In logistics services, the only way to grow is to reduce costs, as this is a cost centre for many managers. On average, in France and in all business sectors, logistics costs represent between 8 and 10% of sales. Data science and associated tools can help reduce these costs while improving customer satisfaction. This means faster delivery, lower costs and fewer errors.
Take, for example, a situation of over-stocking of maintenance parts in all warehouses: this implies high capital costs (space, obsolescence, etc.) and it is normal for companies to seek to optimise their inventory levels while guaranteeing a sufficiently high service rate. Once the best model is found for predicting the consumption of maintenance parts, the importance of business rules is critical for parts that cannot be modeled. Charlotte also discusses the choice of an optimal model by the inventory manager, which parameters are analyzed?
Finally, successful integration of the model into the value chain is based on business ownership. Keep in mind that a model intelligence requires human intelligence to be adapted to a specific context.
Today, I'm gonna show you how data will change logistics jobs and especially for inventory managers. For those who are not initiated to logistics, basically it is a cost center for all companies that sell products. Thus in logistics services, the only way to grow is to reduce costs. On average, in France and in all business sectors, logistics costs represent between 8 and 10% of sales. Data science and associated tools can help reduce these costs while improving customer satisfaction.
Indeed, there is a balance to be found between the stored value and the service rate since fewer parts in stock inevitably implies a decrease in the service rate.
In the example I'm gonna present, the goal was to optimize stored value for maintenance parts for an industrial company. The method we used was to predict demand, which is stock consumption here, and then make a recommendation of inventory level.
So what does stock consumption look like for these maintenance parts? As is often the case, maintenance parts have a sporadic consumption type, i.e. parts are consumed at very long and irregular intervals.
Given the particular type of stock consumption, the Croston model is the most appropriate. We found it in the literature of intermittent demand forecasting.
The Croston model allows two independent predictions: one for the value of each consumption and another for the interval between each consumption. Prediction is an arbitration between the last prediction and the most recent observation; thanks to this exponential smoothing principle, historical data receive variable weights according to their age in time.
Business knowledge of the type of parts stored and their consumption characteristics makes it possible to add rules for parts that cannot be modelled (not enough historical values). You’ll end up with a flow chart with a specific rule for each category of items. For example, for a very expensive item, not sensitive, and which is rarely consumed, it will be more reasonable not to keep stock because being able to satisfy all orders in this category would explode the capital tied up.
The custom algorithm generates a number of models that test different combinations of the smoothing parameters. The performance of the calculated models is evaluated against the two KPIs initially defined: service rate and stored value. And you need the inventory manager knowledge to choose the optimal model. I'm sure you noticed that this one is objectively the best one. But maybe at first, to build trust with warehouse teams the inventory manager might want to choose this one. It's not the best model from a mathematical point of view that fits, it's the model that best fits a business reality.
To conclude, the AI is not self supporting as its use and exploitation require, on the one hand, adaptation to a business context and, on the other hand, the intervention of the business to analyse the results and implement them. Inventory managers will have to work with AI under the form of a decision support tool, instead of being replaced by AI.