1
$\begingroup$

i'm currently doing dual apprenticeships. My main mission is to represent the health of a company based on accounting records for multiple companies over multiple years. The part of an accounting record i'm interested in is simply the date and the amount spend or receive. We're in a context of time series.

The objective is to build scores, each one representing a part of the company. I have a few indicators for each one of this company's part.

I didn't find any scientific articles about it, so i proposed an algorithm to do so. Here is it :

For each indicator

  • Build the mean for this indicator for each company
  • Use a regressive model to get the trend
  • Compute the difference to this trend (the objective is to increase the amount of data)
  • Make a forecast
  • Build the mean for this forecast
  • Use a regressive model on this forecast to build the trend
  • Compute the difference to this trend for each company
  • Based on this difference, build the score

I know this is not perfect, but after a lot of thinking, that's what i ended up with.

For the forecast part, i'm planning to use an LSTM.

I would like to hear your thought about my algorithm, even if you think it is completely crap. My objective is to improve my skills and to build the best system i can. I'm still a bit lost : should i try to transform the problem and go for more classical, like a simple classifier, or should i do that just for the last part ? Well, i don't know.

Thank you.

EDIT As Julio made me realize, i forgot to mention one important thing. In France, each one of the record has an identifier, which represent what type of of income/outcome it is. For example : each accounting record with the id 40 are accounting charges. In this example i showed you a simple 2 digits identifier, but it can go up to 8.

$\endgroup$
4
  • $\begingroup$ do you have information for those companies such as income, outcome, payroll, providers by date? $\endgroup$
    – Multivac
    Commented Nov 25, 2020 at 0:43
  • $\begingroup$ Yes, I'm sorry I forgot something important. I'm French, and in France, each accounting record has an identifier that tells us what this is (example : a record that has the code 40 is an outcome, etc). A record can be identified by up to a 8 digits identifier. The digitxthe more precise the financial operation it describes $\endgroup$ Commented Nov 25, 2020 at 6:54
  • $\begingroup$ It is unclear what you are trying to predict. what do you mean by health of a company ? What are you trying to measure with your score ? $\endgroup$ Commented Nov 25, 2020 at 12:39
  • $\begingroup$ What i want to do is use forecasting to predict different indicator. Each indicator can be assigned to a category. Each one of this categories represent an aspect of a company. I want to build a score for each category based on the forecast. $\endgroup$ Commented Nov 25, 2020 at 13:27

1 Answer 1

0
$\begingroup$

To me it looks like your question is more about expert knowledge than ML approach: the main problem is how to build a sensible score based on the available indicators. I have zero knowledge about this domain, but in general this kind of problem requires a lot of back and forth with domain experts in order to correctly specify the task in detail.

However I have one piece of advice which might help (or not, maybe you already know this) for the evaluation of any system you build for this task: assuming that you obtain a predicted score which is meant to represent the future of the company, you could evaluate how well this score matches "the future" as follows:

  1. Split the data at a particular point in time in the past, say 2010
  2. Use only the data before 2010 to train and test a model, i.e. obtain predictions based only on data before 2010
  3. Evaluate how well the predicted scores match the "future" of the companies, i.e. the actual data after 2010. For example a low score should lead to poor indicators, a high score should lead to good indicators (you might need to devise a specific method to correlate the scores with the indicators).
$\endgroup$
1
  • $\begingroup$ Thanks for your answer. That's what i was planning to do. Because i'll base my scoring system on forecasting, i need to be sure that my forecasts are as strong as i can get them. I'm also currently in discussion with expert of the domain to extract the different indicators. The objective is simple : look at n scores, and based on this n scores, you can say 'this is a healthy company' or not. $\endgroup$ Commented Nov 25, 2020 at 12:03

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