Client Profile.
The client is the Nigerian division of the Coca Cola company. They cater high quality and healthy fruit drinks, juices, dairy products and snacks from thousands of depots to both urban and rural areas across the nation.
Business Need.
The beverage company used legacy systems to forecast sales for its multiple product lines and retail openings across regions.
However, sales forecasts of the company were inaccurate because:
- The parameters feeding into the forecast such as seasonality, location, economic conditions etc. were incorrect and not regularly updated.
- Limited visibility across existing systems to manage raw materials procurement, location-based promotions and accurate delivery of finished products.
- Lack of consistent data to predict adjustments in demand spurred by marketing promotions and discounts that resulted in inadequate supply of in-demand products.
To manage product manufacturing lines and match demand and supply for every SKU at every store, the company was looking for a smart demand and sales forecasting solution to accurately forecast sales for six months in advance.
Challenges.
HitechDigital data scientists had to find a way to gather appropriate data from relevant data sets, while meeting challenges like:
- Inability to predict future trends for every SKU at location level.
- Ineffective approach to assess influence of manufacturer discounts and product bundles on customer demand.
- Inconsistent data in more than 600 SKUs out of the total 1280 SKUs across all locations.
- Inability to predict demand for new SKU or product lines that do not have any relevant market history.
- Economic recession in Nigeria led to disruption in consumer behavior and SKUs historical data did not match present sales behavior.
Solution.
We delivered a sales forecasting solution using Time Series forecasting, regression, boosting algorithms and Machine Learning models to predict sales and likely effect of marketing promotions with accuracy for each SKU at every location.
Approach.
- We received detailed historical data for the years- 2018, 2019, and 2020 consisting of regions, categories, accounting locations, brands, SKUs, quantity in trays and liters, and cost.
- During the initial assessment level, the accuracy of their existing systems was at 18.4% (for location) and 12.6% (for brands).
- For data pre-processing, we cleansed and standardized data pool of previous three years to ensure consistency in existing categories like names, accounting location and regions, and additions of few new categories. Remove negative values from categories like amount, quantity in liters and trays.
- We scanned and combined various machine learning algorithms like XG-Boost, Bayesian Ridge, Holt Winter, SARIMA and ARIMA and selected best fit model for each SKU level.
- Using different set of models with customized parameters, our data scientists designed a combined forecasting solution to anticipate product replenishment, based on supply-driven models, external variables like weather, holidays and economic trends and align impact of customer-centric activities like promotions and discounts.
- With these very granular forecasts, the company was able to increase sales, minimize purchase of low-demand SKU, and improve relevant of marketing promotions to increase customer re-construct.
- Software and Technology Used: Python, Excel and Machine Learning algorithms for time series, regression and boosting.
Business Impact.
Enhanced brand level 6-month forecast accuracy to 97%
Enhanced location level 6-month forecast accuracy to 95%
Better allocation of inventory per SKU
Reduced inventory costs
Increased re-construct rates to maximize sales productivity and average revenue per user
