Resampling Time Series

Resampling changes the frequency of a time series. Downsampling reduces frequency — for example, converting daily data to monthly totals. Upsampling increases frequency — for example, converting monthly data to daily, filling gaps with interpolated values. Both operations require a DatetimeIndex and are performed with the resample() method, which is Pandas' time-aware version of groupby().

df.resample(rule) creates a DatetimeIndexResampler object, where rule is a frequency offset alias. Like groupby(), nothing is computed until you chain an aggregation method. Common rules: 'D' (day), 'W' (week), 'ME' (month end), 'QE' (quarter end), 'YE' (year end). The DataFrame must have a DatetimeIndex.

To downsample daily data to monthly, call resample('ME') and chain an aggregation. sum() gives monthly totals; mean() gives monthly averages; last() gives the last value in each period. The result has one row per period, with the period-end date as the index label.

Resample with 'W' to aggregate by calendar week ending on Sunday. Use 'W-MON' if you want weeks ending on Monday. Pandas groups all dates within each week together and applies the aggregation. This is useful for weekly reporting where you want one summary row per week.

Just like groupby(), you can chain .agg() on a resampler to compute multiple statistics in a single pass. Pass a list of function names or a dict for named aggregations. This is the most efficient way to build a comprehensive time-series summary table.

For financial time series, the .ohlc() method resamples to Open, High, Low, Close (OHLC) values per period — the standard candlestick representation. This is only meaningful for numeric data representing a price or level. Each column in the input gets four OHLC columns in the output.

Upsampling increases frequency, which creates rows for timestamps that did not exist in the original data. These new rows initially contain NaN. You then fill them with a method that makes sense for your data: ffill() (forward fill — carry the last known value forward) or bfill() (backward fill — use the next known value), or interpolate() for smooth interpolation.

After upsampling, fill the NaN gaps created for new timestamps. ffill() propagates the last known value forward until the next real observation — good for prices where the last traded price remains valid. interpolate(method='linear') fills with linearly spaced values between observations — good for smooth continuous variables.

You can combine groupby() and resample() to resample within each group separately. Call groupby(column).resample(rule) on a DataFrame with a DatetimeIndex. This produces a result with a MultiIndex where the outer level is the group key and the inner level is the resampled time period — very useful for product-level or region-level time series analysis.

By default, 'ME' anchors to calendar month-end dates. For non-standard fiscal periods, use offsets like 'QS-APR' (quarter starting April) or 'YS-OCT' (year starting October). Pandas supports many anchored offset aliases. Check the Pandas documentation for the full list when working with non-calendar fiscal periods.

After resampling, always verify the result makes sense. Check that the number of output periods matches expectations, that no periods are missing (look for NaN in the output), and that the sum of monthly totals equals the sum of the original daily data when downsampling with sum(). These sanity checks catch off-by-one errors in frequency strings and missing date ranges.

Test your understanding of resampling time series from this lesson.

In this lesson you learned: resample() is the time-aware equivalent of groupby() for changing time series frequency; downsampling aggregates data (daily to monthly with sum/mean); upsampling creates new timestamps that must be filled with ffill() or interpolate(); and you can combine groupby() with resample() for group-level time aggregation. Next up we explore shifting and lag features.