febama: Feature-based Bayesian Forecasting Model Averaging

Notice to users:

A better maitained R version is now available at coauthor's github https://github.com/feng-li/febama

Python Port

The FEBAMA workflow has been ported to Python as gsm.febama, a submodule of the general smooth-mixture package gsm. The Python port keeps the FEBAMA idea of feature-driven Bayesian forecast averaging, but places it inside the broader GSM/MoE codebase with JAX-based scoring, standard Python data tooling, and pluggable predictive distributions.

New Python development should target the gsm.febama module. This R package remains the original implementation and a reference for API behavior, S&P 500 examples, feature construction, and paper replication checks.

In the Python gsm repository, install and run the current FEBAMA example with:

python -m pip install -e ".[dev,febama]"
python scripts/run_febama_example.py --max-origins 4 --test-size 1 --max-iter 100

Short Introduction

In this work, we propose a novel framework for density forecast combination by constructing time-varying weights based on time series features, which is called FEature-based BAyesian forecasting Model Averaging (FEBAMA). Our framework estimates weights in the forecast combination via Bayesian log predictive scores, in which the optimal forecasting combination is determined by time-series features from historical information. In particular, we use an automatic Bayesian variable selection method to weight the importance of different features. To this end, our approach has better interpretability compared to other black-box forecasting combination schemes.

Installation

You can install the package febama from GitHub Repository with:

devtools::install_github("lily940703/febama")

Usage

This part explains how to generate forecasts based on FEBAMA framework.

Packages we need

library(febama)
library("tsfeatures")
library("M4metalearning")
library("forecast")
library("mvtnorm")
library(parallel)
library(doParallel)
library(foreach)

Example data

# M3 monthly
library(Mcomp)
data_m3_mon <- Filter(function(l) l$period == "MONTHLY", M3)

# Pick a time series randomly
set.seed(2021-7-29)
id = sample(1:length(data_m3_mon), 1)
data_example = data_m3_mon[[id]]

Training period

# 1. Training period
# 1.1 Compute predictive densities and features
model_conf_default = model_conf_default()
lpd_features = lpd_features_multi(data = data_example, model_conf = model_conf_default)

# If parallel
model_conf_curr = model_conf_default
model_conf_curr$lpd_features_parl = list(par = T, ncores = 2)
lpd_features = lpd_features_multi(data = data_example, model_conf = model_conf_curr)

# The features used
lpd_features <- feature_clean(list(lpd_features))[[1]]
fe <- lpd_features$feat
fm <- lpd_features$feat_mean
fs <- lpd_features$feat_sd
lpd_features$feat<- fe[, unique(unlist(model_conf_curr$features))]
lpd_features$feat_mean <- fm[unique(unlist(model_conf_curr$features))]
lpd_features$feat_sd <- fs[unique(unlist(model_conf_curr$features))]

# Up to now, we obtain log predictive densities and features for training.
head(lpd_features$lpd)
head(lpd_features$feat)

# 1.2 Inference procedure
# (1) For FEBAMA
parameters = febama_mcmc(data = lpd_features, model_conf = model_conf_curr)
# (2) For FEBAMA+VS
model_conf_curr$algArgs$nIter = 50
parameters_vs = febama_mcmc(data = lpd_features, model_conf = model_conf_curr)

Forecasting period

# 2. Forecasting period
# (1) For FEBAMA
model_conf_curr$algArgs$nIter = 1
data_example_fore = forecast_feature_results_multi(ts = data_example, model_conf = model_conf_curr,
                                                   data = lpd_features, beta_out = parameters)

cat("The average LS and MASE of the example data based on FEBAMA method are \n", data_example_fore$err_feature[1]/18, data_example_fore$err_feature[2])

# (2) For FEBAMA+VS
model_conf_curr$algArgs$nIter = 50
data_example_fore_vs = forecast_feature_results_multi(ts = data_example, model_conf = model_conf_curr,
                               data = lpd_features, beta_out = parameters_vs)
cat("The average LS and MASE of the example data based on FEBAMA+VS method are \n", data_example_fore_vs$err_feature[1]/18,data_example_fore_vs$err_feature[2])