Intelligent and adaptable demand sensing solution improves forecast accuracy by 60%

Challenges.

Inability to maintain optimized inventory levels due to:

• Use of traditional demand forecast methods that used only historical demand data for SKUs across markets, geographies etc. were ineffective in providing accurate weekly forecasting results

Solution.

HitechDigital delivered an intelligent demand sensing solution which navigated the dependence on historic demand and leaned on more parameters such as inventory levels, number of customers, customer feedback, open orders etc. to enable more accurate forecasting to save non-compliance penalties along with ensuring customer delight.

The solution helped to:

• Automate data collection for parameters other than demand like customer visits, feedback, inventory status etc. to optimize the algorithm for accurate demand estimation.
• Select best algorithm among various algorithms to provide best results.
• Handle changing requirements such as increased demand volatility with good accuracy.
• Ensure scalability and flexibility across different geographies / product / market.

Approach.

Assessment

• HitechDigital data scientists and analysts conducted extensive brain storming sessions with the company’s product and forecast managers to understand the gaps caused by existing algorithmic analysis and data models and concluded that:
  • In a stable market scenario, existing standard time series algorithm used parameters like seasonality, cyclicity and patterns of historical sales data. But these algorithms cannot function with any dynamic inputs like promotion, changes in local environment, change in consumer behavior etc.
  • To meet online sales channel needs and volatile market, our data scientists needed to widen the database with parameters for promotions, environment, consumer behavior along with find a way to use these additional parameters and still forecast the future.

Implementation

• Collected data like total inventory, total number of customers visiting the site daily, total number of customers visiting the company’s products / SKUs on site, customer feedbacks on individual SKU’s etc. from the online retailer to widen the database
• We converted time series method to regression. As algorithms need to be trained on the specific data set – we transformed parameters to lag parameters for data analysis.  Also, established that the correlation between time duration and impact of the parameter.
• Setting a breakthrough for demand sensing, we built intelligence in Machine learning algorithms and regression algorithms like XG Boost and Decision tree for shorter time duration forecasts.
• Changed methodology from times series to a regression analysis to incorporate demand impacts due to parameters other than historic demand along with advance algorithms such as XG Boost, Decision tree etc.