Time Series Forecasting Package

This is a python package containing Various Forecasting Algorithms, Forecasting Datasets and, Plotting, Preprocessing and Utility Tools.

Forecasting Algorithms

This package provides (or would provide) various algorithms which work on data containing multivariate target time series, univariate target time series - both with or without exogenous time series. The algorithms are provided in the model subpackage of the time series package ts.

Currently, this package provides the following Forecasting Algorithms,

MLP (DNN) based multivariate forecasting algorithm which has support for exogenous time series. You may import this in your application by doing as follows

from ts.model import DeepNN

RNN based multivariate forecasting algorithm with support for multivariate exogenous time series. This is a recurrent model which takes in any recurrent layer (and parameters) and stacks it required number of times.

from ts.model import RnnForecast

Simple RNN based multivariate forecasting algorithm with support for multivariate exogenous time series. This model is built by stacking multiple simple RNN layers.

from ts.model import SimpleRnnForecast

GRU based multivariate forecasting algorithm with support for multivariate exogenous time series. This model is built by stacking multiple GRU layers.

from ts.model import GruForecast

LSTM based multivariate forecasting algorithm with support for multivariate exogenous time series. This model is built by stacking multiple LSTM layers.

from ts.model import LstmForecast

The Extreme Time Model - which focuses on forecasting target series with extreme values - i.e. values which have a very large deviation from the time series trend. It supports univariate time series and multivariate exogenous time series.

from ts.model import ExtremeTime

The Extreme Time Model 2 - It is another model focussed on forecasting time series with extreme values. It supports univariate time series and multivariate exogenous time series.

from ts.model import ExtremeTime2

A GMM-HMM based forecasting model which predicts by finding maximum posterior probability over a discrete set of observation values (since maximum over the contiuous observation space seems infeasible)

from ts.model import GmmHmmForecast

A GMM-HMM based forecasting model which predicts by finding the observation in the training set which is has closest log likelihood value to the current observation's log likelihood value

from ts.model import GmmHmmLikelihoodSimilarity

Forecasting Data

This package provides Data Generators as well as Datasets. The data subpackage of the time series package ts provides two subpackages named generate and dataset, the first one contains data generators and the second one contains real world datasets.

Currently, we provide the following,

ARMA Generated data - generates only univariate target series without support for exogenous series.

from ts.model.univariate.nonexo import ArmaGenerator

Standard Generators - provides generators for simple data, long term dependency data and extreme valued data

from ts.model.univariate.nonexo import StandardGenerator

Periodic Generator - generates periodic data but with support for only univariate target series without exogenous series.

from ts.model.univariate.nonexo import PeriodicGenerator

Polynomial Generator - generates data which is a polynomial function of time. This allows one to generate data with a varying trend.

from ts.model.univariate.nonexo import PolynomialGenerator

Difficult Generator - generates difficult data, i.e. data which has a varying trend, periodicity (seasonality) and noise (ARMA). One can introduce extreme values into the data by providing the appropriate contructor arguments

from ts.model.univariate.nonexo import DifficultGenerator

Plotting, Preprocessing and Utility Tools

This package contains plotting tools for plotting losses, plotting training data and comparing prediction with true (using plots). The plotting tools are available in the plot subpackage of the time series package ts. It also provides utility tools for in the utility subpackage of ts. Access to the global logger and local loggers is provided in the log subpackage of ts.

Package Structure

ts/
|__ data/
|    |__ dataset/
|         |__ AmazonStockPrice
|         |__ RetailSales
|         |__ JaipurWeather
|
|    |__ generate/
|         |__ univariate/
|              |__ nonexo/
|                  |__ ArmaGenerator
|                  |__ StandardGenerator
|                  |__ PeriodicGenerator
|                  |__ PolynomialGenerator
|                  |__ DifficultGenerator
| 
|__ model/
|    |__ RnnForecast
|    |__ SimpleRnnForecast
|    |__ GruForecast
|    |__ LstmForecast
|    |__ DeepNN
|    |__ ExtremeTime
|    |__ ExtremeTime2
|    |__ GmmHmmForecast
|    |__ GmmHmmLikelihoodSimilarity
|
|__ plot/
|    |__ Plot
|
|__ log/
|    |__ ConsoleLogger
|    |__ FileLogger
|    |__ GlobalLogger
|
|__ test/
|    |__ model/..
|    |__ utility/..
|
|__ utility/
|    |__ Utility
|    |__ ForecastDataSequence
|    |__ SaveCallback
|    |__ DatasetUtility
|    |__ Metric

Repository Structure

This repository is structured as follows:

Forecast
|__ other/...
|
|__ notebooks/..
|
|__ ts/...

other: contains deprecated codes and codes that do not work. The contents of this directory would be removed as soon as they have been analyzed thoroughly.
notebooks: contains notebooks containing experiments, tests and examples
ts: The time series forecasting package

Existing experiments, tests and example notebooks

The notebooks directory of this repository contains notebooks containing experiments, tests and examples. To be able to run these notebooks, we need to do as follows:

Go to your .bashrc file and add the following line:

export PYTHONPATH="$PYTHONPATH:<location_of_this_repository>"

where <location_of_this_repository> is the location of this repository in your filesystem.

tymefighter / Forecast