ForecastGA

A Python tool to forecast GA data using several popular time series models.

About

Welcome to ForecastGA

ForecastGA is a tool that combines a couple of popular libraries, Atspy and googleanalytics, with a few enhancements.

• The models are made more intuitive to upgrade and add by having the tool logic separate from the model training and prediction.
• When calling am.forecast_insample(), any kwargs included (e.g. learning_rate) are passed to the train method of the model.
• Google Analytics profiles are specified by simply passing the URL (e.g. https://analytics.google.com/analytics/web/?authuser=2#/report-home/aXXXXXwXXXXXpXXXXXX).
• You can provide a data dict with GA config options or a Pandas Series as the input data.
• Multiple log levels.
• Auto GPU detection (via Torch).
• List all available models, with descriptions, by calling forecastga.print_model_info().
• Google API info can be passed in the data dict or uploaded as a JSON file named identity.json.
• Created a companion Google Colab notebook to easily run on GPU.
• A handy plot function for Colab, forecastga.plot_colab(forecast_in, title="Insample Forecast", dark_mode=True) that formats nicely and also handles Dark Mode!

Models Available

• ARIMA : Automated ARIMA Modelling
• Prophet : Modeling Multiple Seasonality With Linear or Non-linear Growth
• ProphetBC : Prophet Model with Box-Cox transform of the data
• HWAAS : Exponential Smoothing With Additive Trend and Additive Seasonality
• HWAMS : Exponential Smoothing with Additive Trend and Multiplicative Seasonality
• NBEATS : Neural basis expansion analysis (now fixed at 20 Epochs)
• Gluonts : RNN-based Model (now fixed at 20 Epochs)
• TATS : Seasonal and Trend no Box Cox
• TBAT : Trend and Box Cox
• TBATS1 : Trend, Seasonal (one), and Box Cox
• TBATP1 : TBATS1 but Seasonal Inference is Hardcoded by Periodicity
• TBATS2 : TBATS1 With Two Seasonal Periods

How To Use

Find Model Info:

forecastga.print_model_info()

Initialize Model:

Google Analytics:

data = { 'client_id': '',
         'client_secret': '',
         'identity': '',
         'ga_start_date': '2018-01-01',
         'ga_end_date': '2019-12-31',
         'ga_metric': 'sessions',
         'ga_segment': 'organic traffic',
         'ga_url': 'https://analytics.google.com/analytics/web/?authuser=2#/report-home/aXXXXXwXXXXXpXXXXXX',
         'omit_values_over': 2000000
        }

model_list = ["TATS", "TBATS1", "TBATP1", "TBATS2", "ARIMA"]
am = forecastga.AutomatedModel(data , model_list=model_list, forecast_len=30 )

Pandas DataFrame:

# CSV with columns: Date and Sessions
df = pd.read_csv('ga_sessions.csv')
df.Date = pd.to_datetime(df.Date)
df = df.set_index("Date")
data = df.Sessions

model_list = ["TATS", "TBATS1", "TBATP1", "TBATS2", "ARIMA"]
am = forecastga.AutomatedModel(data , model_list=model_list, forecast_len=30 )

Forecast Insample:

forecast_in, performance = am.forecast_insample()

Forecast Outsample:

forecast_out = am.forecast_outsample()

Ensemble Performance:

all_ensemble_in, all_ensemble_out, all_performance = am.ensemble(forecast_in, forecast_out)

Pretty Plot in Google Colab

forecastga.plot_colab(forecast_in, title="Insample Forecast", dark_mode=True)

Installation

Windows users may need to manually install the two items below via conda :

1. conda install pystan
2. conda install pytorch -c pytorch
3. !pip install --upgrade git+https://github.com/jroakes/ForecastGA.git

otherwise, pip install --upgrade forecastga

This repo support GPU training. Below are a few libraries that may have to be manually installed to support.

pip install --upgrade mxnet-cu101
pip install --upgrade torch 1.7.0+cu101

Acknowledgements

1. Majority of forecasting code taken from https://github.com/firmai/atspy and refactored heavily.
2. Google Analytics based off of: https://github.com/debrouwere/google-analytics
3. Thanks to richardfergie for the addition of the Prophet Box-Cox model to control negative predictions.

Contribute

The goal of this repo is to grow the list of available models to test. If you would like to contribute one please read on. Feel free to have fun naming your models.

1. Fork the repo.
2. In the /src/forecastga/models folder there is a model called template.py. You can use this as a template for creating your new model. All available variables are there. Forecastga ensures each model has the right data and calls only the train and forecast methods for each model. Feel free to add additional methods that your model requires.
3. Edit the /src/forecastga/models/__init__.py file to add your model's information. Follow the format of the other entries. Forecastga relies on loc to find the model and class to find the class to use.
4. Edit requirments.txt with any additional libraries needed to run your model. Keep in mind that this repo should support GPU training if available and some libraries have separate GPU-enabled versions.
5. Issue a pull request.

If you enjoyed this tool consider buying me some beer at: Paypalme