Desplazamientos y variables de retardo

In time series analysis, the value at a previous time step (a lag) is often one of the best predictors of the current value. For example, yesterday's sales are informative about today's sales. Creating lag columns lets you use historical values as features in machine learning models or in statistical analysis of autocorrelation. Pandas provides shift() to create lag features in a single vectorised call.

Series.shift(n) moves all values down by n positions (positive n creates a lag), filling the first n rows with NaN. Passing a negative n moves values up (creates a lead, shifting future values back to the current row). The index remains unchanged — only the values are displaced.

You typically create several lag features at once for machine learning — lag-1, lag-7 (same day last week), and lag-30 (same day last month). The most readable way is to create them in a loop and assign each to a new column. The resulting DataFrame has the original series plus all its lagged versions ready for modelling.

Instead of shifting by an integer number of rows, you can shift by a time offset using the freq parameter. df.shift(1, freq='ME') shifts the index forward by one month-end, keeping the values in place but changing the index. This is useful for aligning two time series that are offset by a fixed time period without rearranging the values.

Series.diff(n) computes the difference between a value and the value n positions before it: x[t] - x[t-n]. This is commonly used to compute day-over-day changes, week-over-week differences, or year-over-year deltas. The first n rows are NaN since there are no previous values to subtract from.

Series.pct_change(n) computes the relative percentage change: (x[t] - x[t-n]) / x[t-n]. This is essential for financial analysis (daily returns), growth rate computation, and any situation where the absolute change is less meaningful than the relative change. Multiply by 100 to get percentage points.

You can combine shift() with arithmetic to compute derived time features. For example: the ratio of today's value to last week's value, the cumulative sum since a fixed start, or the difference from the year-ago value. All of these follow the same pattern: shift the original series and perform element-wise arithmetic with the current values.

Series.cumsum() computes the running total from the first row up to each row, while Series.cumprod() computes the cumulative product. These are useful for cumulative revenue, cumulative returns (compound growth), and year-to-date totals. They require no arguments and work on any numeric Series or DataFrame column.

When preparing time series data for machine learning, a standard feature engineering step is to create a matrix of lag features. Each column represents the target variable at a different past time step. After creating lags, drop rows with NaN (which appear at the start due to missing history) and split into train and test sets without shuffling, respecting the temporal order.

When your data contains multiple entities (e.g., multiple products or regions in one DataFrame), you must compute shifts within each entity's group separately. Use groupby().shift() to ensure that the lag for product A's first row is NaN and not the last value of product B. Mixing groups without this step is a common and subtle data leakage bug.

A positive pct_change() value means the current value is higher than the previous one; a negative value means it is lower. Watch out for division-by-zero when the previous value is zero — the result is NaN or inf depending on the sign of the current value. Use replace([float('inf'), float('-inf')], float('nan')) to clean infinite values after pct_change().

Test your understanding of shifting and lag features from this lesson.

In this lesson you learned: shift(n) creates lag features by moving values down by n positions; diff(n) computes period-over-period absolute change; pct_change(n) computes relative percentage change; and when working with multiple groups you must use groupby().shift() to avoid data leakage. Next up we extract temporal features using the .dt accessor.