Model retraining

Suppose we are working on timeseries data (like predicting the stock prices for the next day). We are using Day1 to Day 100 as our training data and our model is trained with the 100 points

And we deployed our machine learning model. For any machine learning model we use some performance metrics like MSE(mean squared error).

For day 101,102 our model predicted correct(MSE value is low) but from day 103 to day 109 the stocks of the company got low due to some external factors ,then for predicting the stock prices we need to retrain our model with the latest data.

This process is called model retraining

Model retraining should happen periodically if we dont have so much training data size.

Otherwise we should retrain the model when the performance of the model is decreasing and data it self is changing

How do you know data itself is changing?

When the cross validation performance is high and test performance is less. Then mostly it may be because the data is changing with the time