Reading time: 5 min
Easy question, tough answer
When it comes to leveraging a company’s data assets, the first use-case that pops into your head is probably not the optimization of opening hours. However, if we consider an enterprise which operates a number of offline shops, the potential of such a project becomes clear.
Setting the goals of the optimization is easy as pie: if a shop operates too long then the operating costs are higher than optimal. If it does too short, additional revenues may be realized. In the first case, the open hours should be decreased while in the latter, they should be increased.
We might as well think of two things that question an optimization project like this:
- is it indeed possible to “play” with the shop hours?
- isn’t the shop manager adaptive enough based on business operation?
The answer to the former concern is the usual “it depends”. During our project, we have dealt with shops that had both legal (some shops’ working hours were regulated by law) and business constraints (employment is only possible with conventional hours/week for employees) on working hours. However, scenarios like the following may be accounted for:
- 24/7 vs closing at night
- 12 vs 16 hours of operation
- opening 1 hour earlier/later
- closing 1 hour earlier/later
A savvy manager might realize what customer volume to expect in a given hour and he might take this into account when deciding on the operation hours. Nevertheless, can he accurately guess the (additional) expected revenue when operating longer? And does he make an optimal decision on the network level? By using machine learning, we can definitely predict more accurately and with more objectivity. It is quite sure that, by using only human intuition, we will not be able to find an optimal solution on company level due these two factors:
- one does not know whether, by operating longer, one actually steals revenue from another shop of ours (cannibalization)
- one (e.g. shop manager) might have personal incentives not to operate less hours or close a shop entirely
In the following, we present you the method that can tackle this problem in a data-driven way.
Modelling customer decisions in a simulation framework
Simulations are extensively used when optimizing complex, real-world processes (e.g. traffic design). In this case, the outputs such as a financial metric (e.g. EBITDA) of working hour combinations may be simulated. To make this efficient, we should model the decisions of the customers, thus we may see the effects of changing the operation hours on micro-level. The two most important effects are:
- how many additional customers do we attract by operating longer and how many of them are taken away from another shop of ours (cannibalization)
- how (where) do customers substitute when we close a shop
The animation above demonstrates how customers make decisions when we close two of our shops close to each other. We can see, that by closing shop No. 2, we lose 5 customers (they choose the blue competitor) but we also save operational costs. Our simulation framework models such scenarios in more iterations and larger scale. It assigns an output (e.g. EBITDA) to all inputs (operation hour combination of our shops). Having that, we can easily conclude the optimal set of working hours for our company.
The strength of our solution lies in its ability to account for various factors that influence customer choice. In the animation, the customers chose solely based on distance: after closing their shop, they choose the one that still operates and is the closest (it may also be a competitor). Such additional factors are:
- loyalty: customers of different loyalty have different willingness to travel
- competition: the heterogeneity of the competition may be quantified. Substitution with competitors of price level significantly higher or lower is less likely
- travel time: it may be more important than distance alone
Summary
Most companies operating a big network of shops base their decision of operation hours on instincts and conventions. By using sophisticated machine learning solutions, we can support these decisions. Such a project can be really beneficial but it also entails challenges. The biggest challenges is, by no means, the collection and processing of accurate data. If we can tackle this, we are on the right path to create vale with data.
Author: Gerold Csendes – Data Scientist