README.md

In snarf-snarf/karma: Complementary tools for automated time-series machine learning.

karma

A library for automated time series prediction

Statistical forecasting with automated model selection and hyperparameter optimization - One-line ARIMA selection and fitting with both stochastic and deterministic training options (meant as an alternative to auto.arima) - One-line Box-Jenkins model constuction with training options for approximation metaheuristic

Manual time series analysis - Backwards compatibility with package 'forecast' (for Hyndman's ARIMA, neural network autoregression, and exponential smoothing) - In-sample and out-of-sample model validation (diagnose overfitting/underfitting via train-test splitting)

AutoML for univariate series - Train multiple weak learners to make stronger predictions (ensemble learning / stacking) - Train all models in 'karma' and 'forecast' on a single time series dataset - Train all models in 'karma' and 'forecast' on multiple time series datasets and benchmark their results

Model selection algorithms

• MS-ARIMA (Markov Selection ARIMA)
• RW-ARIMA (Random Walk ARIMA)
• LD-ARIMA (Local Descent ARIMA)
• Nonparametric Markov Selection ARIMA (NMS-ARIMA)
• Stochastic Box-Jenkins (SBJ-ARIMA)
• Stochastic Multiseasonal Box-Jenkins (SMBJ-ARIMA)
• NP-Hard ARIMA (NPH-ARIMA)
• Stacked Stochastic Forecast (SSF)

Instructions

You first install these packages:

install.packages(c('devtools', 'forecast', 'stringr'))

Then you load devtools:

library(devtools)

Then you install this package:

install_github("snarf-snarf/karma")

Then you load all packages:

library(karma)

library(forecast)

library(stringr)

Examples:

Search for optimal SARIMA model using auto.karma with cross-validation:

sfit <- auto.karma( mdeaths )   # train SARIMA/ARIMA model

Retrain selected model (training set) and validate prediction accuracy on held out data (test set):

karma.cv( sfit )   # will plot predicted vs. actual test-set data and print out test MAPE

Forecast future periods:

plot(forecast( sfit ))    # plot projection

Use karma API with 'forecast' package:

afit <- auto.arima(mdeaths)   # auto.arima
karma.cv( afit )
plot(forecast( afit ))    
nfit <- nnetar(mdeaths)       # shallow neural network autoregression
karma.cv( nfit )
plot(forecast( nfit ))   
efit <- ets(mdeaths)          # exponential smoothing state space model
karma.cv( efit )
plot(forecast( efit ))    
                              # ...etc...

Train multiple models on a single series:

kmodels <- magic.karma(mdeaths)    # train all models
karma.cv(kmodels[[1]]$fit_obj)     # validate 1st model
karma.cv(kmodels[[2]]$fit_obj)     # validate 2nd model
                                   # ...etc...

Or you can follow the examples on this script:

https://github.com/snarf-snarf/karma/blob/master/karma_demo.r

Use help() on any function to see a more detailed documentation.

snarf-snarf/karma documentation built on May 24, 2019, 7:19 a.m.