0
$\begingroup$

I have a financial dataset which has 10 years worth of data. The aim is to build a regressor capable of predicting next year sales. So, if I want to predict sales for 2024, I could use data from 2023, 2022, 2021, and so on...

If I use all my features, I'm able to get a decent model, but with some feature selection/engineering, I believe we can have much better model, because some of my features are very noisy. So, I was trying to apply MDA (Mean Decrease accuracy) and analyze feature importance for a Random Forest, but the big thing is, the feature importance keeps varying a lot from year to year, so its really hard to select which features are noise and which are not.

For instance, for 2023 the inflation is a very important feature, but back in 2018 it was not that important and was only adding noise. I thought about adding more features or trying to differentiate my features in order to have something more stable, but no success yet.

So my question is, what are the best ways to tackle this?

Thanks for reading!

Improve this question
$\endgroup$

2 Answers 2

Reset to default
1
$\begingroup$

I think you can derive new features that should be more stable over time using PCA. A more simple approaches is to calculate the technical indicators of these numerical features such as RSI, moving average as they will represent the trend more effectively.

Improve this answer
$\endgroup$
0
$\begingroup$

The aim is to build a regressor capable of predicting next year sales. So, if I want to predict sales for 2024, I could use data from 2023, 2022, 2021, and so on...

it is not a Regressor in full-sence -- RegressionAnalysis & TimeSeries Analysis have different aims & different algorithms & different conclusions... In regression analysis you investigate Dependency and you should analyse Independent and identically distributed random variables, but in TS present state is (can be) dependent on previous state -- it is Markov Process, that can describe State-Space Models, aka TSAnalysis is -- and its aim is just to analyse time(I.V.) as feature and response(D.V.) depending on time, NOT any other features/causes for the certain response, like Inflation or etc. that you can imagine... all of them depending on time will have there own trend & seasonal components, that can be used in the prediction... nothing about regressing -- do not mix two different kinds of analysis

Improve this answer
$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .