An interactive online course on how an analyst gets from raw data to a trustworthy, explainable machine-learning model: data quality, tidy data and exploratory data analysis (EDA), predictive models and data leakage, model evaluation and validation, statistical evidence for comparing models, and explainable machine learning (SHAP, PDP, ICE). Built on real customer datasets (such as churn) with pandas and scikit-learn.
Price
19 900 Ft (one-off; seller is VAT-exempt)
Access
12 months from payment
Length
6 modules, about 13 hours
Contents
78 video clips, 23 quizzes, 39 code examples, 6 practice tasks, 6 decision scenarios
Level
Intermediate
Language
English materials, English and Hungarian captions
Format
online, self-paced, with Python exercises that run in the browser
University students who know basic Python and are learning data analysis and machine learning.
Analysts who want more reliable models and results they can defend.
Professionals who need to explain how a model works to decision-makers.
Prerequisites
Basic Python (variables, conditions, loops) and introductory statistics. If you are just starting to program, take “Introduction to Data Science and Programming” first.
Modules and learning outcomes
Week 1 — From Raw Data to Reproducible Analysis (120 min)
Decide whether a dataset is fit for modelling: data vs. information vs. knowledge, the pandas toolkit you need, a step-by-step data-quality investigation, a decision scenario and a hands-on practice task on a real messy customer dataset.
By the end you will be able to:
Distinguish a data-quality problem from a legitimate unusual observation.
Define and test simple data-quality rules from the meaning of the variables.
Quantify missing values and duplicates with pandas.
Check range, categorical and temporal consistency.
Explain why data quality affects downstream modelling.
Document cleaning decisions so another analyst can reproduce them.
Week 2 — Variables, Tidy Data and EDA as Hypothesis Generation (120 min)
Recognise kinds of data and variables, read an experimental design, put data into tidy form, and use exploratory data analysis to generate — not prove — testable hypotheses on a real customer-behaviour dataset.
By the end you will be able to:
Distinguish observational, experimental and virtual data, and nominal, ordinal, Boolean and numerical variables.
Identify dependent and independent variables, factors, levels and conditions in an experimental design.
Decide whether a table is tidy and name the operation that fixes it.
Choose a strategy for missing values that fits the meaning of the data.
Use EDA to generate rather than prematurely confirm hypotheses; separate observation, interpretation and hypothesis.
Recognise a possible confounder and turn an observation into one testable hypothesis.
Week 3 — From Clean Data to Predictive Models: Correlation, Prediction Error, Leakage and Validation (140 min)
Move from a clean dataset to a first predictive model: define X and y, separate prediction from causal analysis, measure association with correlation, understand how a model learns weights from training data and is judged by its prediction error on live data — and check whether a high score is trustworthy or the result of data leakage.
By the end you will be able to:
Identify the target y and the input variables x in a dataset and state whether a question is predictive or causal.
Explain conditioning, conditional and joint probability, covariance and correlation, and interpret a correlation coefficient.
Use proxy variables for latent concepts and name the kinds of prediction (quantitative, probability, classification, forecast).
Distinguish original (training) data from live data, and explain prediction error, model weights and loss functions such as MSE and RMSE.
Define the prediction moment and identify target, temporal, post-outcome and preprocessing leakage.
Choose a validation strategy (train/test, cross-validation, group-aware or temporal split) that reflects how the model will be used.
Week 4 — Model Evaluation and Validation (135 min)
Move from a business question to a trustworthy performance estimate: measure prediction error with MSE and RMSE, weigh errors by their consequences, recognise under- and overfitting, and build a fair evaluation of a churn model with a stratified split, a preprocessing pipeline, a dummy baseline, balanced accuracy and k-fold cross-validation.
By the end you will be able to:
Place evaluation and validation between model building and the business decision.
Compute prediction errors, MSE and RMSE, and explain why errors are squared.
Explain why over- and under-prediction (false positives and false negatives) can have different costs.
Recognise underfitting and overfitting and relate them to bias and variance.
Explain why accuracy misleads on an imbalanced target and compare a model with a dummy baseline using balanced accuracy.
Build a stratified train–test split and a preprocessing pipeline with imputation, scaling and one-hot encoding.
Use k-fold cross-validation on the training data to compare models, and keep the test set for the final evaluation.
Week 5 — Statistical Evidence for Data Scientists: Is the Difference Real? (140 min)
Understand how a classifier's threshold drives accuracy, precision, recall and F1, read ROC curves and AUC, and then ask the harder question: is one model's higher AUC real, or sampling variability? Compare a logistic regression and a random forest on a churn dataset using one split, repeated cross-validation, bootstrap intervals, a paired difference and a permutation test.
By the end you will be able to:
Explain how moving the classification threshold changes true/false positives and negatives, and why the right threshold depends on the cost of errors.
Compute and interpret accuracy, precision, recall and F1, and explain why accuracy misleads on imbalanced data such as churn.
Explain how a ROC curve is built from all thresholds and what the AUC summarises.
Recognise sampling variability in model evaluation and describe the sampling distribution, standard error and confidence intervals.
Compare models with repeated cross-validation and read the distribution of the B − A difference, not just two means.
Bootstrap a paired difference and interpret a permutation test against a no-signal null.
Distinguish statistical evidence from practical significance and write a short, uncertainty-aware model recommendation.
Week 6 — Explainable and Interpretable Machine Learning (130 min)
Move from 'how good is the model?' to 'what has the model learned?': distinguish interpretability from explainability and global from local explanations, compare impurity-based and permutation importance, read PDP, ICE and SHAP, and explain a churn model without turning association into causation.
By the end you will be able to:
Explain why accuracy-type metrics are not enough and what model explanations are used for.
Distinguish interpretability from explainability, intrinsic from post-hoc, and global from local explanations.
Compare impurity-based (MDI) and permutation feature importance and say why their rankings can differ.
Recognise how correlated predictors and random seeds make importance rankings unstable.
Interpret partial dependence, ICE curves and SHAP values as descriptions of model behaviour.
Write a stakeholder explanation that separates model behaviour from causal claims.
The Teringette learning framework
Every Teringette course follows the same proven arc, so you always know where you are and every step prepares you for the next:
Understand: Short video clips and concise concept explanations — only what you really need.
Step by step: Worked demonstrations that show how a professional solves a task.
Real decisions: Decision scenarios from real situations: what would you do, and what follows?
Practise: Code examples and practice tasks, with Python running right in your browser, nothing to install.
Check: Quizzes and automatically checked tasks with instant, explanatory feedback.
Consolidate: A checklist and a final assessment at the end of every module, with progress you can track.
About the author
Ádám Uzonyi — Programmer, educator, founder of Teringette
Ádám Uzonyi is a programmer and educator and the founder of Teringette. He teaches Python, Java and C++, runs AI training and builds custom AI-based software. He builds the Teringette Academy courses on his own Teringette learning framework, so that learning turns into real, usable skills rather than memorisation: from the concepts through step-by-step demonstrations and real decision situations to independent practice with instant feedback.
How much does the course cost and how long can I use it?
The course costs 19 900 Ft (Hungarian forints) as a one-off payment; the seller is VAT-exempt, so this is the total. Access lasts 12 months from payment, with no subscription and no automatic renewal.
How can I pay?
By card, Apple Pay or Google Pay on Stripe’s secure payment page, or by bank transfer. Buyers abroad can pay in their own currency. The invoice is sent by e-mail.
What language is the course in?
The texts and exercises are in English. The videos have English and Hungarian captions, and the interface is available in English and Hungarian.
Do I need to install anything?
No. The Python exercises run directly in your browser, tasks are checked instantly and your progress is saved. An up-to-date Chrome, Edge, Firefox or Safari is enough.
How long does it take?
The course has 6 modules, about 13 hours of study in total. You learn at your own pace and can continue where you left off.
Can I withdraw within 14 days?
This is digital content: when ordering you expressly ask for access to open immediately after payment and acknowledge that you thereby lose the 14-day right of withdrawal. If the material is faulty or does not work, your rights under the Terms remain.