Our model apparently presented a good enough result during training and testing.įor the forecast we have to consider that, since we didn’t had a big enough time range in our training data, we had to decrease how far in the future our model could predict. The idea behind this story was to get to work a bit with time series and Recurrent Neural Networks. Since there is no data for our forecast period we can’t really measure how well our forecast was. If we plot the correlation between stores for the same department and between departments within the same store we can see that departments tend to have a stronger correlation Therefore, we could either generate a model for each store or for each department. So maybe a unique model will have problems finding a unique generalization function that works for all stores and departments.
We need to keep in mind that at the end of the day, Machine Learning algorithms for supervised problems are all about finding a generalization function. Second, we need to think how store sales are influenced and what influence it, for example, can we do a single model that will predict independent of store and department? Or maybe we need a model for each store or one for each department? The first thing to have in mind is that we are dealing with time series. I decided to reduce the scope of the analysis focusing primarily only in the data related to the Top Stores and Departments, i.e, those that contribute more to the final revenue of the company.īefore we begin dealing with preparing the data for modeling we need first to stop and think about our model and what we want to predict at the end. Now let’s deal with the data as a whole, dropping the Markdown columns. Total of 14 Stores contributed for 50% of the Revenue Analyzing Data for Forecasting Modeling
0 kommentar(er)
