I am currently working on a project in which I have to develop a model to predict how much money other companies will make by using the services provided by my company. The money made is a type of tax return. The model should predict a range (e.g 100k\$-200k\$).

Basically, if I give the model the data of the target company, it will give me how much they can potentially make with us.

The data I have is the financial statements of the companies we worked with. This includes their general and financial data such as capital, number of employees, city, type of industry, and of course the amount of money they made with us.

After researching online, I found that most solutions are forecasting and time series based. While the data I have has the year in which we worked with the clients, I do not think this is the right approach for me. I experimented with decision trees and a regression model but I do not seem to get a good result.

Any tips on where to look to solve this problem?

  • 2
    $\begingroup$ Have you discussed the problem with a subject matter expert, yet? That may help to quickly identify if there are key unknown variables (unknown as in you don't have them in your financial statements) as well as to help zero in on relevant variables among the many in the financial statements. You'll want to be careful with financial statements as those are typically backward looking. $\endgroup$
    – C8H10N4O2
    Apr 28 at 17:56
  • $\begingroup$ For now I am just testing if there is any correlation between the data I have and the target. I cannot contact an subject matter expert yet. $\endgroup$ Apr 28 at 19:51

If you have many available input parameters/features (financial data like balance, depreciations, tax for several years for several companies, the business areas where those companies are working in (e.g. telco, banks, media, etc) and you already have "labelled" data (I mean, you already have the "tax" saving for you previous customers).... well, a good start could be a standard process of feature selection and designing a neural network.

  • $\begingroup$ so you think I should start with neural networks? $\endgroup$ Apr 29 at 13:44
  • $\begingroup$ Seems to me just a simple clasification problem, isn´t it? I mean... I dont see trelationship between the time when you achieved tax savings and those savings... A better or worse tax saving has nothing to do whith the time (the year) you achieved it but with the financials of the company, your advice and the regulations (that may change from year to year that´s true). $\endgroup$ Apr 29 at 14:38
  • $\begingroup$ Yes, that is exactly the case $\endgroup$ Apr 30 at 7:08
  • $\begingroup$ Well, then you can use just a vanilla deep neural network like this one to predict the range of "tax Savings" for each of your customers based on their financial statements and your "labelled" data based on your previous customers: analyticsindiamag.com/… $\endgroup$ May 1 at 8:47
  • $\begingroup$ I don´t think a time series analysis will help you here as "a time series allows one to see what factors influence certain variables from period to period. Time series analysis can be useful to see how a given asset, security, or economic variable changes over time". In this case your tax saving is not depending on time (I dont think it changes depending on the month or on the year when you calculate it). investopedia.com/terms/t/timeseries.asp unles you take into account "past" tax regulations that have changed and that don´t apply to forecasting anymore.. $\endgroup$ May 1 at 9:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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