Forecasting with XGBoost embedded in Quasi-Randomized Neural Networks

This article was first published on T. Moudiki's Webpage - Python , and kindly contributed to python-bloggers. (You can report issue about the content on this page here)
Want to share your content on python-bloggers? click here.

Next week, I’ll present nnetsauce’s (univariate and multivariate probabilistic) time series forecasting capabilities at the 44th International Symposium on Forecasting (ISF) (ISF) 2024. ISF is the premier forecasting conference, attracting the world’s leading forecasting (I don’t only do forecasting though) researchers, practitioners, and students. I hope to see you there!

In this post, I illustrate how to obtain predictive simulations with nnetsauce’s MTS class using XGBoost as base learner, and I give some intuition behind the method "kde" employed for uncertainty quantification in this case. The time series used here is the USAccDeaths dataset, a univariate time series of monthly deaths from motor vehicle traffic in the United States from 1973 to 1978.

(Command line)

!pip install nnetsauce --upgrade --no-cache-dir

Import Python packages

import nnetsauce as ns
import numpy as np
import pandas as pd
import xgboost as xgb
import matplotlib.pyplot as plt
import seaborn as sns

# import data 
url = ""
df = pd.read_csv(url)
df.index = pd.DatetimeIndex(
df.drop(columns=['date'], inplace=True)

Adjusting XGBoost regressors with different number of estimators

# number of estimators for the base learner
n_estimators_list = [5, 20, 50, 100]
estimators = []
residuals = []

for n_estimators in n_estimators_list:
  # XGBoost regressor as base learner
  regr_xgb = ns.MTS(obj=xgb.XGBRegressor(n_estimators=n_estimators,
  # in sample residuals 
  # out-of-sample predictions

residuals_df = pd.DataFrame(np.asarray(residuals).T,
                            columns=["n5", "n20", "n50", "n100"])
100%|██████████| 1/1 [00:00<00:00,  4.55it/s]
100%|██████████| 1000/1000 [00:00<00:00, 1797.56it/s]
100%|██████████| 1000/1000 [00:00<00:00, 1583.30it/s]
100%|██████████| 1/1 [00:00<00:00,  2.30it/s]
100%|██████████| 1000/1000 [00:01<00:00, 582.86it/s]
100%|██████████| 1000/1000 [00:01<00:00, 882.61it/s]
100%|██████████| 1/1 [00:01<00:00,  1.83s/it]
100%|██████████| 1000/1000 [00:01<00:00, 808.35it/s]
100%|██████████| 1000/1000 [00:00<00:00, 1555.30it/s]
100%|██████████| 1/1 [00:02<00:00,  2.37s/it]
100%|██████████| 1000/1000 [00:01<00:00, 578.83it/s]
100%|██████████| 1000/1000 [00:00<00:00, 1436.02it/s]

for est in estimators:


for i in range(4):
  sns.kdeplot(residuals_df.iloc[:,i], fill=True, color="red")


In order to obtain predictive simulations with "kde" method (and as seen last week in #143), a Kernel Density Estimator (KDE) is adjusted to in-sample residuals. The most intuitive piece I found on KDEs is the following presentation:

When using a high number of estimators (with the other parameters kept constant), the XGBRegressor base learner will overfit the training set, so that the in-sample residuals will be very small, and the uncertainty can’t be captured/estimated adequately: the predictions will consist of point forecasts. A compromise needs to be found by using cross-validation on the base learner’s hyperparameters, with an uncertainty quantification metric. Other types of predictions intervals/predictive simulation methods will be available in future versions of nnetsauce.

To leave a comment for the author, please follow the link and comment on their blog: T. Moudiki's Webpage - Python .

Want to share your content on python-bloggers? click here.