Convolutional Models for M4 time series forecasting

About the Project

This project is based on the M4 Time Series Forecasting Competition 'btrotta' entry. In particular I trained and run her convolutional neural networks using her instructions in her repo : (https://github.com/btrotta/m4).

Built With

The code and models were built using :

• python
• Keras
• Tensorflow

(back to top)

Prerequisites

The models were tested with :

python==3.8.10
tensorboard==2.13.0
tensorboard-data-server==0.7.1
tensorflow==2.13.0
tensorflow-estimator==2.13.0
tensorflow-io-gcs-filesystem==0.32.0
scikit-learn==1.2.2
scipy==1.10.1
sklearn==0.0.post1
pandas==2.0.3
numpy==1.24.2
matplotlib==3.7.1
matplotlib-inline==0.1.6

In a clean python virtual environment you can just run pip install <The above list in a txt file> to install them altogether.

Regarding Tensorflow in particular the models were implemented using V1 compatibility mode.

Usage

To download the data you can visit (https://github.com/Mcompetitions/M4-methods/tree/master/Dataset/Test) and (https://github.com/Mcompetitions/M4-methods/tree/master/Dataset/Train) to download the train and test data. Next Move them into a directory called data.

1. Run shuffle.py to shuffle the series in each frequency and get the resulting shuffled .csv to use for training.
2. To build, train the models and get the metric results run main.py. The original entry evaluation metrics are being used (smape/mase). This will generate a log file with the metric results, as well as train the models. The trained models will be saved in a directory called trained_models and each one is named with the convention frequency_training length_horizon step.h5.For example the model Yearly_length_10_step_0.h5 is the model that was trained on the yearly series dataset with training length of 10 to predict the first horizon time step.
3. After having trained the models you can make predictions and receive the .csv with the predictions for each frequency by running predict.py. This will read the trained models and make predictions for all the time series of each frequency. It removes the final horizon values of each frequency's dataset and using the history before those values and the training length given , it predicts those horizon values. The predictions will be saved in the predictions directory with a different directory for each frequency.
4. If you want to make the visualizations you must download the directory tree original_trained_models/results in your current directory since the scripts read the original author's reproduced results from this path. Alternatively create the same directory tree and copy the content inside. Then you can make visualizations of the original models and the modified models by running compare_predictions.py which will read the first series from each frequency from the *fc*.csv results and visualize the predictions.

(back to top)

Experimentation/Improving the Models

The models created by the author and the models i created are both single step models. They are trained in a history of observations in the time series that are being given and each one predicts a single time step in the future. I experimented with the architecture of the convolutional models created and their parameters and i tried to improve the models for each frequency . Some functions used are taken from the original (https://github.com/btrotta/m4) and others were added. In particular besides the new models, functions to observe the training loss and visualize the final predictions are implemented, whereas the final predictions are located in the 'predictions' directory and are split according to frequency.

Models

If you want to take a look at the models or make modifications and update them see build_models.py where each model for each frequency is created. Additionally you can change model parameters like series_length, yearly_count, filter_count, units_count, epochs, bs from the train_models.py in the utils module, in the line where each model is constructed.

Regarding Visualizations

To make predictions I used the provided test dataset as the ground truth since we have a number of observations equal to the horizon provided for each frequency and after having trained the author's original designed models i plot the predictions from them aligned with the predictions from our models. For each frequency i picked the first series so this can not be considered representative but i include it just for visualization purposes.