Stock-Out Prediction System for Fashion Brand
goals
Forecast sales/demand for the products in different stores to reduce the Stock out events. The forecasts need to be weekly and for next 8 weeks.
challenges
There were 1000s of products to forecast which would be difficult using statistical/ classical methods of Evaluation Metrics; how to evaluate the model's noisy data.
solution
Deep Learning Models that can learn from sales time series of multiple skus as well as from other features in data. The team came up with a number of evaluation metrics , but it is difficult to evaluate the forecasts more descriptively. They employed a Dashboard using dash to compare the model outputs on testset.
results
The team was able to produce forecasting for 8 weeks by using 16 weeks previous sales history. This is 3x better than the Naive Average Method for Outlet channel forecast, and 50% better MAPE than the existing statistical/seasonal sales forecasting software package for sku-store forecasts.