Improvements to Forecasting Accuracies projects an estimated $2Million impact by allowing for a refined production schedule.**

The automated model framework and consumption dashboard will accelerate the scale up of this

This initiative has enabled the client to gain better visibility into business processes and data landscape

The artefacts created during the 8-week sprint are modular and can be re-used in initiatives beyond this project

Discord Deep-dive’s a mode of root-causing that renders post-gaming insightful, with potential to reduce gaps in forecasts beyond what’s allowed by just math.


The client is one of the world’s largest manufacturers of plumbing products. In the absence of an effective demand forecasting model, they faced problems such as less-than-desirable accuracy, additional cost in Inventory management (Insufficient/ Excessive inventory), inability to meet retailer demands on time, last-minute staffing and operational changes and expedited interplant and customer delivery.

The two major problems they faced with their existing demand forecasting solution were: The forecasting accuracy was low for some SKUs (Range of 44% to 66%) It was a black box solution, so it was unable to breakdown the forecasts to the component factors such as seasonality, promotions and market factors.

There were also other challenges in data collection, processing and consumption such as: Data silos (promo, price, market category, etc., all isolated from each other and from different sources, with different, non-standard formats) The account managers spent a lot of time estimating and making manual adjustments to these forecasts

Each manager had his/her own logic to make these adjustments. Hence there was lack of consistency across accounts and there were no means to track or log these adjustments. They realized that there was a lot of scope for improvement. They wanted a forecasting solution that not only had better accuracy but also was easy to interpret, consume and incorporate into their manufacturing process.


TheMathCompany partnered with the client to create an automated framework for demand forecasting that builds, evaluates and fine-tunes models based on an exhaustive list of model-parameters combinations.


The automated model framework created tries all these model-parameter combinations for each SKU and picks the combination that best captures the nature of that SKU. We included multiple models (6) in our solution and an exhaustive list of parameters (~5400) across all these models.


The Power BI dashboard (for consumption by account managers) is an interactive tool that is easy to interpret and consume and enables the user to visualize forecasts at multiple levels (As granular as SKU level and can be rolled up to a market category level). Our solution was designed in such a way that it could be deployed even in case of an expansion (From 250 SKUs in phase 1 to maybe 10000 SKUs in the future). Macroeconomic Indicators of the US Economy were encoded with automated API Feeds serving data to the model pipelines.”

The following forecasting techniques were applied, Univariate: Autoregressive Integrated Moving Average (ARIMA) Unobserved Components Model (UCM) Holt Winters Moving Average Multivariate: Extreme Gradient Boosting (XGBoost) Unobserved Components Model (UCM)


Want critical data insights to move your business needle?

Get Started
Linkedin Instagram Facebook Twitter
Cookies policy Privacy policy © 2020 TheMathCompany