Overfitting on purpose Would it make sense to overfit a model on purpose?
Say I have a use case where I know the data will not vary much respect to the training data.
I'm thinking here about traffic prediction, where the traffic status follows a fixed set of patterns 


*

*morning commute

*night time activity 

*and so on.


These patterns won't change much unless there is a sudden increase of car users or major changes in the road infrastructure. In this case I would like the model to be as biased as possible towards the patterns it learned in current data, assuming that in the future the pattern and the data will be very similar. 
 A: No, it does not make sense to overfit your data. 
The term overfitting actually refers to a comparison between models: If model_a performance better on the given training data but worse out-of-sample than model_b, model_a is overfitting. Or in other words: "there exists a better alternative".
If the traffic status "will not vary at all with respect to the training data", then you will achieve the best possible results by simply memorizing the training data (again, that's not "overfitting").
But "data will not vary much with respect to the training data" simply equates to having a reasonable representation of the underlying pattern. This is where machine learning works best (stationary environment as Ferdi explained).
A: I would say, that there is a sense to overfit your data, but only for research purposes. (Don't use overfitted model in production!)
In cases when data can be complex and task non-trivial, trying to overfit a model can be an important step!
If you can overfit a model - it means that the data is possible to be described by the model.
If you cannot even overfit - it can give you a clue for investigation:


*

*you data is not ready to be modelled, so you would need do more data preparation / feature engineering

*your model is too simple and cannot capture all data dependencies

A: In General it does not make sense to overfit your data on purpose. The problem is that it is difficult to make sure that the patterns also appear in the part which is not included in your data. You have to affirm that there are pattern in the data. One possibility of doing so is the concept of stationarity.

What you describe reminds me of stationarity and ergodicity. From a contextual side/ business side you assume that your time series follows certain patterns. These patterns are called stationarity or ergodicity.
Definition stationarity:
A stationary process is a stochastic process whose unconditional joint probability distribution does not change when shifted in time. Therefore parameters such as mean and variance also do not change over time.
Definition ergodicity:
An ergodic process is a process relating to or denoting systems or processes with the property that, given sufficient time, they include or impinge on all points in a given space and can be represented statistically by a reasonably large selection of points.

Now you want to make sure that it really follows these certain patterns. You can do this, e.g. with Unit root test (like Dickey-Fuller) or Stationarity test (like KPSS).
Definition Unit root test:
$H_0:$ There is a unit root.
$H_1:$ There is no unit root. This implies in most cases stationarity. 
Definition Stationarity test:
$H_0:$ There is stationarity.
$H_1:$ There is no stationarity. 
Further reading:
What is the difference between a stationary test and a unit root test?

It the time-series really follows these patterns forecasting and predicting will be "easier from a statistical point of view", for example you can apply econometric models for forecasting like ARIMA or TBATS. My answer relates to univariate and also multivariate time series if you have cross-sectional data stationarity and unit roots are not common concepts.
