Common mistake 1 with Change frequency (resample)

Summing a rate. An interest rate averaged over three months is meaningful; the sum of three monthly interest rates is not a number that exists.

Common mistake 2 with Change frequency (resample)

Averaging a flow. Quarterly GDP is the SUM of the months, not their average. Averaging silently divides your series by three.

Common mistake 3 with Change frequency (resample)

Going the other way - annual up to quarterly - is interpolation, not resampling. It invents observations, inflates your apparent sample size and makes every standard error too small. This module deliberately only goes down in frequency.

Common mistake 4 with Change frequency (resample)

Aggregating first and then differencing gives a different answer from differencing then aggregating. Decide which you mean.

EconStatsCraft › Change frequency (resample)

Change frequency (resample)

Real projects mix frequencies - monthly prices against quarterly GDP against annual budgets. This converts series to a common frequency so they can be modelled together, and it is careful about which rule fits which kind of series: a flow is summed, a stock or a rate is averaged or taken end-of-period.

What it needs

Time column - Filled in automatically when your data has a recognisable date column.
Series to convert - Leave empty to convert every numeric column.

The intuition

Standard errors come from a formula that assumes a lot: a distribution, a sample size large enough for the asymptotics, independence. When any of that is doubtful, resample instead. Draw new samples from the data you have, recompute the statistic on each, and read the uncertainty off the spread of the results. For time series the twist is that you cannot draw single observations, because that destroys the dependence you are trying to preserve.

When to use it

Whenever your sources disagree about frequency, which in applied work is nearly always. Prices are monthly, national accounts are quarterly, budgets are annual. Every multivariate model needs them on one clock first.

Where the estimator comes from

The ordinary bootstrap resamples rows with replacement and is invalid for dependent data - it produces independent draws and therefore understates the variance of anything that depends on the serial correlation. The MOVING BLOCK bootstrap draws overlapping blocks of length \ell and concatenates them, preserving dependence within a block and destroying it only at the joins. The STATIONARY bootstrap (Politis and Romano, 1994) uses geometrically distributed block lengths with mean \ell, which makes the resampled series genuinely stationary at the cost of a little extra variance. Block length matters and scales roughly as \ell\propto n^{1/3}; too short kills the dependence, too long yields too few distinct blocks.

The model

\text{flow: } y^{Q}_{\tau}=\sum_{t\in\tau}y_t \qquad \text{stock or rate: } y^{Q}_{\tau}=\frac{1}{n_\tau}\sum_{t\in\tau}y_t \quad\text{or}\quad y_{t=\max\tau}

What it assumes

The time column is a real date, month, quarter or year - the app parses it and tells you what it found.
You know whether each series is a flow or a stock. Nothing can work this out for you.
The period boundaries mean what you expect. A quarter beginning in April and one beginning in January are different series, and the module's convention has to match the source's or every comparison is shifted.
Partial periods at either end are handled deliberately. A final month with three days of data resampled to a monthly total is not a month, and leaving it in produces a spurious collapse at the end of every chart.

Statistical properties

The block bootstrap estimates the sampling distribution of a statistic, not the distribution of the data. The output is a standard error or an interval, not a synthetic dataset for modelling.
Choosing the block length is the whole difficulty and there is no universally good automatic rule. Report the length used and check sensitivity to it.
Percentile intervals are simple but biased when the statistic is skewed. BCa corrects for bias and skewness and is worth the extra computation.
For a statistic that depends on the extreme tail, the bootstrap fails: the resample can never contain a value larger than the sample maximum, so extreme quantiles are systematically understated.
A thousand replications suffice for a standard error; a 95% interval wants five thousand or more, and a 99% interval more still.
For non-stationary data the block bootstrap is not valid either. Difference or detrend first, resample the stationary residual, and rebuild.

How to read the output

Check the number of periods out against what you expected: 180 months should give 60 quarters or 15 years.
A period built from fewer observations than the rest is a partial period - usually the first or last. Decide whether to keep it.
Resampling a flow by mean and by sum answer different questions and both are defensible. What is not defensible is doing one and describing the other.
Aggregation reduces variance mechanically. A monthly series is smoother than the daily one underneath it, and a volatility computed after resampling is not the volatility of the original process.

A worked reading

On a persistent series, the ordinary iid bootstrap and the moving block bootstrap disagree about the standard error of the mean by a wide margin, and the iid one is always the smaller. The reason is exactly the autocorrelation: with positive serial dependence the effective sample size is far below n, so each observation carries less independent information than the iid bootstrap assumes. The block version keeps that dependence inside each block and recovers a standard error close to the HAC formula, which is the right sanity check to run alongside it.

Symptom → diagnosis → what to do

If you see this	It means	Do this
Bootstrap standard error far below the analytic one	Blocks too short, dependence destroyed	Increase the block length
Results very sensitive to block length	Normal; report the sensitivity	Show the range rather than one number
Interval not containing the point estimate	A strongly biased statistic	Use BCa
Bootstrapping an extreme quantile	The resample cannot exceed the sample maximum	The bootstrap is the wrong tool here
Very slow	Refitting a complex model per replication	A thousand replications for a standard error is usually enough
Series with a trend	Block bootstrap assumes stationarity	Detrend, resample residuals, add the trend back

Common mistakes

Summing a rate. An interest rate averaged over three months is meaningful; the sum of three monthly interest rates is not a number that exists.
Averaging a flow. Quarterly GDP is the SUM of the months, not their average. Averaging silently divides your series by three.
Going the other way - annual up to quarterly - is interpolation, not resampling. It invents observations, inflates your apparent sample size and makes every standard error too small. This module deliberately only goes down in frequency.
Aggregating first and then differencing gives a different answer from differencing then aggregating. Decide which you mean.

Where to read this properly

Efron (1979), "Bootstrap Methods", Annals of Statistics 7(1).
Kunsch (1989), "The Jackknife and the Bootstrap for General Stationary Observations", Annals of Statistics 17(3).
Politis and Romano (1994), "The Stationary Bootstrap", JASA 89(428).
Davison and Hinkley, Bootstrap Methods and Their Application (1997), ch. 8.

Related Time series - one series

Time-varying coefficients (Kalman) · Unit root tests · Correlogram (ACF / PACF) · ARIMA / SARIMA · Rolling-origin cross-validation · Theta method · Unobserved components · STL forecasting · Forecast combination · Markov regime switching · ARCH / GARCH volatility · Decompose & filter · Spectral analysis · Exponential smoothing · Forecast accuracy

All analyses

Data & variables

Each example is simulated from a known model, so you can check whether the estimator recovers the truth - and nothing here is anyone's private data.

Filter variables

Drag a variable into a slot on the right - or just click it.

Choose an example dataset above to begin.

Choose a transformation - it opens in the data editor, with a live preview.

Apply to

Multiply by

Columns ctrl-click for several

Group by ctrl-click for several

Within groups

Statistic

Bands

How many

Cut points

Keep rows where

Value

and

Fill with

Value

Also mark which rows were filled

Rules

Base period = 100

Which part

Percent at each tail

Threshold

Survey weight optional

Direction

Names column

Values column

Largest first

Delete them, not just count them

Read this column as

Tidy the text

New name optional

EconStatsCraft rigorous, plain-spoken econometrics

Econometrics that explains itself.

Regression, panel data, time series and the whole battery of tests. Every result is followed by a verdict you can defend.

Load a dataset to see it here.

Change frequency (resample)

What it needs

The intuition

When to use it

Where the estimator comes from

The model

What it assumes

Statistical properties

How to read the output

A worked reading

Symptom → diagnosis → what to do

Common mistakes

Where to read this properly

See also

Related Time series - one series

All analyses

Charts

Maps

Explore

Regression

Time series - one series

Time series - several series

Panel data

Multivariate statistics

Marketing research

Causal

Finance & risk

Actuarial

Tests & breaks

Machine learning

Econometrics that explains itself.