Older title:
Wording of "time-series" is understood as trend analysis although just a mix of some monthly/weekly/yearly/timeless columns as input is meant
A question of how to build an ML model from columns that are stored monthly in a database and other columns that are overwritten or timeless should be clearly needed on Stack Exchange if you set up the features of an ML model.
Attempts to get the votes to reopen this question might fail since most readers seem to misunderstand the question as asking for a time series analysis that predicts a trend like a classical time series. But this question is not about a time series that you would take to make a seasonal analysis or trend analysis for its decomposition. This is just about columns of some x months back, and columns of some y months back, and some columns of a database that do not have a history at all, that you altogether treat as normal features - as if you did not even know that they were monthly columns. And even then, you will get good predictions. I know this since I ran the model myself, so if anyone now wants to tell me that I have to take a time series analysis instead, they are wrong, proven by example.
Here is what I wrote against the downvotes (-1 at the time of writing):
Since I know from the practical model in the end that this misunderstanding is misleading, and that the model worked, I now try to convince readers to reopen this.
Since this got downvoted and since the first answer seems to see this as an ML time series question: This question is not about a classical time series analysis, but seeks to deal with monthly columns as features. There is no aim to check for a trend or any other decomposition! I have shared this question since I had exactly this challenge at work, and in the end, the model worked fine with this setup, mixing up non-monthly (timeless) features with monthly features. Therefore, this is just a question of using monthly data columns as features, and the model does not care about whether it is December or June, it only cares about how many months these features lie in the past, so that it learns from the pattern of data over the time of some x months before. The features are not called after the months, but just after how many months they lie back in time, like wealth_month_1, wealth_month_2 for the wealth of 1 or 2 months back in time.
But it got closed since it is said not to be a programming question. Then I ask the reader: how would you set up a model without knowing how to treat the monthly saved columns of your database as features?
Example:
Not every dataset covers just the height and colour of a plant. Some might want to take the height and colour three months after it being planted as further input features, some might want to take the height and colour one month before the final measurement as further input features.
Even if this question is so general, it needs to be answered so that you can program a (non-?)time series model on attributes that may or may not change over time. I lack the wording to define this, I guess you would still call this a time series, even if I now stress in the title: "(no time-series!)" and make this only clearer in the body.
I get remarks that even my own answer shows a picture of a "time series". Yes, this answer shows the monthly shift in the dataset that you need if you shift the training of the model by a month. But that does not mean that the whole dataset needs to be treated like a "time series": it does not mean to ask for any decomposition like seasonal or trend analysis. For example, I could also decide to take only some data column six months before, two months before, and that of the last month, and the model would still learn from these input columns as features. Such a model does not care about a gapless time series dataset, and it is not about it anyway, it is about checking for patterns in the feature input that lead to a given output for training. Until I predict this output in the end.
Working model, needed general question without any code, proven by work experience, still closed.
How do I get it reopened?
Should it be on Cross Validated SE instead?
Any other hints?
The word "time series" seems to be misunderstood since most readers and voters seem to misunderstand it as a classical ML time series question of a dataset with gapless columns over time even though it is just about monthly saved columns of the database and other timeless columns as one basket of features. What is a better word for this?
