Data Analytics — Explained Simply

Methods that summarise and describe the main features of a dataset — its centre, spread, and shape — without making inferences beyond the data.

Descriptive summarises the data you have; inferential uses a sample to draw conclusions about a larger population.

Mean (average), median (middle value), and mode (most frequent value).

For skewed data or with outliers — the median resists extreme values, while the mean gets pulled toward them.

The average squared deviation from the mean — a measure of how spread out the values are.

The square root of variance, expressing spread in the same units as the data, which makes it more interpretable.

Q3 − Q1, the spread of the middle 50% of the data; it's robust to outliers.

Minimum, Q1, median, Q3, and maximum — the basis of a box plot.

A measure of asymmetry; right-skew has a long right tail (mean > median), left-skew the opposite.

A measure of how heavy the tails are — high kurtosis means more extreme outliers than a normal distribution.

Common rules: beyond 1.5×IQR from the quartiles, or more than ~3 standard deviations from the mean (z-score).

The value below which a given percentage of observations fall — the 90th percentile is exceeded by only 10% of values.

Population covers everyone; a sample is a subset. Sample variance divides by n−1 (Bessel's correction) to stay unbiased.

Standard deviation divided by the mean — a unitless measure for comparing variability across different scales.

It's sensitive to outliers and skew; a single huge value can make a 'typical' figure unrepresentative.

Visualising the five-number summary and outliers, and comparing distributions across groups at a glance.

Nominal = unordered categories, ordinal = ordered categories, interval = ordered with equal gaps but no true zero, ratio = interval with a true zero.

The full shape of the distribution — modality, skew, and gaps that a mean and standard deviation alone hide.

Max − min; it's simple but driven entirely by the two most extreme values, so it's unstable.

A trend that appears in groups but reverses when the groups are combined — a warning to always segment before concluding.

A function describing how probabilities are spread across the possible values of a random variable.

Discrete take countable values (use a PMF); continuous take any value in a range (use a PDF where probability is area under the curve).

A symmetric bell-shaped distribution defined by its mean and standard deviation; it appears widely thanks to the CLT.

In a normal distribution, ~68% of data lies within 1 SD of the mean, ~95% within 2 SD, and ~99.7% within 3 SD.

The distribution of the number of successes in n independent yes/no trials with fixed success probability p.

Models the count of rare, independent events in a fixed interval, parameterised by an average rate λ.

Every value in a range is equally likely — a flat distribution.

Models the time between events in a Poisson process, e.g. wait times; it's memoryless.

Probability Density Function — for continuous variables, its area over an interval gives the probability of falling in that interval.

Cumulative Distribution Function — the probability that the variable is less than or equal to a given value.

A normal distribution with mean 0 and standard deviation 1; any normal can be converted to it via z-scores.

How many standard deviations a value is from the mean: (x − μ)/σ; it standardises values for comparison.

Sums/averages of many independent variables tend toward a normal distribution regardless of the original shape.

For positive, right-skewed quantities like incomes or response times, whose logarithm is normally distributed.

The long-run average of a random variable — the probability-weighted mean of its possible values.

The mean (location of the centre) and standard deviation (spread).

A Q-Q plot, a histogram, or formal tests like Shapiro-Wilk or Kolmogorov-Smirnov.

A single yes/no trial with success probability p — the building block of the binomial distribution.

Because the CLT makes sample means approximately normal, so normal-based tests are valid for reasonably large samples.

One with more extreme values than normal; it matters because rare large events (risk, fraud) are far more likely than a normal assumption suggests.

Selecting a subset of a population to estimate something about the whole, since measuring everyone is usually impractical.

The distribution of sample means approaches a normal distribution as sample size grows, regardless of the population's shape.

It justifies using normal-based confidence intervals and tests on sample means even when the underlying data isn't normal.

When the sample isn't representative of the population, systematically skewing results — e.g. surveying only website visitors.

Every member of the population has an equal chance of being selected, which minimises selection bias.

Dividing the population into subgroups (strata) and sampling within each to ensure representation of key groups.

Splitting the population into clusters, randomly selecting whole clusters, and sampling within them — cheaper for dispersed populations.

The standard deviation of a sample statistic (e.g. the mean); for the mean it's σ/√n and shrinks as n grows.

SD measures spread of the data; SE measures the precision of an estimate (like the sample mean).

Larger samples reduce standard error, giving more precise estimates and narrower confidence intervals (precision scales with √n).

Drawing conclusions only from cases that 'survived' a process, ignoring those that dropped out — distorting the picture.

Sampling whoever is easiest to reach; it's fast but highly prone to bias and rarely representative.

No — that's the point; the sample mean becomes approximately normal even from skewed populations, given enough samples.

A rule of thumb is n ≥ 30, but heavily skewed populations need larger samples.

When those who don't respond differ systematically from those who do, biasing survey results.

With replacement, a unit can be picked more than once; without replacement, each is chosen at most once (typical for surveys).

The distribution of a statistic (e.g. the mean) over all possible samples of a given size.

Increase sample size and use proper randomisation; note this addresses random error, not systematic bias.

Systematic error from how the sample is chosen, so the sample doesn't reflect the population.

Randomisation removes systematic selection patterns, making the sample representative and inference valid.

A range of plausible values for a population parameter, estimated from a sample, with a stated confidence level like 95%.

If you repeated the sampling many times, about 95% of the constructed intervals would contain the true parameter — not that there's a 95% chance for this one interval.

Estimate ± (critical value × standard error); for a 95% CI on a mean it's roughly mean ± 1.96 × SE.

The half-width of the interval — the critical value times the standard error.

Higher confidence level, greater data variability, and smaller sample size.

Larger samples shrink the standard error, producing a narrower, more precise interval.

Higher confidence means a wider interval; you trade precision for the assurance of capturing the true value.

When the population SD is unknown and the sample is small; the t-distribution's heavier tails account for the extra uncertainty.

If a 95% CI for a difference excludes zero, the result is significant at the 5% level — they're two views of the same inference.

Not in frequentist terms — the parameter is fixed; the interval is random. That probabilistic phrasing belongs to Bayesian credible intervals.

The Bayesian analogue — a range that contains the parameter with a stated probability given the prior and data.

The multiplier (e.g. 1.96 for 95% normal) from the relevant distribution that sets the interval's width for a confidence level.

p̂ ± z × √(p̂(1−p̂)/n), or better methods (Wilson) for small samples or extreme proportions.

It communicates the uncertainty around the estimate, preventing false precision.

High variability increases the standard error, widening the interval and signalling a less precise estimate.

No — it honestly reflects more uncertainty (small sample or high variance); a misleadingly narrow interval is worse.

Roughly: random sampling, independent observations, and approximate normality of the sampling distribution (via CLT).

You need about four times the sample size, since precision improves with √n.

Resampling the data with replacement many times to empirically build the sampling distribution and read off interval bounds — no distribution assumption needed.

Yes — overlapping intervals don't guarantee non-significance; you should test the difference directly.

An approach that updates the probability of a hypothesis as new evidence arrives, combining prior belief with observed data.

P(H|E) = P(E|H)·P(H) / P(E): posterior is proportional to likelihood times prior.

Your belief about a parameter before seeing the new data, expressed as a probability distribution.

The updated belief about the parameter after combining the prior with the observed data.

The probability of the observed data under a given value of the parameter.

Frequentists treat parameters as fixed and data as random; Bayesians treat parameters as random with a probability distribution updated by data.

A prior that yields a posterior of the same family (e.g. Beta prior with binomial likelihood → Beta posterior), making updates analytic.

Informative encodes real prior knowledge; uninformative (flat) expresses little, letting the data dominate.

It directly reports the probability that the variant beats control and the expected loss, instead of a p-value.

A range that contains the parameter with a stated probability given the data and prior — the Bayesian counterpart to a confidence interval.

They're subjective; a strong, poorly-chosen prior can bias conclusions, especially with little data.

It concentrates around the true value and the prior's influence fades — data overwhelms the prior.

Ignoring the prior probability (base rate), e.g. overestimating disease likelihood from a positive test when the disease is rare.

Even with a 99%-accurate test, a positive result for a rare disease often means a low actual probability of disease because the base rate is tiny.

P(E), the total probability of the data across all hypotheses; it normalises the posterior.

Markov Chain Monte Carlo samples from complex posteriors that can't be computed analytically.

The most probable parameter value under the posterior — the Bayesian analogue of a point estimate.

With small data, when prior knowledge exists, or when you need direct probability statements for decisions.

It applies Bayes' theorem with a conditional-independence assumption to compute the posterior probability of each class.

Posterior odds = prior odds × likelihood ratio (Bayes factor); the data shifts the odds by the strength of evidence.

A procedure for deciding whether sample evidence is strong enough to reject a default assumption about a population.

The default 'no effect / no difference' statement that the test tries to disprove (H₀).

The claim you suspect is true — that there is an effect or difference (H₁).

The probability of observing data at least as extreme as yours if the null hypothesis were true.

The result would occur less than 5% of the time under the null, so it's deemed statistically significant at the 5% level.

It is NOT the probability the null is true, nor the probability the result is due to chance — it's conditional on the null being true.

The threshold for rejecting the null and the accepted false-positive rate, set before testing (commonly 0.05).

A false positive — rejecting a true null hypothesis; its rate is α.

A false negative — failing to reject a false null hypothesis; its rate is β.

1 − β, the probability of correctly detecting a real effect; typically targeted at 80%.

Larger sample size, bigger true effect, lower variance, and a higher α.

One-tailed tests for an effect in a specific direction; two-tailed tests for any difference and is the safer default.

To compare the means of one or two groups when the population standard deviation is unknown.

To test association between two categorical variables, or goodness-of-fit to expected counts.

Analysis of Variance — tests whether the means of three or more groups differ, avoiding inflated error from many pairwise t-tests.

Running many tests inflates the chance of a false positive; corrections like Bonferroni or FDR control it.

Dividing α by the number of tests to keep the overall false-positive rate in check (conservative).

A result can be statistically significant but too small to matter in practice; always look at effect size, not just the p-value.

A measure of the magnitude of a difference (e.g. Cohen's d), independent of sample size.

A standardised number (t, z, χ², F) summarising how far the data is from the null, mapped to a p-value.

A randomised controlled experiment comparing two versions (control A and variant B) to see which performs better on a chosen metric.

It balances confounders across groups so any measured difference can be attributed to the change, not pre-existing differences.

The group that sees the existing version, providing the baseline to compare the variant against.

Pick one metric that reflects the goal and is sensitive to the change; too many metrics invite cherry-picking.

The smallest effect you care to detect; smaller MDEs require larger samples.

From the baseline rate, the MDE, the significance level (α), and desired power (usually 80%).

Repeatedly checking results and stopping when significant inflates false positives; fix with a fixed horizon or sequential testing.

A temporary behaviour change just because something is new, which can fade — run the test long enough to see past it.

A metric you don't want to harm (e.g. latency, revenue) monitored alongside the primary metric.

An overall winner can lose within every segment (or vice versa); segment the results to be sure.

The probability the test detects the MDE if it truly exists; underpowered tests miss real effects.

Compare the metric with an appropriate test (e.g. proportions z-test), report the lift with a confidence interval, and check guardrails.

Verifying the actual split matches the intended (e.g. 50/50); a mismatch signals a broken experiment to be discarded.

Running two identical versions to validate the experiment setup — it should show no significant difference.

Early significance is often noise; stopping on it (peeking) greatly inflates the false-positive rate.

Testing multiple element combinations at once to find interactions, requiring much larger samples than a simple A/B test.

When one user's treatment affects another's outcome (social apps), breaking independence; cluster-level randomisation helps.

At least one full business cycle (often a week or two) to cover weekday/weekend behaviour and reach the planned sample size.

Absolute lift is the raw difference (4% → 5% = +1pp); relative lift expresses it as a percentage of baseline (+25%).

Not that there's no effect — only that you couldn't detect one; it may be too small or the test underpowered.

A statistical measure of how two variables move together, ranging from −1 (perfect inverse) to +1 (perfect direct).

A measure of the strength and direction of a linear relationship between two continuous variables.

A rank-based correlation capturing monotonic relationships, robust to outliers and non-linearity.

Because a third variable (confounder), reverse causality, or pure coincidence can produce correlation without a causal link.

A variable that influences both the supposed cause and effect, creating a misleading association.

Ice-cream sales and drowning deaths correlate because both rise with summer heat — the confounder — not because one causes the other.

Best via a randomised controlled experiment; otherwise with causal-inference methods that control for confounders.

When the assumed effect actually causes the assumed cause — e.g. 'support tickets correlate with churn' may run either direction.

No linear relationship — but a strong non-linear relationship (e.g. U-shaped) can still exist.

−1 to +1; the sign gives direction and the magnitude gives strength.

Anscombe's quartet shows datasets with identical correlations but wildly different shapes; a scatter plot reveals the truth.

Stratification, multivariable regression, matching, or propensity scoring to adjust for known confounders.

Association means a statistical relationship exists; causation means changing one variable would change the other.

Situations where an external event randomly assigns treatment, enabling causal estimates without a designed experiment.

If the groups compared differ systematically from the start, any difference can be wrongly attributed to the treatment.

What would have happened to the same unit without the treatment — unobservable, so we estimate it via controls or randomisation.

Another term for an unmeasured confounder that drives an observed correlation.

Simple linear regression's slope is directly related to Pearson's correlation; both summarise a linear relationship.

It isolates the relationship of interest from confounders, moving an analysis closer to a causal claim.

The gold standard for causation: random assignment to treatment/control balances confounders so differences are causal.

A method to model and quantify the relationship between a dependent variable and one or more independent variables.

The expected change in the outcome for a one-unit increase in that predictor, holding the others constant.

Simple uses one predictor; multiple uses several, allowing you to control for confounders.

The proportion of variance in the outcome explained by the model, from 0 to 1.

It penalises adding predictors, so it only increases when a new variable genuinely improves the model.

High correlation among predictors that destabilises coefficient estimates; detect it with the Variance Inflation Factor.

Linearity, independent errors, homoscedasticity, normally distributed residuals, and little multicollinearity.

Constant variance of the residuals across all fitted values; violating it is heteroscedasticity.

It tests whether that predictor's effect differs from zero; small p means the relationship is statistically significant.

The differences between observed and predicted values; their patterns reveal model problems.

Categorical variables are encoded as 0/1 indicators, with one category dropped as the reference baseline.

A product of two predictors that lets the effect of one depend on the level of another.

Linear predicts a continuous outcome; logistic predicts the probability of a binary outcome via the sigmoid.

Non-linearity, heteroscedasticity, and outliers — ideally it shows random scatter around zero.

Adding too many predictors so the model fits noise; it shows high training R² but poor out-of-sample performance.

Add polynomial or transformed terms (log, sqrt), or use a non-linear model.

Penalising large coefficients to reduce overfitting; Lasso can also zero out weak predictors for selection.

Predicting outside the range of the training data, where the fitted relationship may not hold.

An observation that disproportionately changes the fit; detect with leverage and Cook's distance.

Check assumptions via residual plots and assess predictive accuracy on held-out data (cross-validation), not just R².

The initial investigation of data to understand its structure, spot anomalies, test assumptions, and form hypotheses before modelling.

Understand distributions and relationships, find data-quality issues, detect outliers, and generate hypotheses.

Examining one variable at a time — its distribution, central tendency, and spread (histograms, box plots).

Examining the relationship between two variables, e.g. with scatter plots, cross-tabs, or correlation.

Quantify them, understand why they're missing, then decide to drop, impute, or flag — never ignore them silently.

Missing Completely At Random, Missing At Random (explained by other variables), and Missing Not At Random (related to the missing value itself) — each needs different handling.

Box plots, z-scores, the IQR rule, and scatter plots; then investigate whether they're errors or genuine extremes.

Quickly spotting strongly related variables and potential multicollinearity across many features.

Summary stats hide shape; plots reveal skew, multimodality, and gaps that change how you analyse the data.

Duplicates, impossible values, inconsistent categories, wrong types, and out-of-range entries.

Frequency counts, bar charts, and cross-tabulations against the target.

EDA surfaces patterns and relationships that suggest useful new features (ratios, bins, date parts).

Aggregating a metric across segments to compare behaviour and uncover where differences live.

Look for features that wouldn't be available at prediction time or that suspiciously predict the target perfectly.

A grid of scatter plots for every pair of variables, with distributions on the diagonal — a fast multivariate overview.

Aggregates can hide or reverse patterns (Simpson's paradox); segmenting reveals the real story.

Shape, column types, a sample of rows, missing-value counts, and basic summary statistics.

Tools like ydata-profiling generate distributions, correlations, and quality warnings in one report to speed exploration.

EDA is open-ended discovery; confirmatory analysis formally tests pre-specified hypotheses.

They justify cleaning and modelling decisions and prevent others (or future you) from repeating the investigation.

Measuring how many users progress through sequential steps of a process to find where the biggest drop-offs occur.

Grouping users by a shared start characteristic (e.g. signup month) and tracking a metric across their lifetime.

Overall metrics mix old and new users; cohorts isolate how each group behaves over time, revealing real trends.

The percentage of users moving from one step to the next, or from the top to the bottom of the funnel.

Identify the step with the largest drop-off relative to its potential impact and focus there.

Tracking what fraction of each signup cohort remains active in week/month 1, 2, 3… to see retention shape over time.

Acquisition cohorts group by when users joined; behavioural cohorts group by an action they took.

Usually a tracking issue (double counting, users entering mid-funnel) — a signal to audit the event data.

They separate 'new users are worse' from 'the whole product is declining' by isolating each group's trajectory.

Measuring how long users take to move through the funnel, revealing friction and natural decision lags.

One where many users drop between steps; plugging the biggest leak yields the largest conversion gain.

Count distinct users at each event step (often with conditional aggregation or sequential joins on timestamps).

A stable core of loyal users — the plateau height is your realistic long-term retention ceiling.

Conversion often varies by device, channel, or geography; segmenting reveals where to focus fixes.

Closed funnels require strict step order; open funnels allow users to enter or skip steps, affecting how you count.

A triangular heatmap (cohorts as rows, periods as columns) coloured by the retention/metric value.

Newer cohorts have less observed time, so comparing them to older ones at the same age — not calendar date — avoids bias.

'Where are we losing users, and which step should we fix first?'

'Are users acquired recently better or worse than before, and how does engagement evolve?'

Compare funnel conversion across cohorts to see whether product changes improved progression for newer users.

A customer-segmentation technique scoring each customer on Recency, Frequency, and Monetary value.

How recently a customer last purchased — recent buyers are more likely to buy again.

How often a customer purchases within a period — frequent buyers are more engaged.

How much a customer spends in total — high spenders are most valuable.

Each dimension is split into quantiles (e.g. 1–5), then combined into a score or segment label.

Different groups need different treatment; targeting beats blanket campaigns on both cost and conversion.

Customers high on all three — recent, frequent, high-value — ideal for loyalty and advocacy programs.

Previously valuable customers who haven't purchased recently — prime targets for win-back campaigns.

RFM is rule-based and interpretable; K-Means clusters on many behavioural features for richer but less transparent segments.

The total value a customer is expected to generate over their relationship — guides acquisition spend.

Distance-based clustering is dominated by large-range features (like monetary), so standardising is essential.

Balance statistical signals (elbow/silhouette for clustering) with how many groups the business can actually action.

Segments that are distinct, sizable, stable, and map to a clear, different action.

Grouping by what users do (usage patterns, features used) rather than who they are demographically.

Each segment gets a tailored message — VIP perks for champions, reactivation offers for lapsed high-spenders.

A transaction log with customer ID, purchase date, and amount.

It's backward-looking and ignores product mix, channel, and engagement beyond purchases.

Profile each segment, check they differ on outcomes, and test whether targeted actions actually lift results.

Often ~80% of revenue comes from ~20% of customers — segmentation finds and protects that 20%.

Regularly (e.g. monthly), because customers move between segments as their behaviour changes.

The rate at which customers stop using a product or cancel a subscription over a period.

The fraction of customers who continue to use the product over time — the inverse of churn.

Acquiring a new customer costs far more than keeping one, so small retention gains compound into large revenue.

With a clear, product-specific action and time window (e.g. logged in within 30 days) — the definition shapes every metric.

A plot of the percentage of a cohort still active at each period after signup; its plateau is the loyal core.

Voluntary is a deliberate cancel; involuntary is failed payments or expired cards — each needs a different fix.

Gross counts lost revenue only; net subtracts expansion (upsells) and can even be negative if expansion outpaces losses.

Train a classifier (e.g. logistic regression or gradient boosting) on behavioural and account features to score churn risk.

Declining usage/engagement, support tickets, time since last activity, tenure, and plan/price changes.

With precision/recall, ROC-AUC or PR-AUC (data is imbalanced), and the business lift from acting on the scores.

Churn is imbalanced; predicting 'no churn' for everyone can be highly accurate yet useless.

Logo (customer) churn counts accounts lost; revenue churn weights by the money those accounts represented.

Improve onboarding, target at-risk users with interventions, fix failed payments, and address the top drivers found in analysis.

The share of users active exactly N days after signup (e.g. Day-7 retention).

Counting a user as retained if active on or after day N, which is more forgiving than exact-day retention.

Comparing churn across signup cohorts to see whether product changes improved retention for newer users.

It models time-until-churn (e.g. Kaplan-Meier, Cox), handling customers who haven't churned yet (censoring).

Customers lost in a period divided by customers at the start of that period.

Prioritise high-risk, high-value customers for targeted retention offers, balancing intervention cost against saved value.

When expansion revenue from existing customers exceeds revenue lost to churn — a strong growth signal.

A Key Performance Indicator — a metric chosen to track progress toward a specific business objective.

Every KPI is a metric, but a KPI is specifically tied to a goal and watched closely; not every metric is a KPI.

A number that looks impressive (total signups, page views) but doesn't inform decisions or reflect real value.

One that ties to a decision — when it moves, you know what to do about it.

The single metric that best captures the core value your product delivers, aligning the whole team.

Leading indicators predict future outcomes (e.g. trial usage); lagging indicators confirm past results (e.g. revenue).

To ensure improving the target doesn't quietly harm something else (e.g. boosting engagement while hurting retention).

Specify the numerator, denominator, time window, and population so it's unambiguous and reproducible.

'When a measure becomes a target, it ceases to be a good measure' — people optimise the number, not the goal.

Aggregates hide divergent group behaviour (Simpson's paradox); segmenting reveals what's really driving the number.

Counts grow with scale and can mislead; ratios (rates, per-user) normalise and are more comparable over time.

Acquisition, Activation, Retention, Referral, Revenue — a funnel framing for product/growth KPIs.

Google's UX metrics: Happiness, Engagement, Adoption, Retention, Task success.

Pick the leading metric most correlated with long-term value and that the team can influence — not raw revenue or signups.

Another name for a guardrail — a metric watched to prevent unintended harm while optimising the primary KPI.

Without defining the action and window, DAU/MAU can be inflated or inconsistent across teams.

Daily over Monthly Active Users — a stickiness measure of how often monthly users return.

Few — one North Star plus a handful of supporting and guardrail metrics; too many dilute focus.

It's relevant to the goal, clearly defined, comparable over time, hard to game, and tied to action.

Output measures activity (features shipped); outcome measures impact (users helped) — outcomes matter more.

A sequence of data points indexed in time order, such as daily sales or hourly temperature.

Trend (long-term direction), seasonality (regular cycles), and residual/noise (random variation).

The long-term increase or decrease in the series, independent of short cycles.

Patterns that repeat at fixed periods, like higher retail sales every December.

Seasonality has a fixed known period; cycles are longer, irregular fluctuations (like economic booms) with no fixed length.

A series whose statistical properties (mean, variance) don't change over time — required by many models.

Visually, plus tests like the Augmented Dickey-Fuller (ADF) or KPSS test.

Differencing, removing trend/seasonality, or transforming (e.g. log) to stabilise variance.

Subtracting the previous value from the current to remove trend and help achieve stationarity.

The correlation of a series with its own past values at various lags.

Autocorrelation and Partial Autocorrelation Functions — plots used to identify the lag structure and choose ARIMA orders.

Additive when seasonal swings are constant in size; multiplicative when they grow proportionally with the level.

Averaging over a sliding window to smooth noise and reveal the underlying trend.

It leaks future information into training; you must split chronologically to respect time order.

A past value of the series used as a predictor for the current value.

A statistic (mean, std) computed over a moving window, used for smoothing and feature engineering.

A series of uncorrelated, zero-mean, constant-variance values — what residuals should look like after good modelling.

Time series tracks one entity over time; cross-sectional captures many entities at a single point in time.

Ignoring it makes you mistake normal cycles for real change and produces poor forecasts.

A sudden, lasting shift in the series' behaviour (e.g. a policy change) that breaks a single fitted model.

Predicting future values of a time series, ideally with a quantified uncertainty range.

A point forecast hides risk; an interval communicates how confident the forecast is and supports planning.

Predicting the next value equals the last observed value — a baseline every model must beat.

Forecasting with weighted averages that give more weight to recent observations; Holt-Winters extends it to trend and seasonality.

AutoRegressive Integrated Moving Average — combines autoregression, differencing for stationarity, and moving-average terms.

p = autoregressive lags, d = differencing order, q = moving-average lags.

ARIMA with added seasonal terms to capture repeating seasonal patterns.

Use time-based splits or rolling/expanding-window backtesting — never random splits.

MAE, RMSE, MAPE, and sMAPE; MAPE is intuitive but breaks near zero values.

Repeatedly training on data up to time t and testing on t+1, advancing through the series to mimic real use.

For business series with strong seasonality, holidays, and missing data, where you want fast, robust forecasts with little tuning.

Yes — gradient boosting or neural nets on lag/rolling/calendar features, but you must prevent leakage with proper time splits.

A model that fits historical noise and seasonal quirks but generalises poorly to the future.

If a complex model can't beat 'use last value' or 'same as last year', it isn't adding value.

Forecasting specifically projects future values of a time-ordered series; prediction is the general term for any model output.

They cause spikes/dips that regular seasonality misses; modelling them explicitly improves accuracy.

How far ahead you predict; accuracy generally degrades as the horizon lengthens.

The shaded range showing plausible future values; it widens with the horizon as uncertainty grows.

Detect the break, down-weight or drop pre-change data, or use adaptive models that adjust to the new level.

Combining multiple models' forecasts (e.g. averaging) to reduce variance and often beat any single model.

The process of fixing or removing incorrect, incomplete, duplicated, or inconsistent data to make it analysis-ready.

Analysis is only as good as the data; errors lead to wrong conclusions — garbage in, garbage out.

Drop rows/columns, impute (mean/median/mode/model-based), or flag missingness as its own signal.

Drop when missingness is tiny or the column is mostly empty; impute when the data is valuable and missingness isn't informative.

Check for exact or fuzzy matches on key fields; decide whether duplicates are true repeats or data-entry errors.

Standardise casing/spelling and map variants ('IN', 'India', 'india') to a single canonical value.

An extreme value; investigate whether it's an error (fix/remove) or genuine (keep, maybe cap/transform).

Converting columns to correct types (dates, numbers, categories) so operations behave correctly.

Dates stored as strings can't be sorted, differenced, or grouped by period correctly until parsed to datetime.

Wide spreads variables across columns; long stacks them into key-value rows — many tools and plots expect long (tidy) data.

Data where each variable is a column, each observation a row, and each cell a single value.

Cap (winsorise), transform (log), or use robust methods that down-weight extremes.

Checking data against rules (ranges, types, uniqueness) to catch errors early and automatically.

A documented, scripted pipeline lets others rerun it, audit decisions, and apply the same steps to new data — unlike manual edits.

It can distort distributions and understate uncertainty if done carelessly (e.g. mean-imputing a skewed variable).

Trimming whitespace, fixing case, and removing special characters so text values match consistently.

Detect and convert all values to a single unit before any calculation.

Often ~80% is spent finding, cleaning, and preparing data, and ~20% on the actual analysis.

Documentation defining each field's meaning, type, and valid values — essential for correct cleaning and analysis.

Fix typos, inconsistent naming, and mislabeled categories during a standardisation step before analysis.

To communicate insight clearly and quickly — making patterns and comparisons obvious to the audience.

Match it to intent: bar for comparison, line for trend over time, scatter for relationship, histogram/box for distribution.

When there are many slices or similar sizes — humans compare angles poorly; a sorted bar chart is clearer.

Tufte's idea of maximising the ink that conveys data and minimising decorative 'chartjunk'.

Starting the axis above zero exaggerates small differences, distorting the reader's perception.

A histogram or box plot (or violin plot) — to reveal shape, spread, and outliers.

A scatter plot, optionally with a trend line.

A line chart, which emphasises trend and continuity.

Purposefully — to encode data or highlight, not decorate — and accessibly for colour-blind viewers.

Unnecessary visual elements (3D effects, heavy gridlines, clip art) that distract from the data.

Direct labels reduce the back-and-forth of decoding a legend, speeding comprehension.

Showing magnitude across a matrix — e.g. correlations or a cohort retention grid.

They can imply relationships that don't exist by arbitrarily scaling two axes; use with care.

Stacked bars or a treemap for several categories; pie only for a few large, distinct slices.

Visual features (colour, size, position) the brain perceives instantly — use them to direct attention to the key point.

Sorting by value reveals ranking immediately, unlike alphabetical order which hides it.

Colour-blind-safe palettes, sufficient contrast, text labels, and not relying on colour alone to convey meaning.

A grid of small charts with the same axes, letting you compare many categories at a glance.

Use transparency, sampling, binning (hexbin), or density contours when points overlap heavily.

'What is the one message I want the audience to take away?' — the chart should serve that.

Most business data lives in relational databases/warehouses, and SQL is the standard way to retrieve and aggregate it.

FROM → WHERE → GROUP BY → HAVING → SELECT → ORDER BY → LIMIT (not the written order).

WHERE filters rows before aggregation; HAVING filters groups after aggregation.

Combining rows from two tables based on a related column.

INNER keeps only matching rows; LEFT keeps all left-table rows, filling unmatched right columns with NULL.

A function that computes across a set of rows related to the current row without collapsing them (e.g. running totals, ranks).

ROW_NUMBER, RANK, DENSE_RANK, LAG, LEAD, and aggregate windows like SUM() OVER(...).

RANK leaves gaps after ties (1,2,2,4); DENSE_RANK doesn't (1,2,2,3).

A Common Table Expression (WITH clause) — a named, readable subquery that structures multi-step logic.

Functionally similar, but CTEs are more readable, reusable within a query, and can be recursive.

SUM(value) OVER (ORDER BY date ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW).

Aggregates rows that share values in specified columns, enabling per-group summaries.

Use IS NULL/IS NOT NULL, COALESCE for defaults, and remember NULL breaks normal comparisons and some aggregates.

COUNT(*) counts all rows; COUNT(column) counts non-NULL values in that column.

Use ROW_NUMBER() OVER (PARTITION BY group ORDER BY metric DESC) and filter to rn <= N.

Restarts the window calculation for each partition (group), like a per-group running calculation.

UNION removes duplicate rows; UNION ALL keeps them and is faster.

Count distinct users at each step using conditional aggregation or sequential timestamp joins, then compute step conversions.

Accessing a previous row's value — e.g. computing period-over-period change without a self-join.

Filter early, select only needed columns, use indexes/partitions, avoid unnecessary DISTINCT, and pre-aggregate where possible.

A visual display of key metrics that lets stakeholders monitor performance and answer questions themselves.

Understand the audience and the decisions they need to make; design around their key questions.

Arranging elements so the most important metric draws the eye first, typically top-left.

Viewers scan briefly; the most important number must be instantly visible before any detail.

Filters and drill-downs that let users answer follow-up questions without asking the analyst.

Limit the number of charts, remove chartjunk, group related items, and use whitespace deliberately.

Operational track real-time day-to-day metrics; strategic track high-level KPIs over longer horizons for leadership.

If 'active user' means different things in different tiles, the dashboard loses trust and creates confusion.

Big-number KPIs, trend lines, ranked bars, and simple tables — clarity over novelty.

Only metrics that are actionable and tied to the audience's goals; everything else is noise.

They let one dashboard serve many questions by slicing by date, region, segment, etc.

Tableau, Power BI, Looker, Metabase, and Google Data Studio (Looker Studio).

Pre-aggregate data, limit live queries, cache where possible, and avoid overly granular visuals.

Automated notifications when a metric crosses a threshold, so issues are caught without constant watching.

Provide a high-level summary view with the option to drill into detailed views per role.

A number alone is meaningless; comparison to a goal, prior period, or benchmark tells whether it's good or bad.

Cramming in every available metric ('chart soup') instead of focusing on the few that drive decisions.

Match the decision cadence — real-time for operations, daily/weekly for strategic reporting.

Reconcile its numbers against a trusted source, test edge cases, and review definitions with stakeholders.

A report is usually static and detailed; a dashboard is live, visual, and focused on monitoring key metrics.