Supply Chain professionals describe the present business environment with 4 drivers : olatility, ncertainty, omplexity and mbiguity (). Firms are working on gaining and adapt their products to behavioral changes, not to say reacting to increasing supply complexity.
In the style sector, of products in stores were . In the identical way, of the automotive catalog is renewed every yr.
Due to this fact, latest products forecasting has grow to be a day-to-day challenge for demand planners who cannot only depend on historical data to make their predictions.
After each launch, Supply Chain teams, along with Merchandising and Marketing, must closely observe the behaviour of consumers to latest products and react accordingly.
- What are the which could explain why a product is successful in region A and never in region B ? Visibility on Instagram? Web traffic volume? Conversion rate?
- Which quantities should we move from region B to region A to in the primary and within the second ?
- Can we observe between a latest product and an old product ? Do we want to of the brand new product ASAP?
- How should we in an effort to enhance a deceptive product launch ?
At IRIS by Argon & Co, we have now teamed up with several to deal with this challenge. We’ve mixed industry best practices — processes and agile tools — with rigorous data-driven approach leveraging AI.
Traditional forecasting softwares use statistical algorithms based on . There are ineffective for launches since the historical data (.
Similarly, mainstream AI forecasting models are found inappropriate for lower than 15 data points’ time series.
We subsequently needed to reconsider the issue as a complete and use a totally latest approach by attempting to that may influence the sales, hence should drive the forecast in return.
After many interviews with Supply, E-commerce, Merchandising and Marketing we agreed on the required output from the model:
– Forecasting the 3-month Sell-Through
– Modelling at SKU x Region Level
– Updating the forecasts each day
Note : the goal sell-though
is the initial forecast for the product at the three months mark, made and committed by the Teams before product launch
To counterbalance the undeniable fact that the scale of the history could be very small, we tried to aggregate a maximum of additional data, to provide more conttext to the model
We built a pipeline on Dataiku that was in a position to provide all the next data sources day by day:
- Actual sales volumes by distribution channel (stores / department shops / retailers / e-commerce)
- CRM data of the primary customers (age, nationality, latest or existing?)
- Traffic data on product web pages (click-through rates, conversion rates, acquisition channels)
- Variety of Instagram posts containing the product family name, variety of likes and reposts generated
- Retail Buy and Stocks Data
- Product attributes (materials, colours, prices) and product images
- Store attributes (geographic location, average sales volume)
We defined the Launch Similarity Rating between two products because the weighted distance between their principal attributes.
We then defined the Launch Similarity Rating as follows:
We had different options to best estimate weights in response to our problem:
1. Select it arbitrarily with business teams, their experience can tell us which attributes is most vital for products similarity calculation
2. Optimize these parameters inside an embedded grid search
Knowing w
weeks of sales, we desired to predict the 3-month Sell-Through. We built 1 model for every product age w in weeks ( w
in [1,6] ).
Conversely to traditional forecasting models where the output is usually absolutely the value of sales, we rebased the info with the the primary days’ sales. In other words, we trained the model not to reply the query “how much units will I sell, in absolute ? ” but “Will I sell roughly on average in the approaching weeks than prior to now few weeks ? ”.
For every product X, we construct the feature table with:
- All the info concerning product X characteristics (attributes, first sales, web traffic, etc. )
- Data concerning other previously launched products which are just like product X (using the Similarity rating previously defined).As these products are older, we will feed the model with all the data concerning their complete launch profile, and specifically their output (= goal) value
We tuned XGBoost hyperparameters with cross-validation on the train set after which used it to suit the info and predict on the test set.
To judge the performance of our model we used the common forecast metric Foracast Accuracy = 1 — WAPE, with A
being Actuals and F
Forecasts.
Listed here are the outcomes obtained in comparison with an easy linear model ( i.e extending first sales linearly). The x-axis is the variety of weeks after the launch.
We see that the , especially for the very first weeks of launch (1,2,3).
We’ve also calculated that . This doesn’t include advantages in storage and obsolescence costs from avoiding overstocks situations.