0
$\begingroup$

First, let me say that I just started my job one month ago with less experience in data science (I graduated from electrical engineering field).

Now, I'm doing the project about demand forecasting for new gas station by using station properties (e.g. type of station, location, area) and geographic properties (e.g. number of banks/residents,shop... around the station, road, competitor, route, traffic) as features to properly locate new station to make it most profitable.

My senior said that former data scientist in my company also used to do this project. However, there are some problems when users choose rural zone (e.g. forest) as a new location. Let say that model's output (demand) is 100K. It seems like there is nothing with the output. But, model's output when users choose urban location is also 100K too.

The problem are

  • How to track difference between urban zone and country zone to get the sensible result? (Normally gas station in city should be more profitable than outside the city) Any good features?
  • Are there any research about these problems? I already searched for them but not yet see the good one.
$\endgroup$

1 Answer 1

0
$\begingroup$

First off, the predictors you list sound reasonable for this kind of model. You might also include stuff like population density or income, but some of your predictors (like banks) are likely already proxies for that.

It's not prima facie obvious that your result does not make sense. Yes, a location in the city should generate more revenue than one in the forest, because there are more customers close by. However, I notice that your model also includes the competition, and it's quite likely that there is a lot more competition in the city than in the forest, which would reduce the likely revenue of a new station.

(Incidentally, if you trust neoclassical economics, this is rather exactly what we should expect: marginal profit for any new opportunity should be equal, because anything that shows a higher profit has already been taken up. If we assume approximately equal costs for both locations - lower wages, but higher costs of trucking in product in the middle of nowhere - we should expect about equal revenue. No, I'm not saying to trust this blindly, but if city locations showed systematically and strongly higher expected revenue, then I would wonder why these opportunities had not already been taken, or conversely, why anyone would still be operating stations in less advantageous locations.)

Research in this direction has to contend with huge selection bias: if I use this kind of model for forecasting revenue or profit and then allocate my limited budget to build new stations at the locations that are forecasted to have high revenues/profits, then I have preselected my sample, so I can't really say whether my forecasting model was any good. It may well have been that revenues/profits would have been even higher at those locations where I didn't build a new station (because they were erroneously forecasted to be unattractive), but I will never know!

One place to possibly look at is a recent review article on retail forecasting by Fildes, Ma & Kolassa (2019, IJF), who explicitly note location decisions as one use case for forecasting. Unfortunately, I don't remember how deeply they went into this kind of literature, which is a bit embarrassing, since I'm one of the authors... we wrote this quite a while ago. Here is an earlier ungated version, or you can ping me on ResearchGate for the paper.

$\endgroup$

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.