Today, give a try to Techtonique web app, a tool designed to help you make informed, data-driven decisions using Mathematics, Statistics, Machine Learning, and Data Visualization. Here is a tutorial with audio, video, code, and slides: https://moudiki2.gumroad.com/l/nrhgb. 100 API requests are now (and forever) offered to every user every month, no matter the pricing tier.

Follow me on LinkedIn

Loading views...

Introduction

In this post, we’ll explore an intriguing comparison between XGBoost’s balanced accuracy with its default hyperparameters on 72 data sets and a GenericBooster with Linear Regression as a base learner, also with its default hyperparameters. Whereas the first one clearly overfits when not tuned (to an extent that I wanted to verify), but gives an overall higher test set balanced accuracy, the second one gives highly correlated training and test set errors, with an overall lower test set balanced accuracy.

What We’ll Cover

Setting up the environment with necessary packages
Loading and preparing the data
Comparing model performances
Analyzing overfitting patterns
Drawing conclusions about which approach might be better for specific scenarios

Let’s dive into the implementation and analysis.

!pip install genbooster
!pip install nnetsauce

Import the necessary packages.

import xgboost as xgb
import joblib
import requests
import io

from genbooster.genboosterclassifier import BoosterClassifier
from genbooster.randombagclassifier import RandomBagClassifier
from sklearn.tree import DecisionTreeRegressor
from sklearn.ensemble import BaggingRegressor
from sklearn.linear_model import Ridge, LinearRegression
from sklearn.gaussian_process import GaussianProcessRegressor
from tqdm import tqdm
from scipy import stats
from sklearn.model_selection import train_test_split
from sklearn.metrics import balanced_accuracy_score
from sklearn.model_selection import cross_val_score
from sklearn.metrics import make_scorer
from time import time

Obtain 72 data sets, a reduced form of the openml-cc18, available at https://github.com/thierrymoudiki/openml-cc18-reduced.

# Fetch the file content from the URL - **Use the raw file URL!**
url = 'https://github.com/thierrymoudiki/openml-cc18-reduced/raw/main/openml-cc18-Xys-2024-05-20.pkl'

# Use a session to handle potential connection interruptions
session = requests.Session()
response = session.get(url, stream=True)
response.raise_for_status()  # Raise an exception for bad responses

# Load the data from the downloaded content in chunks
with io.BytesIO() as buffer:
    for chunk in response.iter_content(chunk_size=1024*1024):  # Download in 1MB chunks
        buffer.write(chunk)
    buffer.seek(0)  # Reset buffer position to the beginning
    clf_datasets = joblib.load(buffer)

1 - xgboost results

results = {}

balanced_accuracy_scorer = make_scorer(balanced_accuracy_score)

for i, dataset in tqdm(enumerate(clf_datasets.items())):
  dataset_name = dataset[0]
  print("\n ----------", (i + 1), "/", len(clf_datasets.items()), ": ", dataset_name)
  try:
    X, y = dataset[1]['dataset'][0], dataset[1]['dataset'][1]
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)
    # Split dataset into training and testing sets
    model = xgb.XGBClassifier()
    model.fit(X_train, y_train)
    y_pred_train = model.predict(X_train)
    y_pred = model.predict(X_test)
    # Calculate balanced accuracy
    balanced_acc_train = balanced_accuracy_score(y_train, y_pred_train)
    balanced_acc_test = balanced_accuracy_score(y_test, y_pred)
    print("Training Balanced Accuracy:", balanced_acc_train)
    print("Testing Balanced Accuracy:", balanced_acc_test)
    results[dataset_name] = {"training_score": balanced_acc_train,
                             "testing_score": balanced_acc_test}
  except:
    print("Error")

0it [00:00, ?it/s]

 ---------- 1 / 72 :  kr-vs-kp

1it [00:00,  4.01it/s]

Training Balanced Accuracy: 0.9793965874420882
Testing Balanced Accuracy: 0.975187969924812

 ---------- 2 / 72 :  letter

2it [00:03,  2.19s/it]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7513736263736264

 ---------- 3 / 72 :  balance-scale

3it [00:04,  1.52s/it]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7260334744908249

 ---------- 4 / 72 :  mfeat-factors

4it [00:09,  2.97s/it]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.86

 ---------- 5 / 72 :  mfeat-fourier

5it [00:16,  4.19s/it]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.86

 ---------- 6 / 72 :  breast-w

6it [00:17,  3.20s/it]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9470108695652174

 ---------- 7 / 72 :  mfeat-karhunen

7it [00:22,  3.84s/it]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.905

 ---------- 8 / 72 :  mfeat-morphological

8it [00:22,  2.76s/it]

Training Balanced Accuracy: 0.9862499999999998
Testing Balanced Accuracy: 0.7950000000000002

 ---------- 9 / 72 :  mfeat-zernike

10it [00:23,  1.46s/it]

Training Balanced Accuracy: 0.99875
Testing Balanced Accuracy: 0.7799999999999999

 ---------- 10 / 72 :  cmc
Training Balanced Accuracy: 0.9798009974256786
Testing Balanced Accuracy: 0.6773695839418791

 ---------- 11 / 72 :  optdigits

13it [00:23,  1.66it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.905

 ---------- 12 / 72 :  credit-approval
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.8353204172876304

 ---------- 13 / 72 :  credit-g
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.6797619047619048

 ---------- 14 / 72 :  pendigits

16it [00:24,  3.01it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9240601503759398

 ---------- 15 / 72 :  diabetes
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7112962962962963

 ---------- 16 / 72 :  spambase
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9041740767862747

 ---------- 17 / 72 :  splice

18it [00:24,  4.30it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9529914529914528

 ---------- 18 / 72 :  tic-tac-toe
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9776119402985075

 ---------- 19 / 72 :  vehicle
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7570075757575757

 ---------- 20 / 72 :  electricity

21it [00:24,  5.32it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.778005115089514

 ---------- 21 / 72 :  satimage
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.8195766960114664

 ---------- 22 / 72 :  eucalyptus

22it [00:25,  5.05it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.6379919821780285

 ---------- 23 / 72 :  sick
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9583333333333333

 ---------- 24 / 72 :  vowel

24it [00:25,  4.56it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.898989898989899

 ---------- 25 / 72 :  isolet

25it [00:26,  2.19it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.6002747252747253

 ---------- 26 / 72 :  analcatdata_authorship
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9850446428571429

 ---------- 27 / 72 :  analcatdata_dmft

27it [00:27,  2.97it/s]

Training Balanced Accuracy: 0.5354622066211941
Testing Balanced Accuracy: 0.1808442851453604

 ---------- 28 / 72 :  mnist_784

30it [00:27,  3.67it/s]

Training Balanced Accuracy: 0.9974358974358974
Testing Balanced Accuracy: 0.5634728124659475

 ---------- 29 / 72 :  pc4
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7357954545454546

 ---------- 30 / 72 :  pc3
Training Balanced Accuracy: 0.9939024390243902
Testing Balanced Accuracy: 0.7138888888888889

 ---------- 31 / 72 :  jm1

33it [00:28,  5.62it/s]

Training Balanced Accuracy: 0.9894845464612907
Testing Balanced Accuracy: 0.6519350215002389

 ---------- 32 / 72 :  kc2
Training Balanced Accuracy: 0.9764705882352941
Testing Balanced Accuracy: 0.6002190580503833

 ---------- 33 / 72 :  kc1
Training Balanced Accuracy: 0.9660003848559195
Testing Balanced Accuracy: 0.7827829738499714

 ---------- 34 / 72 :  pc1

35it [00:28,  7.30it/s]

Training Balanced Accuracy: 0.9818181818181818
Testing Balanced Accuracy: 0.7803379416282642

 ---------- 35 / 72 :  adult
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7384868421052632

 ---------- 36 / 72 :  Bioresponse
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7790603891521323

 ---------- 37 / 72 :  wdbc

39it [00:28,  9.19it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.933531746031746

 ---------- 38 / 72 :  phoneme
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.8967423969227071

 ---------- 39 / 72 :  qsar-biodeg
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9139827179890023

 ---------- 40 / 72 :  wall-robot-navigation
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9885802469135803

 ---------- 41 / 72 :  semeion

43it [00:29,  8.21it/s]

Training Balanced Accuracy: 0.6431816659664762
Testing Balanced Accuracy: 0.63

 ---------- 42 / 72 :  ilpd
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.5405740609496811

 ---------- 43 / 72 :  madelon
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.845

 ---------- 44 / 72 :  nomao

45it [00:29,  9.50it/s]

Training Balanced Accuracy: 0.9890350877192983
Testing Balanced Accuracy: 0.9456508403876824

 ---------- 45 / 72 :  ozone-level-8hr
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7531879884821061

 ---------- 46 / 72 :  cnae-9
Training Balanced Accuracy: 0.7333304959709455
Testing Balanced Accuracy: 0.699604743083004

 ---------- 47 / 72 :  first-order-theorem-proving

50it [00:30,  7.28it/s]

Training Balanced Accuracy: 0.9974161661661661
Testing Balanced Accuracy: 0.44510582010582017

 ---------- 48 / 72 :  banknote-authentication
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 1.0

 ---------- 49 / 72 :  blood-transfusion-service-center
Training Balanced Accuracy: 0.8398196194712132
Testing Balanced Accuracy: 0.6754385964912281

 ---------- 50 / 72 :  PhishingWebsites
Training Balanced Accuracy: 0.9715260585285342
Testing Balanced Accuracy: 0.9516145358841988

 ---------- 51 / 72 :  cylinder-bands

52it [00:30,  8.60it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.8183730715287517

 ---------- 52 / 72 :  bank-marketing
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.6352247605011054

 ---------- 53 / 72 :  GesturePhaseSegmentationProcessed
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.473961038961039

 ---------- 54 / 72 :  har

54it [00:31,  4.44it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7054457054457054

 ---------- 55 / 72 :  dresses-sales
Training Balanced Accuracy: 0.9970238095238095
Testing Balanced Accuracy: 0.5472085385878489

 ---------- 56 / 72 :  texture

57it [00:31,  5.19it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9494949494949495

 ---------- 57 / 72 :  connect-4
Training Balanced Accuracy: 0.8687185238384995
Testing Balanced Accuracy: 0.490677451203767

 ---------- 58 / 72 :  MiceProtein

58it [00:32,  4.66it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9510996240601504

 ---------- 59 / 72 :  steel-plates-fault

59it [00:32,  3.88it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.8302055604850634

 ---------- 60 / 72 :  climate-model-simulation-crashes
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7222222222222222

 ---------- 61 / 72 :  wilt
Training Balanced Accuracy: 0.9880952380952381
Testing Balanced Accuracy: 0.9090909090909092

 ---------- 62 / 72 :  car

62it [00:32,  5.28it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9357142857142857

 ---------- 63 / 72 :  segment

63it [00:34,  1.90it/s]

Training Balanced Accuracy: 0.9974826446647592
Testing Balanced Accuracy: 0.9649894440534835

 ---------- 64 / 72 :  mfeat-pixel

64it [00:35,  2.15it/s]

Training Balanced Accuracy: 0.9012499999999999
Testing Balanced Accuracy: 0.7649999999999999

 ---------- 65 / 72 :  Fashion-MNIST

66it [00:36,  2.41it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.545

 ---------- 66 / 72 :  jungle_chess_2pcs_raw_endgame_complete
Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.7222113037136032

 ---------- 67 / 72 :  numerai28.6

67it [00:36,  2.87it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.5044004400440043

 ---------- 68 / 72 :  Devnagari-Script

68it [00:38,  1.32it/s]

Training Balanced Accuracy: 0.9909686700767264
Testing Balanced Accuracy: 0.2217391304347826

 ---------- 69 / 72 :  CIFAR_10

71it [00:39,  2.08it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.215

 ---------- 70 / 72 :  Internet-Advertisements
Training Balanced Accuracy: 0.967741724282422
Testing Balanced Accuracy: 0.9613787375415282

 ---------- 71 / 72 :  dna
Training Balanced Accuracy: 0.9147506693440429
Testing Balanced Accuracy: 0.9043803418803419

 ---------- 72 / 72 :  churn

72it [00:39,  1.84it/s]

Training Balanced Accuracy: 1.0
Testing Balanced Accuracy: 0.9111295681063123

import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import numpy as np

# Your data
data = results

# Convert to DataFrame - correct approach
df = pd.DataFrame.from_dict({(i,j): data[i][j]
                           for i in data.keys()
                           for j in data[i].keys()},
                          orient='index')

# The index is a MultiIndex with (dataset, score_type)
# Let's properly unpack it
df = df.reset_index()
df.columns = ['index', 'score']  # Temporary column names

# Split the multi-index into separate columns
df[['dataset', 'score_type']] = pd.DataFrame(df['index'].tolist(), index=df.index)
df = df.drop(columns=['index'])

# Now pivot the data
plot_df = df.pivot(index='dataset', columns='score_type', values='score')

# Sort by test score for better visualization
plot_df = plot_df.sort_values('testing_score')

# Create figure
plt.figure(figsize=(14, 10), dpi=100)

# Plot training and test scores
sns.scatterplot(data=plot_df, x=plot_df.index, y='training_score',
                color='blue', label='Training Score', s=100, alpha=0.7)
sns.scatterplot(data=plot_df, x=plot_df.index, y='testing_score',
                color='red', label='Test Score', s=100, alpha=0.7)

# Add connecting lines
for i, dataset in enumerate(plot_df.index):
    plt.plot([i, i], [plot_df.loc[dataset, 'training_score'],
                      plot_df.loc[dataset, 'testing_score']],
             color='gray', alpha=0.3, linestyle='--')

# Add perfect score line
plt.axhline(1.0, color='green', linestyle=':', alpha=0.5, label='Perfect Score')

# Customize plot
plt.title('Training vs Test Scores Across Datasets', fontsize=16, pad=20)
plt.xlabel('Dataset', fontsize=12)
plt.ylabel('Score', fontsize=12)
plt.xticks(rotation=90, fontsize=8)
plt.yticks(np.arange(0, 1.1, 0.1))
plt.ylim(0, 1.05)
plt.legend(loc='upper right', bbox_to_anchor=(1.15, 1))
plt.grid(True, alpha=0.3)

# Adjust layout
plt.tight_layout()
plt.show()

image-title-here

import numpy as np
import pandas as pd
from scipy.stats import ttest_rel, wilcoxon, shapiro
import matplotlib.pyplot as plt
import seaborn as sns

def detect_overfitting(train_scores, test_scores, alpha=0.05, plot=True):
    """
    Performs statistical tests to detect overfitting between training and test scores,
    including BOTH paired t-test and Wilcoxon signed-rank test for robustness.

    Parameters:
    -----------
    train_scores : array-like
        Array of training scores (e.g., accuracies) for multiple datasets
    test_scores : array-like
        Array of test scores for the same datasets
    alpha : float, default=0.05
        Significance level for statistical tests
    plot : bool, default=True
        Whether to generate diagnostic plots

    Returns:
    --------
    dict
        Dictionary containing test results and effect sizes
    """

    # Convert to numpy arrays
    train_scores = np.asarray(train_scores)
    test_scores = np.asarray(test_scores)
    differences = train_scores - test_scores

    # Initialize results dictionary
    results = {
        'n_datasets': len(train_scores),
        'mean_train_score': np.mean(train_scores),
        'mean_test_score': np.mean(test_scores),
        'mean_difference': np.mean(differences),
        'median_difference': np.median(differences),
        'std_difference': np.std(differences, ddof=1),
    }

    # Normality test (for interpretation guidance)
    _, shapiro_p = shapiro(differences)
    results['normality_shapiro_p'] = shapiro_p
    results['normal_distribution'] = shapiro_p >= alpha

    # ========================
    # 1. Paired t-test (parametric)
    # ========================
    t_stat, t_p = ttest_rel(train_scores, test_scores, alternative='greater')
    results['paired_ttest'] = {
        'statistic': t_stat,
        'p_value': t_p,
        'significant': t_p < alpha
    }

    # ========================
    # 2. Wilcoxon signed-rank test (non-parametric)
    # ========================
    w_stat, w_p = wilcoxon(train_scores, test_scores, alternative='greater')
    results['wilcoxon_test'] = {
        'statistic': w_stat,
        'p_value': w_p,
        'significant': w_p < alpha
    }

    # ========================
    # Effect size measures
    # ========================
    # Cohen's d (standardized mean difference)
    results['cohens_d'] = results['mean_difference'] / results['std_difference'] if results['std_difference'] > 0 else 0

    # Rank-biserial correlation (for Wilcoxon)
    n = len(differences)
    results['rank_biserial'] = 1 - (2 * w_stat) / (n * (n + 1)) if n > 0 else 0

    # Effect size classification
    for eff_size in ['cohens_d', 'rank_biserial']:
        val = abs(results[eff_size])
        if val > 0.8:
            results[f'{eff_size}_interpret'] = 'large'
        elif val > 0.5:
            results[f'{eff_size}_interpret'] = 'medium'
        else:
            results[f'{eff_size}_interpret'] = 'small'

    # ========================
    # Generate plots
    # ========================
    if plot:
        plt.figure(figsize=(15, 5))

        # Plot 1: Training vs Test scores
        plt.subplot(1, 3, 1)
        sns.scatterplot(x=train_scores, y=test_scores)
        plt.plot([0, 1], [0, 1], 'k--', alpha=0.5)
        plt.xlabel('Training Scores')
        plt.ylabel('Test Scores')
        plt.title('Training vs Test Scores')
        plt.grid(True, alpha=0.3)

        # Plot 2: Distribution of differences
        plt.subplot(1, 3, 2)
        sns.histplot(differences, kde=True)
        plt.axvline(results['mean_difference'], color='r', linestyle='--', label='Mean')
        plt.axvline(results['median_difference'], color='g', linestyle=':', label='Median')
        plt.xlabel('Train Score - Test Score')
        plt.title('Distribution of Differences')
        plt.legend()
        plt.grid(True, alpha=0.3)

        # Plot 3: Boxplot of scores
        plt.subplot(1, 3, 3)
        pd.DataFrame({
            'Training': train_scores,
            'Test': test_scores
        }).boxplot()
        plt.title('Score Distributions')
        plt.grid(True, alpha=0.3)

        plt.tight_layout()
        plt.show()

    return results

# Run the analysis
training_scores = [result['training_score'] for result in results.values()]
testing_scores = [result['testing_score'] for result in results.values()]
results = detect_overfitting(training_scores, testing_scores)

# Print results in a readable format
print("=== Overfitting Detection Report ===")
print(f"Datasets: {results['n_datasets']}")
print(f"Mean training score: {results['mean_train_score']:.3f}")
print(f"Mean test score: {results['mean_test_score']:.3f}")
print(f"Mean difference: {results['mean_difference']:.3f} (SD={results['std_difference']:.3f})")
print(f"Spearman rho: {stats.spearmanr(training_scores, testing_scores).correlation:.3f}")

print("\n=== Normality Test ===")
print(f"Shapiro-Wilk p-value: {results['normality_shapiro_p']:.4f}")
print("Interpretation:", "Normal distribution" if results['normal_distribution'] else "Non-normal distribution")

print("\n=== Statistical Tests ===")
print("1. Paired t-test:")
print(f"   t = {results['paired_ttest']['statistic']:.3f}, p = {results['paired_ttest']['p_value']:.5f}")
print(f"   Significant: {'✅' if results['paired_ttest']['significant'] else '❌'}")

print("\n2. Wilcoxon signed-rank test:")
print(f"   W = {results['wilcoxon_test']['statistic']:.3f}, p = {results['wilcoxon_test']['p_value']:.5f}")
print(f"   Significant: {'✅' if results['wilcoxon_test']['significant'] else '❌'}")

print("\n=== Effect Sizes ===")
print(f"Cohen's d: {results['cohens_d']:.3f} ({results['cohens_d_interpret']} effect)")
print(f"Rank-biserial correlation: {results['rank_biserial']:.3f} ({results['rank_biserial_interpret']} effect)")

image-title-here

=== Overfitting Detection Report ===
Datasets: 72
Mean training score: 0.975
Mean test score: 0.764
Mean difference: 0.210 (SD=0.175)
Spearman rho: 0.208

=== Normality Test ===
Shapiro-Wilk p-value: 0.0001
Interpretation: Non-normal distribution

=== Statistical Tests ===
1. Paired t-test:
   t = 10.184, p = 0.00000
   Significant: ✅

2. Wilcoxon signed-rank test:
   W = 2556.000, p = 0.00000
   Significant: ✅

=== Effect Sizes ===
Cohen's d: 1.200 (large effect)
Rank-biserial correlation: 0.027 (small effect)

2 - Genbooster

from os import name
from sklearn.linear_model import LinearRegression
from sklearn.tree import DecisionTreeRegressor, ExtraTreeRegressor
from sklearn.neighbors import KNeighborsRegressor
from sklearn.neural_network import MLPRegressor

estimators = [LinearRegression()]
name_estimators = ["LinearRegression"]

balanced_accuracy_scorer = make_scorer(balanced_accuracy_score)

for j, est in enumerate(estimators):
  results = {}
  print("\n ---------- GenBooster +", name_estimators[j])
  try:
    for i, dataset in tqdm(enumerate(clf_datasets.items())):
      dataset_name = dataset[0]
      print("\n ----------", (i + 1), "/", len(clf_datasets.items()), ": ", dataset_name)
      try:
        X, y = dataset[1]['dataset'][0], dataset[1]['dataset'][1]
        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)
        X_train = X_train.astype(np.float64)
        X_test = X_test.astype(np.float64)
        # Split dataset into training and testing sets
        model = BoosterClassifier(base_estimator=est)
        model.fit(X_train, y_train)
        y_pred_train = model.predict(X_train)
        y_pred = model.predict(X_test)
        # Calculate balanced accuracy
        balanced_acc_train = balanced_accuracy_score(y_train, y_pred_train)
        balanced_acc_test = balanced_accuracy_score(y_test, y_pred)
        print("Training Balanced Accuracy:", balanced_acc_train)
        print("Testing Balanced Accuracy:", balanced_acc_test)
        results[dataset_name] = {"training_score": balanced_acc_train,
                                "testing_score": balanced_acc_test}
      except Exception as e:
        print("Error", e)
        continue
  except Exception as e:
    print("Error", e)
    continue

  # Your data
  data = results

  # Convert to DataFrame - correct approach
  df = pd.DataFrame.from_dict({(i,j): data[i][j]
                            for i in data.keys()
                            for j in data[i].keys()},
                            orient='index')

  # The index is a MultiIndex with (dataset, score_type)
  # Let's properly unpack it
  df = df.reset_index()
  df.columns = ['index', 'score']  # Temporary column names

  # Split the multi-index into separate columns
  df[['dataset', 'score_type']] = pd.DataFrame(df['index'].tolist(), index=df.index)
  df = df.drop(columns=['index'])

  # Now pivot the data
  plot_df = df.pivot(index='dataset', columns='score_type', values='score')

  # Sort by test score for better visualization
  plot_df = plot_df.sort_values('testing_score')

  # Create figure
  plt.figure(figsize=(14, 10), dpi=100)

  # Plot training and test scores
  sns.scatterplot(data=plot_df, x=plot_df.index, y='training_score',
                  color='blue', label='Training Score', s=100, alpha=0.7)
  sns.scatterplot(data=plot_df, x=plot_df.index, y='testing_score',
                  color='red', label='Test Score', s=100, alpha=0.7)

  # Add connecting lines
  for i, dataset in enumerate(plot_df.index):
      plt.plot([i, i], [plot_df.loc[dataset, 'training_score'],
                        plot_df.loc[dataset, 'testing_score']],
              color='gray', alpha=0.3, linestyle='--')

  # Add perfect score line
  plt.axhline(1.0, color='green', linestyle=':', alpha=0.5, label='Perfect Score')

  # Customize plot
  plt.title(f'Training vs Test Scores Across Datasets for' + name_estimators[j], fontsize=16, pad=20)
  plt.xlabel('Dataset', fontsize=12)
  plt.ylabel('Score', fontsize=12)
  plt.xticks(rotation=90, fontsize=8)
  plt.yticks(np.arange(0, 1.1, 0.1))
  plt.ylim(0, 1.05)
  plt.legend(loc='upper right', bbox_to_anchor=(1.15, 1))
  plt.grid(True, alpha=0.3)

  # Adjust layout
  plt.tight_layout()
  plt.show()

  # Run the analysis
  training_scores = [result['training_score'] for result in results.values()]
  testing_scores = [result['testing_score'] for result in results.values()]
  results = detect_overfitting(training_scores, testing_scores)

  # Print results in a readable format
  print("=== Overfitting Detection Report ===")
  print(f"Datasets: {results['n_datasets']}")
  print(f"Mean training score: {results['mean_train_score']:.3f}")
  print(f"Mean test score: {results['mean_test_score']:.3f}")
  print(f"Mean difference: {results['mean_difference']:.3f} (SD={results['std_difference']:.3f})")
  print(f"Spearman rho: {stats.spearmanr(training_scores, testing_scores).correlation:.3f}")

  print("\n=== Normality Test ===")
  print(f"Shapiro-Wilk p-value: {results['normality_shapiro_p']:.4f}")
  print("Interpretation:", "Normal distribution" if results['normal_distribution'] else "Non-normal distribution")

  print("\n=== Statistical Tests ===")
  print("1. Paired t-test:")
  print(f"   t = {results['paired_ttest']['statistic']:.3f}, p = {results['paired_ttest']['p_value']:.5f}")
  print(f"   Significant: {'✅' if results['paired_ttest']['significant'] else '❌'}")

  print("\n2. Wilcoxon signed-rank test:")
  print(f"   W = {results['wilcoxon_test']['statistic']:.3f}, p = {results['wilcoxon_test']['p_value']:.5f}")
  print(f"   Significant: {'✅' if results['wilcoxon_test']['significant'] else '❌'}")

  print("\n=== Effect Sizes ===")
  print(f"Cohen's d: {results['cohens_d']:.3f} ({results['cohens_d_interpret']} effect)")
  print(f"Rank-biserial correlation: {results['rank_biserial']:.3f} ({results['rank_biserial_interpret']} effect)")

 ---------- GenBooster + LinearRegression

0it [00:00, ?it/s]

 ---------- 1 / 72 :  kr-vs-kp

1it [00:00,  2.21it/s]

Training Balanced Accuracy: 0.9480928346328172
Testing Balanced Accuracy: 0.9182957393483708

 ---------- 2 / 72 :  letter

2it [00:08,  4.76s/it]

Training Balanced Accuracy: 0.6532232216708022
Testing Balanced Accuracy: 0.6037087912087912

 ---------- 3 / 72 :  balance-scale

3it [00:08,  2.76s/it]

Training Balanced Accuracy: 0.6579835623313884
Testing Balanced Accuracy: 0.6492236337971365

 ---------- 4 / 72 :  mfeat-factors

4it [00:10,  2.50s/it]

Training Balanced Accuracy: 0.8712500000000001
Testing Balanced Accuracy: 0.8649999999999999

 ---------- 5 / 72 :  mfeat-fourier

5it [00:12,  2.35s/it]

Training Balanced Accuracy: 0.8174999999999999
Testing Balanced Accuracy: 0.7849999999999999

 ---------- 6 / 72 :  breast-w

6it [00:13,  1.64s/it]

Training Balanced Accuracy: 0.9810158838019196
Testing Balanced Accuracy: 0.962409420289855

 ---------- 7 / 72 :  mfeat-karhunen

7it [00:15,  1.80s/it]

Training Balanced Accuracy: 0.8699999999999999
Testing Balanced Accuracy: 0.865

 ---------- 8 / 72 :  mfeat-morphological

8it [00:16,  1.70s/it]

Training Balanced Accuracy: 0.75
Testing Balanced Accuracy: 0.7250000000000001

 ---------- 9 / 72 :  mfeat-zernike

9it [00:21,  2.61s/it]

Training Balanced Accuracy: 0.7987499999999998
Testing Balanced Accuracy: 0.74

 ---------- 10 / 72 :  cmc

10it [00:21,  2.00s/it]

Training Balanced Accuracy: 0.539420682309971
Testing Balanced Accuracy: 0.48957420514548927

 ---------- 11 / 72 :  optdigits

11it [00:23,  2.01s/it]

Training Balanced Accuracy: 0.8338839752605576
Testing Balanced Accuracy: 0.865

 ---------- 12 / 72 :  credit-approval

12it [00:24,  1.49s/it]

Training Balanced Accuracy: 0.8800414474732983
Testing Balanced Accuracy: 0.7980625931445604

 ---------- 13 / 72 :  credit-g

13it [00:24,  1.17s/it]

Training Balanced Accuracy: 0.6815476190476191
Testing Balanced Accuracy: 0.7059523809523809

 ---------- 14 / 72 :  pendigits

14it [00:26,  1.46s/it]

Training Balanced Accuracy: 0.9082342205793651
Testing Balanced Accuracy: 0.8939849624060152

 ---------- 15 / 72 :  diabetes

15it [00:27,  1.10s/it]

Training Balanced Accuracy: 0.7434696261682243
Testing Balanced Accuracy: 0.6457407407407407

 ---------- 16 / 72 :  spambase

16it [00:27,  1.09it/s]

Training Balanced Accuracy: 0.8971008789190608
Testing Balanced Accuracy: 0.8953865467099069

 ---------- 17 / 72 :  splice

17it [00:28,  1.19it/s]

Training Balanced Accuracy: 0.8904200133868808
Testing Balanced Accuracy: 0.8835470085470085

 ---------- 18 / 72 :  tic-tac-toe

18it [00:28,  1.40it/s]

Training Balanced Accuracy: 0.7092305954129476
Testing Balanced Accuracy: 0.7153432835820895

 ---------- 19 / 72 :  vehicle

19it [00:29,  1.33it/s]

Training Balanced Accuracy: 0.7512680047291669
Testing Balanced Accuracy: 0.7388257575757575

 ---------- 20 / 72 :  electricity

20it [00:29,  1.54it/s]

Training Balanced Accuracy: 0.7804471490091439
Testing Balanced Accuracy: 0.7631713554987212

 ---------- 21 / 72 :  satimage

21it [00:33,  1.45s/it]

Training Balanced Accuracy: 0.7904065610735302
Testing Balanced Accuracy: 0.7254887702816034

 ---------- 22 / 72 :  eucalyptus

22it [00:34,  1.36s/it]

Training Balanced Accuracy: 0.5992921311574563
Testing Balanced Accuracy: 0.5444643212085073

 ---------- 23 / 72 :  sick

23it [00:34,  1.09s/it]

Training Balanced Accuracy: 0.6421904761904762
Testing Balanced Accuracy: 0.5416666666666666

 ---------- 24 / 72 :  vowel

24it [00:37,  1.45s/it]

Training Balanced Accuracy: 0.7588383838383838
Testing Balanced Accuracy: 0.6565656565656567

 ---------- 25 / 72 :  isolet

25it [00:42,  2.61s/it]

Training Balanced Accuracy: 0.5353598014888339
Testing Balanced Accuracy: 0.4539835164835164

 ---------- 26 / 72 :  analcatdata_authorship

26it [00:43,  2.07s/it]

Training Balanced Accuracy: 0.9691163839829222
Testing Balanced Accuracy: 0.9847136047215497

 ---------- 27 / 72 :  analcatdata_dmft

27it [00:44,  1.79s/it]

Training Balanced Accuracy: 0.31318315240756367
Testing Balanced Accuracy: 0.2114058144165671

 ---------- 28 / 72 :  mnist_784

28it [00:47,  2.33s/it]

Training Balanced Accuracy: 0.6018919115939201
Testing Balanced Accuracy: 0.5142492099814755

 ---------- 29 / 72 :  pc4

29it [00:48,  1.77s/it]

Training Balanced Accuracy: 0.6706163207080265
Testing Balanced Accuracy: 0.5482954545454546

 ---------- 30 / 72 :  pc3

30it [00:48,  1.37s/it]

Training Balanced Accuracy: 0.5466884375956731
Testing Balanced Accuracy: 0.525

 ---------- 31 / 72 :  jm1

31it [00:49,  1.09s/it]

Training Balanced Accuracy: 0.5669233866908285
Testing Balanced Accuracy: 0.5454690237298934

 ---------- 32 / 72 :  kc2

32it [00:49,  1.17it/s]

Training Balanced Accuracy: 0.7185861091424521
Testing Balanced Accuracy: 0.651697699890471

 ---------- 33 / 72 :  kc1

33it [00:50,  1.36it/s]

Training Balanced Accuracy: 0.5635553470919324
Testing Balanced Accuracy: 0.5556403893872877

 ---------- 34 / 72 :  pc1

34it [00:50,  1.52it/s]

Training Balanced Accuracy: 0.5
Testing Balanced Accuracy: 0.5

 ---------- 35 / 72 :  adult

35it [00:50,  1.69it/s]

Training Balanced Accuracy: 0.7400971341967484
Testing Balanced Accuracy: 0.6551535087719298

 ---------- 36 / 72 :  Bioresponse

36it [00:51,  1.84it/s]

Training Balanced Accuracy: 0.7767986723709284
Testing Balanced Accuracy: 0.732331888295191

 ---------- 37 / 72 :  wdbc

37it [00:51,  2.11it/s]

Training Balanced Accuracy: 0.9478328173374613
Testing Balanced Accuracy: 0.9623015873015872

 ---------- 38 / 72 :  phoneme

38it [00:52,  2.38it/s]

Training Balanced Accuracy: 0.707737690038575
Testing Balanced Accuracy: 0.6642024281764636

 ---------- 39 / 72 :  qsar-biodeg

39it [00:52,  2.37it/s]

Training Balanced Accuracy: 0.8324861723727508
Testing Balanced Accuracy: 0.816967792615868

 ---------- 40 / 72 :  wall-robot-navigation

40it [00:53,  1.80it/s]

Training Balanced Accuracy: 0.653627904631567
Testing Balanced Accuracy: 0.5904361071027737

 ---------- 41 / 72 :  semeion

41it [00:55,  1.02s/it]

Training Balanced Accuracy: 0.5926607850658483
Testing Balanced Accuracy: 0.6

 ---------- 42 / 72 :  ilpd

42it [00:55,  1.24it/s]

Training Balanced Accuracy: 0.597766939872203
Testing Balanced Accuracy: 0.5581148121899362

 ---------- 43 / 72 :  madelon

43it [00:56,  1.44it/s]

Training Balanced Accuracy: 0.73625
Testing Balanced Accuracy: 0.7

 ---------- 44 / 72 :  nomao

44it [00:56,  1.53it/s]

Training Balanced Accuracy: 0.8829423602789812
Testing Balanced Accuracy: 0.9193963930806036

 ---------- 45 / 72 :  ozone-level-8hr

45it [00:58,  1.18it/s]

Training Balanced Accuracy: 0.5
Testing Balanced Accuracy: 0.5

 ---------- 46 / 72 :  cnae-9

46it [01:00,  1.45s/it]

Training Balanced Accuracy: 0.7333304959709455
Testing Balanced Accuracy: 0.699604743083004

 ---------- 47 / 72 :  first-order-theorem-proving

47it [01:02,  1.39s/it]

Training Balanced Accuracy: 0.2568945598032055
Testing Balanced Accuracy: 0.27648809523809526

 ---------- 48 / 72 :  banknote-authentication

48it [01:02,  1.06s/it]

Training Balanced Accuracy: 0.9954954954954955
Testing Balanced Accuracy: 1.0

 ---------- 49 / 72 :  blood-transfusion-service-center

49it [01:02,  1.22it/s]

Training Balanced Accuracy: 0.5951321966889054
Testing Balanced Accuracy: 0.611842105263158

 ---------- 50 / 72 :  PhishingWebsites

50it [01:03,  1.41it/s]

Training Balanced Accuracy: 0.9133593601218816
Testing Balanced Accuracy: 0.9122886931875696

 ---------- 51 / 72 :  cylinder-bands

51it [01:03,  1.72it/s]

Training Balanced Accuracy: 0.6772527472527472
Testing Balanced Accuracy: 0.6669004207573632

 ---------- 52 / 72 :  bank-marketing

52it [01:03,  1.86it/s]

Training Balanced Accuracy: 0.5602668372475554
Testing Balanced Accuracy: 0.5917464996315401

 ---------- 53 / 72 :  GesturePhaseSegmentationProcessed

53it [01:04,  1.44it/s]

Training Balanced Accuracy: 0.4052738438032556
Testing Balanced Accuracy: 0.3567099567099567

 ---------- 54 / 72 :  har

54it [01:06,  1.17it/s]

Training Balanced Accuracy: 0.6586891684653934
Testing Balanced Accuracy: 0.6022327522327522

 ---------- 55 / 72 :  dresses-sales

55it [01:06,  1.46it/s]

Training Balanced Accuracy: 0.6180213464696224
Testing Balanced Accuracy: 0.5049261083743842

 ---------- 56 / 72 :  texture

56it [01:08,  1.16s/it]

Training Balanced Accuracy: 0.9217171717171717
Testing Balanced Accuracy: 0.9090909090909093

 ---------- 57 / 72 :  connect-4

57it [01:09,  1.00s/it]

Training Balanced Accuracy: 0.36527957383567194
Testing Balanced Accuracy: 0.36229643372500514

 ---------- 58 / 72 :  MiceProtein

58it [01:13,  2.01s/it]

Training Balanced Accuracy: 0.7886130536130537
Testing Balanced Accuracy: 0.7212312030075188

 ---------- 59 / 72 :  steel-plates-fault

59it [01:15,  1.85s/it]

Training Balanced Accuracy: 0.5668243559177928
Testing Balanced Accuracy: 0.5421214137829045

 ---------- 60 / 72 :  climate-model-simulation-crashes

60it [01:15,  1.38s/it]

Training Balanced Accuracy: 0.6486486486486487
Testing Balanced Accuracy: 0.6111111111111112

 ---------- 61 / 72 :  wilt

61it [01:15,  1.06s/it]

Training Balanced Accuracy: 0.6660061646851607
Testing Balanced Accuracy: 0.6818181818181819

 ---------- 62 / 72 :  car

62it [01:16,  1.08it/s]

Training Balanced Accuracy: 0.5475360592569449
Testing Balanced Accuracy: 0.5132936507936507

 ---------- 63 / 72 :  segment

63it [01:17,  1.09s/it]

Training Balanced Accuracy: 0.821012708763058
Testing Balanced Accuracy: 0.8258268824771289

 ---------- 64 / 72 :  mfeat-pixel

64it [01:23,  2.42s/it]

Training Balanced Accuracy: 0.7024999999999999
Testing Balanced Accuracy: 0.7100000000000001

 ---------- 65 / 72 :  Fashion-MNIST

65it [01:27,  2.94s/it]

Training Balanced Accuracy: 0.5262499999999999
Testing Balanced Accuracy: 0.43499999999999994

 ---------- 66 / 72 :  jungle_chess_2pcs_raw_endgame_complete

66it [01:27,  2.19s/it]

Training Balanced Accuracy: 0.5535725243283519
Testing Balanced Accuracy: 0.5305783752385694

 ---------- 67 / 72 :  numerai28.6

67it [01:28,  1.67s/it]

Training Balanced Accuracy: 0.6070716881814764
Testing Balanced Accuracy: 0.5743074307430743

 ---------- 68 / 72 :  Devnagari-Script

68it [01:38,  4.07s/it]

Training Balanced Accuracy: 0.2718190537084399
Testing Balanced Accuracy: 0.15

 ---------- 69 / 72 :  CIFAR_10

69it [01:40,  3.60s/it]

Training Balanced Accuracy: 0.29875
Testing Balanced Accuracy: 0.205

 ---------- 70 / 72 :  Internet-Advertisements

70it [01:41,  2.66s/it]

Training Balanced Accuracy: 0.9168827257490049
Testing Balanced Accuracy: 0.9555647840531561

 ---------- 71 / 72 :  dna

71it [01:41,  2.07s/it]

Training Balanced Accuracy: 0.8917670682730924
Testing Balanced Accuracy: 0.8936965811965812

 ---------- 72 / 72 :  churn

72it [01:42,  1.42s/it]

Training Balanced Accuracy: 0.5900049020872572
Testing Balanced Accuracy: 0.5955149501661129

image-title-here

=== Overfitting Detection Report ===
Datasets: 72
Mean training score: 0.692
Mean test score: 0.661
Mean difference: 0.031 (SD=0.041)
Spearman rho: 0.972

=== Normality Test ===
Shapiro-Wilk p-value: 0.0153
Interpretation: Non-normal distribution

=== Statistical Tests ===
1. Paired t-test:
   t = 6.302, p = 0.00000
   Significant: ✅

2. Wilcoxon signed-rank test:
   W = 2125.000, p = 0.00000
   Significant: ✅

=== Effect Sizes ===
Cohen's d: 0.743 (medium effect)
Rank-biserial correlation: 0.191 (small effect)

But… Which one is the best in this particular setting?

Comments powered by Talkyard.

Citation

For attribution, please cite this work as:

T. Moudiki (2025-06-14). An Overfitting dilemma: XGBoost Default Hyperparameters vs GenericBooster + LinearRegression Default Hyperparameters. Retrieved from https://thierrymoudiki.github.io/blog/2025/06/14/python/xgboost-default-overfitting

BibTeX citation (remove empty spaces)
@misc{ tmoudiki20250614, author = { T. Moudiki }, title = { An Overfitting dilemma: XGBoost Default Hyperparameters vs GenericBooster + LinearRegression Default Hyperparameters }, url = { https://thierrymoudiki.github.io/blog/2025/06/14/python/xgboost-default-overfitting }, year = { 2025 } }

Under License Creative Commons Attribution 4.0 International

Previous publications

Combining any model with GARCH(1,1) for probabilistic stock forecasting Sep 23, 2025

Generating Synthetic Data with R-vine Copulas using esgtoolkit in R Sep 21, 2025

Reimagining Equity Solvency Capital Requirement Approximation (one of my Master's Thesis subjects): From Bilinear Interpolation to Probabilistic Machine Learning Sep 16, 2025

Transfer Learning using ahead::ridge2f on synthetic stocks returns Pt.2: synthetic data generation Sep 9, 2025

Transfer Learning using ahead::ridge2f on synthetic stocks returns Sep 8, 2025

I'm supposed to present 'Conformal Predictive Simulations for Univariate Time Series' at COPA CONFERENCE 2025 in London... Sep 4, 2025

external regressors in ahead::dynrmf's interface for Machine learning forecasting Sep 1, 2025

Another interesting decision, now for 'Beyond Nelson-Siegel and splines: A model-agnostic Machine Learning framework for discount curve calibration, interpolation and extrapolation' Aug 20, 2025

Boosting any randomized based learner for regression, classification and univariate/multivariate time series forcasting Jul 26, 2025

New nnetsauce version with CustomBackPropRegressor (CustomRegressor with Backpropagation) and ElasticNet2Regressor (Ridge2 with ElasticNet regularization) Jul 15, 2025

mlsauce (home to a model-agnostic gradient boosting algorithm) can now be installed from PyPI. Jul 10, 2025

A user-friendly graphical interface to techtonique dot net's API (will eventually contain graphics). Jul 8, 2025

Calling =TECHTO_MLCLASSIFICATION for Machine Learning supervised CLASSIFICATION in Excel is just a matter of copying and pasting Jul 7, 2025

Calling =TECHTO_MLREGRESSION for Machine Learning supervised regression in Excel is just a matter of copying and pasting Jul 6, 2025

Calling =TECHTO_RESERVING and =TECHTO_MLRESERVING for claims triangle reserving in Excel is just a matter of copying and pasting Jul 5, 2025

Calling =TECHTO_SURVIVAL for Survival Analysis in Excel is just a matter of copying and pasting Jul 4, 2025

Calling =TECHTO_SIMULATION for Stochastic Simulation in Excel is just a matter of copying and pasting Jul 3, 2025

Calling =TECHTO_FORECAST for forecasting in Excel is just a matter of copying and pasting Jul 2, 2025

Random Vector Functional Link (RVFL) artificial neural network with 2 regularization parameters successfully used for forecasting/synthetic simulation in professional settings: Extensions (including Bayesian) Jul 1, 2025

R version of 'Backpropagating quasi-randomized neural networks' Jun 24, 2025

Backpropagating quasi-randomized neural networks Jun 23, 2025

Beyond ARMA-GARCH: leveraging any statistical model for volatility forecasting Jun 21, 2025

Stacked generalization (Machine Learning model stacking) + conformal prediction for forecasting with ahead::mlf Jun 18, 2025

An Overfitting dilemma: XGBoost Default Hyperparameters vs GenericBooster + LinearRegression Default Hyperparameters Jun 14, 2025

Programming language-agnostic reserving using RidgeCV, LightGBM, XGBoost, and ExtraTrees Machine Learning models Jun 13, 2025

Exceptionally, and on a more personal note (otherwise I may get buried alive)... Jun 10, 2025

Free R, Python and SQL editors in techtonique dot net Jun 9, 2025

Beyond Nelson-Siegel and splines: A model-agnostic Machine Learning framework for discount curve calibration, interpolation and extrapolation Jun 7, 2025

scikit-learn, glmnet, xgboost, lightgbm, pytorch, keras, nnetsauce in probabilistic Machine Learning (for longitudinal data) Reserving (work in progress) Jun 6, 2025

R version of Probabilistic Machine Learning (for longitudinal data) Reserving (work in progress) Jun 5, 2025

Probabilistic Machine Learning (for longitudinal data) Reserving (work in progress) Jun 4, 2025

Python version of Beyond ARMA-GARCH: leveraging model-agnostic Quasi-Randomized networks and conformal prediction for nonparametric probabilistic stock forecasting (ML-ARCH) Jun 3, 2025

Beyond ARMA-GARCH: leveraging model-agnostic Machine Learning and conformal prediction for nonparametric probabilistic stock forecasting (ML-ARCH) Jun 2, 2025

Permutations and SHAPley values for feature importance in techtonique dot net's API (with R + Python + the command line) Jun 1, 2025

Which patient is going to survive longer? Another guide to using techtonique dot net's API (with R + Python + the command line) for survival analysis May 31, 2025

A Guide to Using techtonique.net's API and rush for simulating and plotting Stochastic Scenarios May 30, 2025

Simulating Stochastic Scenarios with Diffusion Models: A Guide to Using techtonique.net's API for the purpose May 29, 2025

Will my apartment in 5th avenue be overpriced or not? Harnessing the power of www.techtonique.net (+ xgboost, lightgbm, catboost) to find out May 28, 2025

How long must I wait until something happens: A Comprehensive Guide to Survival Analysis via an API May 27, 2025

Harnessing the Power of techtonique.net: A Comprehensive Guide to Machine Learning Classification via an API May 26, 2025

Quantile regression with any regressor -- Examples with RandomForestRegressor, RidgeCV, KNeighborsRegressor May 20, 2025

Survival stacking: survival analysis translated as supervised classification in R and Python May 5, 2025

'Bayesian' optimization of hyperparameters in a R machine learning model using the bayesianrvfl package Apr 25, 2025

A lightweight interface to scikit-learn in R: Bayesian and Conformal prediction Apr 21, 2025

A lightweight interface to scikit-learn in R Pt.2: probabilistic time series forecasting in conjunction with ahead::dynrmf Apr 20, 2025

Extending the Theta forecasting method to GLMs, GAMs, GLMBOOST and attention: benchmarking on Tourism, M1, M3 and M4 competition data sets (28000 series) Apr 14, 2025

Extending the Theta forecasting method to GLMs and attention Apr 8, 2025

Nonlinear conformalized Generalized Linear Models (GLMs) with R package 'rvfl' (and other models) Mar 31, 2025

Probabilistic Time Series Forecasting (predictive simulations) in Microsoft Excel using Python, xlwings lite and www.techtonique.net Mar 28, 2025

Conformalize (improved prediction intervals and simulations) any R Machine Learning model with misc::conformalize Mar 25, 2025

My poster for the 18th FINANCIAL RISKS INTERNATIONAL FORUM by Institut Louis Bachelier/Fondation du Risque/Europlace Institute of Finance Mar 19, 2025

Interpretable probabilistic kernel ridge regression using Matérn 3/2 kernels Mar 16, 2025

(News from) Probabilistic Forecasting of univariate and multivariate Time Series using Quasi-Randomized Neural Networks (Ridge2) and Conformal Prediction Mar 9, 2025

Word-Online: re-creating Karpathy's char-RNN (with supervised linear online learning of word embeddings) for text completion Mar 8, 2025

CRAN-like repository for most recent releases of Techtonique's R packages Mar 2, 2025

Presenting 'Online Probabilistic Estimation of Carbon Beta and Carbon Shapley Values for Financial and Climate Risk' at Institut Louis Bachelier Feb 27, 2025

Web app with DeepSeek R1 and Hugging Face API for chatting Feb 23, 2025

tisthemachinelearner: A Lightweight interface to scikit-learn with 2 classes, Classifier and Regressor (in Python and R) Feb 17, 2025

R version of survivalist: Probabilistic model-agnostic survival analysis using scikit-learn, xgboost, lightgbm (and conformal prediction) Feb 12, 2025

Model-agnostic global Survival Prediction of Patients with Myeloid Leukemia in QRT/Gustave Roussy Challenge (challengedata.ens.fr): Python's survivalist Quickstart Feb 10, 2025

A simple test of the martingale hypothesis in esgtoolkit Feb 3, 2025

Command Line Interface (CLI) for techtonique.net's API Jan 31, 2025

Gradient-Boosting and Boostrap aggregating anything (alert: high performance): Part5, easier install and Rust backend Jan 27, 2025

Just got a paper on conformal prediction REJECTED by International Journal of Forecasting despite evidence on 30,000 time series (and more). What's going on? Part2: 1311 time series from the Tourism competition Jan 20, 2025

Techtonique is out! (with a tutorial in various programming languages and formats) Jan 14, 2025

Univariate and Multivariate Probabilistic Forecasting with nnetsauce and TabPFN Jan 14, 2025

Just got a paper on conformal prediction REJECTED by International Journal of Forecasting despite evidence on 30,000 time series (and more). What's going on? Jan 5, 2025

Python and Interactive dashboard version of Stock price forecasting with Deep Learning: throwing power at the problem (and why it won't make you rich) Dec 31, 2024

Stock price forecasting with Deep Learning: throwing power at the problem (and why it won't make you rich) Dec 29, 2024

No-code Machine Learning Cross-validation and Interpretability in techtonique.net Dec 23, 2024

survivalist: Probabilistic model-agnostic survival analysis using scikit-learn, glmnet, xgboost, lightgbm, pytorch, keras, nnetsauce and mlsauce Dec 15, 2024

Model-agnostic 'Bayesian' optimization (for hyperparameter tuning) using conformalized surrogates in GPopt Dec 9, 2024

You can beat Forecasting LLMs (Large Language Models a.k.a foundation models) with nnetsauce.MTS Pt.2: Generic Gradient Boosting Dec 1, 2024

You can beat Forecasting LLMs (Large Language Models a.k.a foundation models) with nnetsauce.MTS Nov 24, 2024

Unified interface and conformal prediction (calibrated prediction intervals) for R package forecast (and 'affiliates') Nov 23, 2024

GLMNet in Python: Generalized Linear Models Nov 18, 2024

Gradient-Boosting anything (alert: high performance): Part4, Time series forecasting Nov 10, 2024

Predictive scenarios simulation in R, Python and Excel using Techtonique API Nov 3, 2024

Chat with your tabular data in www.techtonique.net Oct 30, 2024

Gradient-Boosting anything (alert: high performance): Part3, Histogram-based boosting Oct 28, 2024

R editor and SQL console (in addition to Python editors) in www.techtonique.net Oct 21, 2024

R and Python consoles + JupyterLite in www.techtonique.net Oct 15, 2024

Gradient-Boosting anything (alert: high performance): Part2, R version Oct 14, 2024

Gradient-Boosting anything (alert: high performance) Oct 6, 2024

Benchmarking 30 statistical/Machine Learning models on the VN1 Forecasting -- Accuracy challenge Oct 4, 2024

Automated random variable distribution inference using Kullback-Leibler divergence and simulating best-fitting distribution Oct 2, 2024

Forecasting in Excel using Techtonique's Machine Learning APIs under the hood Sep 30, 2024

Techtonique web app for data-driven decisions using Mathematics, Statistics, Machine Learning, and Data Visualization Sep 25, 2024

Parallel for loops (Map or Reduce) + New versions of nnetsauce and ahead Sep 16, 2024

Adaptive (online/streaming) learning with uncertainty quantification using Polyak averaging in learningmachine Sep 10, 2024

New versions of nnetsauce and ahead Sep 9, 2024

Prediction sets and prediction intervals for conformalized Auto XGBoost, Auto LightGBM, Auto CatBoost, Auto GradientBoosting Sep 2, 2024

Quick/automated R package development workflow (assuming you're using macOS or Linux) Part2 Aug 30, 2024

R package development workflow (assuming you're using macOS or Linux) Aug 27, 2024

A new method for deriving a nonparametric confidence interval for the mean Aug 26, 2024

Conformalized adaptive (online/streaming) learning using learningmachine in Python and R Aug 19, 2024

Bayesian (nonlinear) adaptive learning Aug 12, 2024

Auto XGBoost, Auto LightGBM, Auto CatBoost, Auto GradientBoosting Aug 5, 2024

Copulas for uncertainty quantification in time series forecasting Jul 28, 2024

Forecasting uncertainty: sequential split conformal prediction + Block bootstrap (web app) Jul 22, 2024

learningmachine for Python (new version) Jul 15, 2024

learningmachine v2.0.0: Machine Learning with explanations and uncertainty quantification Jul 8, 2024

My presentation at ISF 2024 conference (slides with nnetsauce probabilistic forecasting news) Jul 3, 2024

10 uncertainty quantification methods in nnetsauce forecasting Jul 1, 2024

Forecasting with XGBoost embedded in Quasi-Randomized Neural Networks Jun 24, 2024

Forecasting Monthly Airline Passenger Numbers with Quasi-Randomized Neural Networks Jun 17, 2024

Automated hyperparameter tuning using any conformalized surrogate Jun 9, 2024

Recognizing handwritten digits with Ridge2Classifier Jun 3, 2024

Forecasting the Economy May 27, 2024

A detailed introduction to Deep Quasi-Randomized 'neural' networks May 19, 2024

Probability of receiving a loan; using learningmachine May 12, 2024

mlsauce's `v0.18.2`: various examples and benchmarks with dimension reduction May 6, 2024

mlsauce's `v0.17.0`: boosting with Elastic Net, polynomials and heterogeneity in explanatory variables Apr 29, 2024

mlsauce's `v0.13.0`: taking into account inputs heterogeneity through clustering Apr 21, 2024

mlsauce's `v0.12.0`: prediction intervals for LSBoostRegressor Apr 15, 2024

Conformalized predictive simulations for univariate time series on more than 250 data sets Apr 7, 2024

learningmachine v1.1.2: for Python Apr 1, 2024

learningmachine v1.0.0: prediction intervals around the probability of the event 'a tumor being malignant' Mar 25, 2024

Bayesian inference and conformal prediction (prediction intervals) in nnetsauce v0.18.1 Mar 18, 2024

Multiple examples of Machine Learning forecasting with ahead Mar 11, 2024

rtopy (v0.1.1): calling R functions in Python Mar 4, 2024

ahead forecasting (v0.10.0): fast time series model calibration and Python plots Feb 26, 2024

A plethora of datasets at your fingertips Part3: how many times do couples cheat on each other? Feb 19, 2024

nnetsauce's introduction as of 2024-02-11 (new version 0.17.0) Feb 11, 2024

Tuning Machine Learning models with GPopt's new version Part 2 Feb 5, 2024

Tuning Machine Learning models with GPopt's new version Jan 29, 2024

Subsampling continuous and discrete response variables Jan 22, 2024

DeepMTS, a Deep Learning Model for Multivariate Time Series Jan 15, 2024

A classifier that's very accurate (and deep) Pt.2: there are > 90 classifiers in nnetsauce Jan 8, 2024

learningmachine: prediction intervals for conformalized Kernel ridge regression and Random Forest Jan 1, 2024

A plethora of datasets at your fingertips Part2: how many times do couples cheat on each other? Descriptive analytics, interpretability and prediction intervals using conformal prediction Dec 25, 2023

Diffusion models in Python with esgtoolkit (Part2) Dec 18, 2023

Diffusion models in Python with esgtoolkit Dec 11, 2023

Julia packaging at the command line Dec 4, 2023

Quasi-randomized nnetworks in Julia, Python and R Nov 27, 2023

A plethora of datasets at your fingertips Nov 20, 2023

A classifier that's very accurate (and deep) Nov 12, 2023

mlsauce version 0.8.10: Statistical/Machine Learning with Python and R Nov 5, 2023

AutoML in nnetsauce (randomized and quasi-randomized nnetworks) Pt.2: multivariate time series forecasting Oct 29, 2023

AutoML in nnetsauce (randomized and quasi-randomized nnetworks) Oct 22, 2023

Version v0.14.0 of nnetsauce for R and Python Oct 16, 2023

A diffusion model: G2++ Oct 9, 2023

Diffusion models in ESGtoolkit + announcements Oct 2, 2023

An infinity of time series forecasting models in nnetsauce (Part 2 with uncertainty quantification) Sep 25, 2023

(News from) forecasting in Python with ahead (progress bars and plots) Sep 18, 2023

Forecasting in Python with ahead Sep 11, 2023

Risk-neutralize simulations Sep 4, 2023

Comparing cross-validation results using crossval_ml and boxplots Aug 27, 2023

Reminder Apr 30, 2023

Did you ask ChatGPT about who you are? Apr 16, 2023

A new version of nnetsauce (randomized and quasi-randomized 'neural' networks) Apr 2, 2023

Simple interfaces to the forecasting API Nov 23, 2022

A web application for forecasting in Python, R, Ruby, C#, JavaScript, PHP, Go, Rust, Java, MATLAB, etc. Nov 2, 2022

Prediction intervals (not only) for Boosted Configuration Networks in Python Oct 5, 2022

Boosted Configuration (neural) Networks Pt. 2 Sep 3, 2022

Boosted Configuration (_neural_) Networks for classification Jul 21, 2022

A Machine Learning workflow using Techtonique Jun 6, 2022

Super Mario Bros © in the browser using PyScript May 8, 2022

News from ESGtoolkit, ycinterextra, and nnetsauce Apr 4, 2022

Explaining a Keras _neural_ network predictions with the-teller Mar 11, 2022

New version of nnetsauce -- various quasi-randomized networks Feb 12, 2022

A dashboard illustrating bivariate time series forecasting with `ahead` Jan 14, 2022

Hundreds of Statistical/Machine Learning models for univariate time series, using ahead, ranger, xgboost, and caret Dec 20, 2021

Forecasting with `ahead` (Python version) Dec 13, 2021

Tuning and interpreting LSBoost Nov 15, 2021

Time series cross-validation using `crossvalidation` (Part 2) Nov 7, 2021

Fast and scalable forecasting with ahead::ridge2f Oct 31, 2021

Automatic Forecasting with `ahead::dynrmf` and Ridge regression Oct 22, 2021

Forecasting with `ahead` Oct 15, 2021

Classification using linear regression Sep 26, 2021

`crossvalidation` and random search for calibrating support vector machines Aug 6, 2021

parallel grid search cross-validation using `crossvalidation` Jul 31, 2021

`crossvalidation` on R-universe, plus a classification example Jul 23, 2021

Documentation and source code for GPopt, a package for Bayesian optimization Jul 2, 2021

Hyperparameters tuning with GPopt Jun 11, 2021

A forecasting tool (API) with examples in curl, R, Python May 28, 2021

Bayesian Optimization with GPopt Part 2 (save and resume) Apr 30, 2021

Bayesian Optimization with GPopt Apr 16, 2021

Compatibility of nnetsauce and mlsauce with scikit-learn Mar 26, 2021

Explaining xgboost predictions with the teller Mar 12, 2021

An infinity of time series models in nnetsauce Mar 6, 2021

New activation functions in mlsauce's LSBoost Feb 12, 2021

2020 recap, Gradient Boosting, Generalized Linear Models, AdaOpt with nnetsauce and mlsauce Dec 29, 2020

A deeper learning architecture in nnetsauce Dec 18, 2020

Classify penguins with nnetsauce's MultitaskClassifier Dec 11, 2020

Bayesian forecasting for uni/multivariate time series Dec 4, 2020

Generalized nonlinear models in nnetsauce Nov 28, 2020

Boosting nonlinear penalized least squares Nov 21, 2020

Statistical/Machine Learning explainability using Kernel Ridge Regression surrogates Nov 6, 2020

NEWS Oct 30, 2020

A glimpse into my PhD journey Oct 23, 2020

Submitting R package to CRAN Oct 16, 2020

Simulation of dependent variables in ESGtoolkit Oct 9, 2020

Forecasting lung disease progression Oct 2, 2020

New nnetsauce Sep 25, 2020

Technical documentation Sep 18, 2020

A new version of nnetsauce, and a new Techtonique website Sep 11, 2020

Back next week, and a few announcements Sep 4, 2020

Explainable 'AI' using Gradient Boosted randomized networks Pt2 (the Lasso) Jul 31, 2020

LSBoost: Explainable 'AI' using Gradient Boosted randomized networks (with examples in R and Python) Jul 24, 2020

nnetsauce version 0.5.0, randomized neural networks on GPU Jul 17, 2020

Maximizing your tip as a waiter (Part 2) Jul 10, 2020

New version of mlsauce, with Gradient Boosted randomized networks and stump decision trees Jul 3, 2020

Announcements Jun 26, 2020

Parallel AdaOpt classification Jun 19, 2020

Comments section and other news Jun 12, 2020

Maximizing your tip as a waiter Jun 5, 2020

AdaOpt classification on MNIST handwritten digits (without preprocessing) May 29, 2020

AdaOpt (a probabilistic classifier based on a mix of multivariable optimization and nearest neighbors) for R May 22, 2020

AdaOpt May 15, 2020

Custom errors for cross-validation using crossval::crossval_ml May 8, 2020

Documentation+Pypi for the `teller`, a model-agnostic tool for Machine Learning explainability May 1, 2020

Encoding your categorical variables based on the response variable and correlations Apr 24, 2020

Linear model, xgboost and randomForest cross-validation using crossval::crossval_ml Apr 17, 2020

Grid search cross-validation using crossval Apr 10, 2020

Documentation for the querier, a query language for Data Frames Apr 3, 2020

Time series cross-validation using crossval Mar 27, 2020

On model specification, identification, degrees of freedom and regularization Mar 20, 2020

Import data into the querier (now on Pypi), a query language for Data Frames Mar 13, 2020

R notebooks for nnetsauce Mar 6, 2020

Version 0.4.0 of nnetsauce, with fruits and breast cancer classification Feb 28, 2020

Create a specific feed in your Jekyll blog Feb 21, 2020

Git/Github for contributing to package development Feb 14, 2020

Feedback forms for contributing Feb 7, 2020

nnetsauce for R Jan 31, 2020

A new version of nnetsauce (v0.3.1) Jan 24, 2020

ESGtoolkit, a tool for Monte Carlo simulation (v0.2.0) Jan 17, 2020

Search bar, new year 2020 Jan 10, 2020

2019 Recap, the nnetsauce, the teller and the querier Dec 20, 2019

Understanding model interactions with the `teller` Dec 13, 2019

Using the `teller` on a classifier Dec 6, 2019

Benchmarking the querier's verbs Nov 29, 2019

Composing the querier's verbs for data wrangling Nov 22, 2019

Comparing and explaining model predictions with the teller Nov 15, 2019

Tests for the significance of marginal effects in the teller Nov 8, 2019

Introducing the teller Nov 1, 2019

Introducing the querier Oct 25, 2019

Prediction intervals for nnetsauce models Oct 18, 2019

Using R in Python for statistical learning/data science Oct 11, 2019

Model calibration with `crossval` Oct 4, 2019

Bagging in the nnetsauce Sep 25, 2019

Adaboost learning with nnetsauce Sep 18, 2019

Change in blog's presentation Sep 4, 2019

nnetsauce on Pypi Jun 5, 2019

More nnetsauce (examples of use) May 9, 2019

nnetsauce Mar 13, 2019

crossval Mar 13, 2019

test Mar 10, 2019