A statistics-first review of an A/B test that goes past a single p-value before you ship the result.
- Inputs: control and treatment group results, primary metric, test duration, and optional segments
- Checks: post-hoc power, statistical significance with confidence interval, absolute and relative effect size
- Flags: novelty-effect risk, sample ratio mismatch, peeking problems, and multiple-comparison issues
- Output: a ship/kill/keep-testing call with the reasoning behind it
Goes past a raw correlation matrix to flag the relationships in a dataset that actually matter before modeling starts.
- Inputs: key variables of interest plus dataset columns, types, and row count
- Analysis: correlation matrix, non-obvious pairings, multicollinearity risk, categorical-to-numeric associations, and time-lagged correlations
- Warns: spurious correlations that look causal but aren't
- Output: plain-language findings plus runnable Python code
Walks from a target variable, candidate predictors, dataset size, and analysis purpose through a full regression build: exploratory relationships, feature selection, model choice with reasoning, assumption checks, and performance metrics like R², RMSE, or AUC depending on the target type. Delivers plain-language coefficient interpretation, residual analysis, and sample predictions with confidence intervals, plus complete scikit-learn or statsmodels code.
Forecasts a chosen metric over a set horizon from historical data, granularity, known patterns, and external factors, with a justified choice of model — moving average, exponential smoothing, ARIMA/SARIMA, or Prophet.
Delivers 80% and 95% confidence intervals, a trend/seasonality/residual decomposition, holdout accuracy metrics like MAPE and RMSE, and optimistic, baseline, and pessimistic scenarios, so the forecast carries honest uncertainty bounds instead of a single line.
Examines how key metrics actually moved over a date range and granularity before anyone tries to forecast them.
Covers overall trend direction, seasonality at weekly through annual scales, anomalies or change points, day-of-week patterns, and year-over-year growth rates, reporting the time window and magnitude behind every finding so it feeds directly into an anomaly report or forecast.
Feed in a target variable, its type, a list of candidate predictor columns,
and the dataset's row and column counts, and get back a rigorous
drivers-of-X analysis instead of a hand-wavy guess. The prompt runs
univariate correlation, mutual information, tree-based feature importance,
and permutation importance side by side, explains why the four rankings
disagree, flags features that are likely proxies for the same underlying
driver, and closes with a plain-language summary naming the top three
drivers and what they mean for the business.
List the columns with missing data — name, type, approximate percent
missing, and likely cause — and this prompt classifies each one as MCAR,
MAR, or MNAR before recommending a fix.
It weighs whether the missingness itself is informative, considers adding
a was_missing flag, justifies the chosen imputation method against
alternatives for every column, and outputs the Python code implementing
the strategy column by column.
Give the numeric columns to check and describe what the dataset
represents, and this prompt detects outliers with IQR, Z-score, and
isolation forest in parallel, then classifies each one as a data-entry
error, a genuine extreme value, or part of a systemic cluster pointing to
an upstream process bug. It shows each outlier alongside its row context,
quantifies how much removing them shifts the mean and median, and
generates Python code that flags the outliers and outputs a cleaned
dataset with the recommended handling applied.
Describe the research question, the two groups being compared with their
sample sizes, the outcome variable, and the significance level, and this
prompt runs a complete hypothesis test end to end.
- States the null and alternative hypotheses in plain language
- Checks the chosen test's assumptions (normality, variance, independence)
- Selects and justifies the right test, reporting the statistic, p-value
- Reports a confidence interval alongside the p-value
- Computes an effect size and comments on practical significance
- Delivers the full Python/R workflow as code