shape, dtypes, duplicates, describe, value counts — problems named, not treated

a repeating entity ID inflates every metric under a random split

histogram-boxplot and bar plot per variable, target included, no-information baseline stated

heatmap first, then target relationships; a null result stated as confidently as a strong one

strong predictors and weak relationships, each carrying its number

drop, encode, split, then treat — reversing the last two invalidates every metric

only with a stated reason; a weak bivariate gap is not grounds

dummies by default; ordinal and high-cardinality columns get a decision, not a default

stratified holdout, grouped where an entity repeats, chronological for a series

options and costs presented, parameters learned on train only

anything that learns a number from the data is fit on train alone

RMSE, MAE, R², adjusted R², MAPE, with three silent-failure guards

accuracy, precision, recall, F1, ROC-AUC, all read against class balance

the cost of a false negative against a false positive, asked not assumed

a test figure alone hides overfitting; the gap between them is the finding

train on the earlier period; a shuffled split trains on the future

line plot, rolling statistics, decomposition, ADF with its null stated

explicit shifts, verified by the pattern of leading NaNs

last observed value, or one season back, reported beside every model

reasoned from principle, not validated against a finished project

agreed once, opening with a simple baseline the rest has to beat

fit, evaluate on train and test, read the gap rather than the two numbers

total fits stated out loud before any grid search is proposed

scored on the agreed metric, judged on whether the train/test gap shrank

top features reconciled explicitly against what EDA showed

the 1% rule and metric conflict, named rather than defaulted

restated as a point-weighted checklist, divergences named

"with comments" means a written decision; separately listed means separately marked

unused imports, mid-notebook imports, cells that ran out of order

explain, structure, debug — the decisions stay with the person submitting

one unit of work, then stop and wait for the real output

each finding its own headed cell, grounded only in what was reported

target column first, then regression, classification, or time series

stop after EDA when there was no modelling intent; join at the split with clean data

paste-ready notebook cells, or code run directly and interpreted as it goes

rebuilt at each section boundary, so there is something readable throughout