changsik choi的动态

查看changsik choi的档案
changsik choi

Data Product Manager | Data Analyst

Solving the E-commerce Long-tail Inventory Puzzle Recently, I've encountered a very interesting problem. This fascinating problem is about creating a model that meets the following conditions: 1. Predict sales of long-tail products in e-commerce (products that sell sporadically, a few at a time) 2. Minimize stock-outs (when demand exceeds inventory and products can no longer be sold) based on these predictions 3. The operations team should be able to tune the model What makes this truly interesting is that the sales of most long-tail products are literally '0'. If the goal was simply to 'predict sales accurately', we could just set all future sales to 0. However, if we create a model like this, while the difference between predicted and actual sales would be minimized, all products would end up in a 'stock-out' state when deciding how much to store in the warehouse. In other words, the accuracy of the prediction model becomes practically meaningless. How interesting is this problem! (Can you hear the sound of my head exploding?) To make matters more complex, most of the time series prediction models we know (e.g., ARIMA) assume that products sell at least a little every day and have sales records for at least a few months (ideally more than 3 years). Therefore, when using these models, the predictions aren't very accurate. Then some people might say: 'AI is hot these days, can't we predict using AI or machine learning?' Of course we can, but to utilize machine learning and AI, we need to satisfy the following additional requirements: 1. There's a lot~~~~ of clean and diverse data. 2. The machine learning algorithm and AI don't need to explain 'why they predicted sales that way'. However, in reality, clean and diverse data often doesn't exist, or even if it does, it really requires more than a quarter of data cleaning and design. Moreover, for operational goals and efficiency, we often need to explain and tune the logic, making it difficult to utilize machine learning. So, coming back to our three constraints (1) Improving prediction accuracy. (2) Minimizing stock-outs. (3) Allowing fine-tuning for operational purposes. How should we create a prediction model to achieve these? Let's go through the following 5 steps! Step 1: ...... https://lnkd.in/gkcnarrr

要查看或添加评论，请登录

最相关的动态