot=True, save=True)
# df_result = sr.multiple_sklearn_predict(series_close, run_times=10, predict_method='single')

Discussion

1. Look-ahead bias

As the predictor will decompose the entire series first before splitting the training and test set, there is a look-ahead bias. It is still an issue about how to avoid the look-ahead bias.

2. VMD decompose

The vmdpy module can only decompose the even-numbered length time series. When forecasting an odd-numbered length one, this module will delete the oldest data point. It is still an issue how to modify VMD decomposition. Moreover, selecting the K parameters is important for the VMD method, and hence, I will add some methods to choose a suitable K, such as OVMD, REI, SampEn, and so on.

3. Rolling forecasting

Rolling forecasting costs a lot of time. Like a 30-forecast-length prediction, it will run 30 times cl.hybrid_keras_predict(), so I am not sure if it is really effective or not.