.png){kind=logo}
Statistics remains central to data science because machine-learning models, experiments and business decisions all depend on understanding uncertainty. This cheat sheet summarises the formulas, tests and interpretation rules data scientists use most frequently.
The mean is calculated by dividing the sum of all observations by the number of observations. It works well for approximately symmetric data but is sensitive to outliers. The median is the middle value and is generally more reliable for skewed variables such as income, transaction value, and delivery time.
Variance measures average squared deviation from the mean, while standard deviation expresses dispersion in the variable’s original unit. A low standard deviation indicates tightly grouped observations; a high value signals greater variability.
For approximately normal data, nearly 68% of observations fall within one standard deviation of the mean, 95% fall within two, and 99.7% fall within three.
Conditional probability measures the likelihood of event A after observing event B:
P(A|B) = P(A ∩ B) / P(B)
Bayes’ theorem updates an earlier probability when new evidence becomes available. It is particularly useful in fraud detection, medical diagnosis, spam filtering, and probabilistic machine learning.
Expected value represents the probability-weighted average of possible outcomes. It helps compare decisions involving uncertain returns, costs or risks.
A sample should represent the wider population without systematic bias. Increasing sample size generally reduces the standard error, which is calculated as:
Standard Error = Standard Deviation / √n
A 95% confidence interval is commonly estimated as the sample mean plus or minus 1.96 standard errors when normality assumptions are appropriate. However, a 95% interval does not mean there is a 95% probability that the fixed population parameter lies inside one calculated interval. It describes the long-run success rate of the estimation method.
The null hypothesis usually assumes no effect or difference. The alternative hypothesis represents the claim being investigated.
A p-value below 0.05 is often treated as statistically significant, but it does not measure effect size or the probability that the null hypothesis is true. Analysts should report confidence intervals and practical impact alongside p-values.
Use a t-test to compare two means, ANOVA for three or more means, a chi-square test for categorical relationships, and non-parametric tests when distributional assumptions are unsuitable.
Pearson correlation ranges from -1 to +1 and measures linear association. Correlation does not establish causation.
Linear regression estimates how a target variable changes with one or more predictors. Key evaluation metrics include R², adjusted R², MAE, MSE, and RMSE. Residuals should be checked for non-linearity, unequal variance, dependence and influential outliers.
Accuracy can be misleading with imbalanced data. Data scientists should compare precision, recall, specificity, F1-score and ROC-AUC. For rare-event problems such as fraud detection, precision-recall analysis is often more informative than accuracy alone.
Also Read: Best Python Tools for Large-Scale Data Analytics and Processing
Statistics gives data scientists the tools to summarize information, to measure uncertainty, and to check if patterns really are meaningful or just noise. When you pick the right statistical method and interpret the results beyond those p-values, things tend to get more stable. This kind of careful approach helps build more reliable models, run stronger experiments, and make better data-driven decisions.
Statistics helps data scientists understand datasets, identify meaningful patterns, and quantify uncertainty. It also supports model evaluation, experimentation, and evidence-based decision-making across machine learning projects.
Descriptive statistics summarize and describe the characteristics of a dataset using measures like mean and standard deviation. Inferential statistics use sample data to draw conclusions or make predictions about a larger population.
Hypothesis testing is useful when determining whether observed differences or relationships are statistically significant. It is commonly applied in A/B testing, product experiments, healthcare research and predictive model validation.
Accuracy can be misleading when datasets are highly imbalanced, such as fraud detection or disease diagnosis. Precision and recall provide a clearer picture of how well a model identifies positive cases while minimizing false predictions.
Every data scientist should understand descriptive statistics, probability, sampling methods, confidence intervals, hypothesis testing, correlation, regression and classification metrics. These concepts form the foundation for building, evaluating and interpreting machine learning models.
