Time-Series Cross-Validation

Time-series data cannot use random K-Fold; walk-forward (rolling-window) validation must be used instead: the training set contains only past time steps and the validation set is the immediately following future window — step 1 trains on $[0, t]$ and validates on $[t, t+h]$, step 2 rolls to train on $[0, t+s]$ and validate on $[t+s, t+s+h]$, and so on until the data is covered. The iron rule is no future information: any preprocessing computed with future statistics (means, targets, global normalization parameters) causes temporal leakage and inflates validation scores. For hyperparameter tuning, nested CV is used: an outer rolling loop for model evaluation and an inner loop for hyperparameter search, avoiding selection bias.

forecasting stocks/sales/traffic, risk control and recommendation — anything with timestamps; interviews ask “how to CV time series”, “why random K-Fold overestimates”, and “which future information must be blocked”.

random K-Fold mixes future samples into training so the model “peeks” into the future, inflating CV scores that collapse in production; walk-forward mimics the real online setting (every step predicts using only history) and gives an honest generalization estimate — it sacrifices some training data (early windows are small) in exchange for temporal integrity that matches deployment.

Why does random K-Fold fail on time series? Trace the concrete leakage path (e.g. predicting t+1 with samples from t+1 onward)?

Describe the walk-forward rolling-window procedure: training span, validation span and step size for step k?

How to normalize, impute and target-encode in time-series modeling without leakage (history-only statistics)?

What is nested CV? What does each loop do, and which bias does it fix?

How does walk-forward differ from plain K-Fold in bias-variance and cost? When is the data loss acceptable?