descartesresearch / telescope

Telescope is a hybrid multi-step-ahead forecasting approach based on time series decomposition.

Telescope

Telescope is a machine learningbased forecasting approach that automatically retrieves relevant information from a given time series..

Details of the Telescope approach can be found at [1,2]. The details of the recommendation approach can be found at [3].

Installation

This package can be installed in R by using the following commands:

install.packages("devtools")
devtools::install_github("DescartesResearch/telescope")

For unknown reasons, install_github does not work under all Windows versions. Therefore the package can alternatively be installed in R with the following commands:

install.packages("remotes")
remotes::install_url(url="https://github.com/DescartesResearch/telescope/archive/master.zip", INSTALL_opt= "--no-multiarch")

Example without Recommendition system

library(telescope)
forecast <- telescope.forecast(forecast::taylor, horizon = 1000)

Example with Recommendition system

install.packages('Mcomp')
library(Mcomp)
library(telescope)

ts.list <- list()
for(i in 1:length(M3)){
  ts.list[i] <- list(ts(c(M3[[i]]$x,M3[[i]]$xx),frequency = frequency(M3[[i]]$x)))` <br />
}

model <- telescope.trainrecommender(ts.list)
telescope.forecast(forecast::taylor, horizon = 1000, rec_model = model)

Further Information and References

For more information on this forecasting method, please visit our homepage.

[1] Bauer, A., Züfle, M., Herbst, N., Kounev, S. & Curtef, V. (2020). Telescope: An Automatic Feature Extraction and Transformation Approach for Time Series Forecasting on a Level-Playing Field. Proceedings of the 36th International Conference on Data Engineering (ICDE) (p./pp. 1902-1905).

[2] Bauer, A., Züfle, M., Herbst, N., Zehe, A., Hotho, A. & Kounev, S. (2020). Time Series Forecasting for Self-Aware Systems. Proceedings of the IEEE.

[3] Bauer, A., Züfle, M., Grohmann, J., Schmitt, N., Herbst, N. & Kounev, S. (2020). An Automated Forecasting Framework based on Method Recommendation for Seasonal Time Series. Proceedings of the 2020 ACM/SPEC International Conference on Performance Engineering (p./pp. 48-55), April, New York, NY, USA: ACM.

telescope's Issues

Add plot also for time series with frequency == 1

Add policy for NaNs

Fix Error "Error in auto.arima(ts, stepwise = TRUE, seasonal = season)"

Change naming in XGB.R

Flag for Plotting

Introduce a flag parameter indicating whether plots of the forecast are wanted or not

Add recommendation system for choosing the best ML method based on the given time series

Plots

Remove double history in forecast plot

Add Fourier terms

Remove categorical information

RMSE fitting_models

Mean and sqrt for RMSE calculation in fitting_models are switched

Fix bug in doANN

Uncaught error during forecast of simple time series

Code:

library(telescope)
inp <- c(1, 0, 0, 0, 0, 0, 0)
outp <- telescope.forecast(inp, horizon=1)

Produced error:

Error in kernapply.vector(x, k, circular = circular) : 'x' is shorter than kernel 'k'
Calls: telescope.forecast ... spec.pgram -> kernapply -> kernapply.default -> kernapply.vector

Same error generated for the input c(1,0,0,0,1,0,0). I would be thankful for any help or hint for any input restrictions to telescope.

Train recommender with the same data fitted to the prediction

Thank you for sharing your insights and codes. I am trying to apply this framework in financial time series and wonder if it is valid to use the same data used in training the forecast model to train the recommender.

Add prediction/confidence intervals

Improve Telescope for "trendy" time series

Spilt telescope in model and forecast methods