0
$\begingroup$

Assume we have data for two years, 2021 and 2022.

When training, we use 2021 data to predict 2022 outcomes, resulting in model 1.

Now during inference, we use 2022 data to predict 2023 results.

Assuming the trend changes in 2022 and we have few months of data in 2023, is it possible to make model1 more accurate?

A plausible method is to retrain model using the combined data from 2021, 2022, and the available months in 2023; however, this will not work because we need to use 2022 as target variable (which make 2022 unusable)

$\endgroup$

1 Answer 1

0
$\begingroup$

In my opinion this is very open-ended and highly model dependent.

If you used an ARMA/ARIMA based model originally to tune the coefficients, then you would need to start it from scratch.

If you used some kind of Bayesian structural time series you may or may not need to re-train it depending on how much data cardinality there is, because ideally Bayesian-based models (and other online learners), should progressively "learn to adapt over time".

If you used one of those fancy neural network one or zero shot learners then you may not even need to technically re-train. You can just give it the observed data and "should in principle" be able to shoot forward in time again based on the changed observations in 2022.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.