Overview
The stackr package provides an easy way to combine predictions from individual time series or panel data models to an ensemble. stackr stacks models according to the Continuous Ranked Probability Score (CRPS) over k-step ahead predictions. It is therefore especially suited for time-series and panel data. A function for leave-one-out CRPS may be added in the future. Predictions need to be predictive distributions represented by predictive samples. Usually, these will be sets of posterior predictive simulation draws generated by an MCMC algorithm.
CRPS Stacking
Given some training data with true observed values as well as predictive samples generated from different models, stackr finds the optimal (in the sense of minimizing expected cross-validation predictive error) weights to form an ensemble of these models. Using these weights, stackr can then provide samples from the optimal model mixture by drawing from the predictive samples of those models in the correct proportion. This gives a mixture model solely based on predictive samples and is in this regard superior to other ensembling techniques like Bayesian Model Averaging. More information can be found in the package vignette.
Weights are generated using the crps_weights function. With these weights and predictive samples, the mixture_from_samples function can be used to obtain predictive samples from the optimal mixture model.
Usage
Load example data and split into train and test data
splitdate <- as.Date("2020-03-28")
traindata <- example_data[date <= splitdate]
testdata <- example_data[date > splitdate]Get weights and create mixture
weights <- crps_weights(traindata)
test_mixture <- mixture_from_samples(testdata, weights = weights)Score predictions
library("scoringutils")
# combine data.frame with mixture with predictions from other models
score_df <- rbindlist(list(testdata, test_mixture), fill = TRUE)
# score all predictions using from github.com/epiforecasts/scoringutils
score_df[, crps := crps(unique(observed), t(predicted)),
by = .(geography, model, date)
]
# summarise scores
score_df[, mean(crps), by = model][, setnames(.SD, "V1", "CRPS")]References
- Using Stacking to Average Bayesian Predictive Distributions, Yuling Yao, Aki Vehtari, Daniel Simpson, and Andrew Gelman, 2018, Bayesian Analysis 13, Number 3, pp. 917–1003 DOI 10.1214/17-BA1091
- Strictly Proper Scoring Rules, Prediction, and Estimation, Tilmann Gneiting and Adrian E. Raftery, 2007, Journal of the American Statistical Association, Volume 102, 2007 - Issue 477 DOI 10.1198/016214506000001437
- Comparing Bayes Model Averaging and Stacking When Model Approximation Error Cannot be Ignored, Bertrand Clarke, 2003, Journal of Machine Learning Research 4
- Bayesian Model Weighting: The Many Faces of Model Averaging, Marvin Höge, Anneli Guthke and Wolfgang Nowak, 2020, Water, DOI 10.3390/w12020309
- Bayesian Stacking and Pseudo-BMA weights using the loo package, Aki Vehtari and Jonah Gabry, 2019, https://mc-stan.org/loo/articles/loo2-weights.html
Contributors
All contributions to this project are gratefully acknowledged using the allcontributors package following the all-contributors specification. Contributions of any kind are welcome!
