This section details the best practices and considerations for refreshing your
Meridian model as new data becomes available.

## Refresh Frequency

Model refreshes can be done as frequently as you would like. Model selection and
tuning is typically an iterative process, which may need to be refreshed along
with new data. You might consider updating the model quarterly, annually, or at
a frequency that matches your marketing budget decision making process.

## Recommended: incorporate new data by adding it to the older data

We recommend adding the new data to the older data and running Meridian.
One ought to consider whether or not to discard the oldest data to accommodate
the new data. This may be necessary to stay in the 2-3 year data window that's
common in an MMM. Meridian doesn't model media effectiveness as
time-varying. So, the decision to discard old data when appending new data is a
bias-variance trade-off. A longer time window reduces variance because you have
more data, but it can increase bias if media effectiveness and strategies have
changed drastically over time.

Recognize that MMM estimates often exhibit high variance. This can mean that
incorporating even a relatively small amount of new data may have a noticeable
effect on the model's results. For this reason, there can be valid business
reasons to set the priors in the new model to encourage the posterior of the new
model to match the posterior of the old model. We recommend that you set priors
based on prior knowledge and intuition, and it is reasonable for this intuition
to be informed by past MMM results. It is your decision as to how strongly you
want past MMM results to inform your prior knowledge and intuition. However,
consider that setting priors that match an old MMM's results effectively counts
the old data twice.

Priors are not the only way to limit the influence of a relatively small amount
of new data on a model's results. Another way is to ensure consistent
hyperparameters are used for the original model and the refreshed model.
`knots` are an important hyperparameter often selected by Meridian's
[automatic knot selection (AKS)](https://developers.google.com/meridian/docs/advanced-modeling/setting-knots#automatic-knot-selection)
algorithm. AKS includes advanced features to ensure knot location consistency
across model refreshes. Without these features, AKS might select entirely
different knots during a refresh. For more information and a code example, see
[Model refresh with hybrid knot selection](https://developers.google.com/meridian/docs/advanced-modeling/setting-knots#model-refresh-hybrid).

## Alternative: model new data disjointly and use priors

Some may consider incorporating new data by fitting a model to just that new
data, disjointly from the data used in old models. Although technically
possible, even for a small amount of data such as a quarter, this is generally
not recommended.

Modeling the new data completely disjointly from the old data won't properly
consider lagged effects. Meridian allows media data to include more (older) time
periods than the KPI and controls data. This allows the lagged effects to be
more accurately modeled beginning with the first time period of KPI data. It is
best to include max_lag time periods of media data prior to the first time
period of KPI data whenever possible.

A small amount of new data is likely not informative enough for the model to
make conclusions (see Amount of data needed). One may want to incorporate the
information from the old data by using a prior distribution informed by the
posterior of the older model. While the full joint posterior distribution of all
parameters theoretically contains all information from older data (and using it
as a prior for new data would be similar to fitting a new model that combines
both old and new data), Meridian uses independent prior distributions for
individual parameters. Therefore, even if the posterior distribution for each
individual parameter were carried over as its prior, it might not fully capture
the complete joint posterior distribution, which accounts for interdependencies
between parameters. Additionally, Bayesian models require a parametric prior
distribution for each parameter. MCMC sampling provides an empirical sample from
the posterior, which may or may not have a suitable parametric approximation for
direct use as a prior.