We discussed yield management algorithms in this week’s Tuesday Tutorial episode (link below). We also touched upon the newsvendor model as one such algorithm.
If you are not familiar with the Newsvendor model, you can review the episode linked above for an overview. The underlying concept is simple. The model derives its name from one of the use cases it caters to: a vendor selling newspapers. When they reach the newspaper depot every morning, assume that the vendor must decide how many copies they want to buy to sell on the streets. Knowing that unsold copies are worthless, what will be an optimal number?
Examples of key parameters leveraged in the model are:
- Cost of overage (what are leftovers worth?)
- Cost of underage (Opportunity cost of missing demand)
- Percentage of demand you want to cover
- Historical demand distribution
A key input in the newsvendor model is demand data. For items that leverage yield management approaches like the newsvendor model, demand forecasting is equally uncertain. The newsvendor model is popular because it provides a more structured way to plan for such uncertain demand scenarios. But how can we bring more certainty into the demand forecast to further increase the newsvendor model’s efficacy?
This is one area where deep learning models can help much better than forecasting for consumer-packaged goods. For example, if you place a one-time order for seasonal fashion apparel, historical demand data will only add to the uncertainty unless you are a fashion trendsetter like ZARA. Fashion evolves quickly, so historical data from previous years need to be “adjusted” to account for these trend evolutions. Deep learning can help improve forecasts for such items significantly, thereby, improving the yield management algorithm accuracy.
A similar postulate holds for other parameters like underage costs. Underage cost is not always about lost revenue only. How do you account for lost brand loyalty? If a consumer decides to start shopping for those specific seasonal items at another chain because they are not in stock at your chain, they may also become loyal to that other chain for all their fashion needs. While it is impossible to capture this precisely, AI algorithms can be leveraged to incorporate these nuances, hence improving the model’s overall accuracy.
There is another aspect to underage cost. If an item you ordered becomes out of stock, should the underage cost be calculated based on the initial set price, or what could have been charged because the item was hot? This is another area where ML algorithms can bring more accuracy. Using a more accurate underage cost can significantly improve your overall model accuracy.
You can keep increasing the list of factors. What if your policy is that you will get that product at any cost through flexible suppliers and expedited shipping? As you can assume, your underage cost evolves significantly in this scenario. If you want to put a more scientific number together to account for the expenses, ML algorithms can come to your rescue here as well. The list can be long on the overage cost side as well, and for each of those factors, AI and ML algorithms can help you.
In the “Friday Fun” episode this evening, we will cover a News end or model use case to understand what it is and the key parameters involved. We will use a fashion apparel example but will keep the use case limited to a simple introduction to the model.
The episode can be watched here:
An excel file and a python notebook with calculations will be uploaded on 04/13.
Designed Analytics Blog 