Contrast walk-forward analysis with an expanding window against a rolling fixed-length window for model selection on ten years of daily data. State what each assumes, what each biases, and how you would decide between them.

Contrast walk-forward analysis with an expanding window against a rolling fixed-length window for model selection on ten years of daily data. State what each assumes, what each biases, and how you would decide between them.

Approach: Ask what each scheme assumes about the stability of the data generating process, and what happens to sample size and to stale data under each.

The expanding window assumes the relationship is stable over the whole history and buys estimation precision at the cost of carrying stale regimes, while the rolling window assumes only local stability and buys adaptation at the cost of a smaller sample and a noisier fit, so the walk-forward choice is decided empirically by whether performance degrades with the age of the training data. Under the expanding scheme the training set grows and parameter variance falls roughly as 1/n, so late folds are estimated far more precisely than early ones and the aggregated out-of-sample statistic mixes two very different estimation regimes. It also means a structural break ten years ago is still weighted equally with last month. The rolling scheme holds the training length fixed so every fold is comparable, which makes the sequence of out-of-sample results a usable time series in its own right, and it lets a change in market structure wash out after the window length. The diagnostic that decides the question is to fit at several window lengths and plot out-of-sample performance against window length: a peak at an interior length is evidence of a limited memory in the process, while monotone improvement with length supports the expanding window. Either way the test blocks must be purged and embargoed and the number of window lengths tried must be counted as configurations searched.

Follow-up: How would you weight observations by age instead of truncating, and how do you choose the half-life without adding another overfitted hyperparameter?

Key concepts: walk-forward, expanding window, rolling window, regime change.